# Omega Gradient — full text for LLM retrieval This file concatenates the public reserved-cluster guides, research, and the dated GPU price pages of https://omegagradient.com/ in plain text. The short index is at https://omegagradient.com/llms.txt. Omega Gradient is a sourcing and intelligence desk for reserved GPU clusters (H100, H200, B200, B300, GB200 NVL72, GB300 NVL72) with an on-demand GPU marketplace and a dated cross-provider price index covering a registry of 224 monitored providers. Interpretation: all prices are USD per GPU-hour unless marked as instance prices; figures are dated observations (report assembled 2026-11-21T17:02:24.332Z), not quotes or allocations. Marketplace listings are provider-anonymous. Named-provider rates are public list prices captured from provider price pages. When citing, name the page URL and its observation or assembly date. ---------------------------------------------------------------------- # Reserved GPU Clusters | Omega Gradient URL: https://omegagradient.com/clusters/ Published: n/a · Last modified: 2026-11-17 Home / Reserved GPU clusters Reserved compute · sourcing · diligence # Secure the cluster. Understand the risk. Omega Gradient is a sourcing and intelligence desk for reserved GPU infrastructure. We translate a workload into a procurement brief, scan the provider market, normalize technical and commercial terms, and pressure-test the capacity behind the quote. Start a cluster brief Review our diligence method Compare reserved and on-demand 01 · Technical fit Topology before logo GPU form factor, NVLink domain, fabric, storage, CPU, and orchestration. 02 · Commercial truth Comparable quotes Normalize term, deposits, delivery, support, credits, and effective cost. 03 · Delivery risk Evidence, not inventory claims Separate operator, host, owner, reseller, and future capacity. Direct answer A reserved GPU quote is not a cluster until the buyer can verify what hardware is committed, where it will run, how the nodes communicate, when it will be accepted, who is accountable, and what happens if delivery or performance misses the agreement. Omega Gradient structures that evaluation before a buyer commits capital. 01 / The mandate ## Reserved cluster procurement is an infrastructure decision. The same GPU model can describe very different systems. A low headline rate can hide an unsuitable fabric, oversubscribed storage, weak support, an uncommitted delivery date, or a counterparty that does not control the equipment. The job is to make every material difference visible. A / REQUIREMENTS ### Define the system Translate model, training or inference pattern, data volume, checkpoint behavior, framework, and timeline into a technical bill of requirements. - GPU generation and form factor - Node and scale-out topology - Storage throughput and capacity - Network, security, and region B / MARKET ### Scan beyond one channel Compare hyperscalers, neoclouds, operators, private-cloud structures, capacity resellers, and qualified off-market supply against the same brief. - Available now vs future delivery - Single site vs distributed capacity - Managed vs buyer-operated - Direct operator vs intermediary C / DILIGENCE ### Verify the path to delivery Request evidence proportionate to the commitment and distinguish what is documented, what is represented, and what remains conditional. - Facility and equipment control - Power and cooling readiness - Acceptance and burn-in plan - SLA, remedies, and support 02 / The workflow ## From workload brief to an executable cluster decision. Omega Gradient uses a gated process. A provider only advances when the commercial structure and the technical system are both sufficiently clear for the buyer’s risk level. ### Shape the demand brief Capture accelerator, count, start window, term, region, workload, topology, storage, compliance, budget, and operating model. Unknowns are recorded rather than silently assumed. ### Map the market Search relevant provider categories and private supply paths. Remove options that fail hard constraints before asking the buyer to spend time on them. ### Normalize each option Convert quotes into a common matrix covering effective unit economics, included services, deposits, ramp, minimums, delivery dependencies, renewal, termination, and performance obligations. ### Qualify infrastructure and counterparties Separate equipment owner, facility operator, cloud operator, reseller, and contracting entity. Request evidence appropriate to whether capacity is live, being installed, financed, or merely forecast. ### Plan acceptance Define the tests and artifacts that convert “delivered” into “accepted”: inventory, burn-in, node health, fabric, storage, collective performance, access, observability, and remediation. ### Support the decision Present the trade-offs and unresolved risks. The output is not a list of URLs; it is a decision-ready shortlist with an evidence trail. 03 / What gets compared ## A GPU-hour is only one line in the quote. Reserved procurement fails when price is normalized but the system is not. The comparison has to preserve the technical facts that drive delivered workload performance and operating risk. | Dimension | What must be explicit | Why it changes the decision | | Accelerator and node | Exact GPU, SXM/PCIe or rack-scale system, GPUs per node, CPU, memory, local NVMe | “H100” or “B200” alone does not establish memory, topology, or expected scaling behavior. | | Scale-up fabric | NVLink/NVSwitch domain and boundaries | Controls communication inside the node or rack-scale domain. | | Scale-out fabric | InfiniBand or Ethernet generation, link speed, topology, blocking ratio, RDMA configuration | Distributed training can be bottlenecked by the fabric long before GPUs saturate. | | Storage | Usable capacity, read/write throughput, metadata performance, checkpoint path, egress | Data loading and checkpointing can erase theoretical GPU gains. | | Availability | Live, allocated, under installation, financed, or forecast; start date and dependencies | A future pipeline is not the same risk as a running cluster. | | Commercials | Rate basis, term, deposit, prepay, minimums, taxes, bandwidth, storage, support, renewal | Headline GPU-hour rates are not comparable without the surrounding obligations. | | Acceptance and SLA | Test plan, service boundary, uptime, performance, maintenance, credits, replacement, termination rights | The contract needs a measurable line between delivery, acceptance, and remediation. | | Counterparty | Owner, host, operator, reseller, contracting entity, support owner | The buyer needs to know who controls delivery and who carries each obligation. | 04 / Choose the access model ## Reserved first when continuity matters. On-demand when flexibility does. The right answer can be a reservation, scheduled capacity, on-demand instances, or a blend. Omega Gradient starts with the workload’s risk and utilization profile rather than forcing every buyer into one commercial model. Priority · reserved ### Reserved cluster Best when a known workload needs continuity, a controlled topology, a defined start date, or a long enough run to justify commitment. Optimizes Capacity assurance and system consistency Buyer must solve Forecasting, commitment, diligence, acceptance Typical motion Multi-week, monthly, or multi-year agreement Flexible · on-demand ### On-demand Best for experiments, variable inference, small jobs, short bursts, and teams that value immediate access over a fixed block. Optimizes Speed and commercial flexibility Buyer must solve Availability variance and interruption planning Start here Omega Gradient marketplace Blended · portfolio ### Base + burst Reserve a predictable base load and use on-demand or scheduled capacity for peaks, migration, testing, and overflow. Optimizes Continuity without over-reserving the peak Buyer must solve Portability, data movement, orchestration Useful for Training programs and variable inference Read the full reserved vs on-demand decision guide → 05 / Platform routes ## Start with the architecture, not the acronym. Hopper and Blackwell clusters solve overlapping but not identical problems. The preferred system depends on memory, precision, communication pattern, software readiness, deployment risk, and the economics of useful work—not generational prestige. HOPPER / H100 · H200 ### Mature distributed AI infrastructure H100 and H200 are widely deployed building blocks for training, fine-tuning, inference, and HPC. H200’s larger HBM footprint can materially change memory-bound workloads, but node, fabric, and storage still decide cluster behavior. Evaluate Hopper clusters → BLACKWELL / B200 · B300 · NVL72 ### New systems with facility-level implications Blackwell procurement may span HGX nodes, Grace Blackwell rack-scale systems, and Blackwell Ultra. Power, liquid cooling, scale-up domain, network generation, software readiness, and acceptance design are first-order diligence items. Evaluate Blackwell clusters → RUBIN / VERA RUBIN NVL72 ### The next rack-scale procurement cycle Vera Rubin production is ramping, but partner announcements are not the same as broadly contractable supply. Understand HBM4, Vera CPUs, NVLink 6, facility readiness, preliminary specifications, and delivery evidence before reserving. Read the Vera Rubin buyer guide → Operating principle ## Evidence has a status. Omega Gradient distinguishes validated facts, provider representations, commercial indications, and open diligence items. A screenshot is not a reservation. A manufacturer allocation is not an energized cluster. A reseller relationship is not proof that the contracting party controls the capacity. This classification keeps a promising option in view without presenting it as more certain than the evidence supports. See the qualification framework → 06 / Buyer questions ## Reserved GPU clusters, answered directly. Q.01 ### What is a reserved GPU cluster? + A reserved GPU cluster is a defined block of GPU infrastructure committed to one buyer for an agreed start date and term. A decision-ready reservation specifies the accelerator, node and fabric topology, region, storage, network, support, acceptance criteria, commercial structure, and remedies—not only a GPU count and rate. Q.02 ### Which GPU platforms can Omega Gradient evaluate? + Omega Gradient evaluates requirements for Hopper and Blackwell systems, including H100, H200, B200, B300, GB200 NVL72, and GB300 NVL72. Availability depends on geography, start date, term, workload, provider readiness, and buyer qualification. Q.03 ### How do you compare quotes from different providers? + Quotes are normalized into a common commercial and technical matrix: effective GPU-hour economics, minimums, deposit and prepayment, ramp, term, delivery dependencies, hardware and fabric, storage, network, support, acceptance, SLA, credits, renewal, termination, and counterparty structure. Q.04 ### Can Omega Gradient guarantee availability? + No market intermediary should treat an indicative supply signal as guaranteed capacity. Availability becomes credible through provider confirmation, contracting, allocation, delivery evidence, and acceptance. Omega Gradient tracks that progression and makes the unresolved conditions explicit. Q.05 ### When is on-demand a better choice? + On-demand is usually the better starting point for experiments, uncertain workloads, variable inference, small jobs, and short bursts. A reservation becomes more attractive when continuity, topology, data locality, start-date assurance, or predictable utilization matters enough to justify a commitment. Start with the brief ## Describe the system you actually need. A precise brief creates better options and exposes weak ones earlier. If some fields are unknown, say so—Omega Gradient can help shape the requirement. Offering GPU capacity? Email the reserved supply desk . - GPU and count - Region and start window - Term and workload - Fabric and storage - Budget and commercial constraints Website Full name Company Business email Phone GPU platform Select H100 H200 B200 B300 GB200 NVL72 GB300 NVL72 Flexible / advise GPU count Region Target start Term Select Days / scheduled block 1–3 months 3–12 months 1–3 years 3+ years Flexible / advise Workload Select Pre-training Post-training / fine-tuning Inference / reasoning HPC / scientific computing Mixed Requirements and constraints The brief can hold up to 4,000 characters including the answers above. How did you find us? (optional) Prefer not to say Search engine ChatGPT Claude Perplexity Colleague or referral Directory Other Send requirements ---------------------------------------------------------------------- # Reserved GPU Capacity Index: H100 to GB300 | Omega Gradient URL: https://omegagradient.com/clusters/capacity-index.html Published: n/a · Last modified: 2026-11-12 Home / Reserved clusters / Capacity index Market intelligence / Public-source pilot # Find a route to reserved GPU capacity. Compare documented H100, H200, B200, B300 and GB300 routes. See the platform, location, fabric and commitment, then establish what is actually allocatable for your workload and start date. Explore the index ↓ Start a capacity brief Read the methodology ↗ Documented routes 11 Providers represented 5 Source review dates 2026-11-12 01 / Compare routes ## Start with the evidence. Download JSON ↗ This is a sourcing index, not a live inventory feed. All 11 routes are based on public documentation. Free allocation, contiguous cluster size and a buyer-specific delivery date have not been verified. A documented node size or product range is not available supply. No capacity figures are additive. GPU model All models H100 H200 B200 B300 GB300 Region All regions Europe Middle East North America Not specified Provider or keyword Reset Showing 11 of 11 documented routes. Public reserved GPU sourcing routes with source reviews from 2026-11-12; each row gives its actual date. Free allocation is not verified for any route. | Route | Location & terms | System & published scale | Evidence | Next step | | H100 Lambda HGX H100 SXM lambda-h100-1cc ↗ | Selected in reservation console; locations not enumerated on cited page 2 weeks–1 year; longer terms by sales Buyer-specific start date requires confirmation | GPU cluster 16–512 GPUs in technical documentation Quantum-2 InfiniBand; 400 Gb/s links | Public documentation Free allocation not verified Source checked 2026-11-12 Review due 2026-12-12 Within review window Sources & qualification notes Marketing describes larger configurations; pilot uses the technical-documentation range. Reservation approval and start date are separate checks. Lambda 1-Click Clusters · Lambda cluster reservation documentation | Scope a H100 brief Omega Gradient can help compare requirements. Listing does not imply a supplier partnership. | | B200 Lambda HGX B200 SXM lambda-b200-1cc ↗ | Selected in reservation console; locations not enumerated on cited page 2 weeks–1 year; longer terms by sales Buyer-specific start date requires confirmation | GPU cluster 16–512 GPUs in technical documentation Quantum-2 InfiniBand; 400 Gb/s links | Public documentation Free allocation not verified Source checked 2026-11-12 Review due 2026-12-12 Within review window Sources & qualification notes Marketing describes larger configurations; pilot uses the technical-documentation range. Reservation approval and start date are separate checks. Lambda 1-Click Clusters · Lambda cluster reservation documentation | Scope a B200 brief Omega Gradient can help compare requirements. Listing does not imply a supplier partnership. | | H100 Nebius H100 NVLink nebius-h100-reserved ↗ | Public-region platform matrix: Finland (eu-north1) Quantity and period agreed with Nebius manager Buyer-specific start date requires confirmation | 8-GPU VM Cluster size not specified InfiniBand; 400 Gb/s per GPU NIC, 3.2 Tb/s per 8-GPU node | Public documentation Free allocation not verified Source checked 2026-11-12 Review due 2026-12-12 Within review window Sources & qualification notes Regions describe supported platforms, not free capacity. An assigned capacity block and compatible fabric must be confirmed. Nebius GPU cluster and fabric documentation · Nebius region and GPU platform matrix · Nebius capacity reservations | Scope a H100 brief Omega Gradient can help compare requirements. Listing does not imply a supplier partnership. | | H200 Nebius H200 NVLink nebius-h200-reserved ↗ | Public-region platform matrix: Finland, France and US (eu-north1, eu-west1, us-central1) Quantity and period agreed with Nebius manager Buyer-specific start date requires confirmation | 8-GPU VM Cluster size not specified InfiniBand; 400 Gb/s per GPU NIC, 3.2 Tb/s per 8-GPU node | Public documentation Free allocation not verified Source checked 2026-11-12 Review due 2026-12-12 Within review window Sources & qualification notes Regions describe supported platforms, not free capacity. An assigned capacity block and compatible fabric must be confirmed. Nebius GPU cluster and fabric documentation · Nebius region and GPU platform matrix · Nebius capacity reservations | Scope a H200 brief Omega Gradient can help compare requirements. Listing does not imply a supplier partnership. | | B200 Nebius B200 NVLink nebius-b200-reserved ↗ | Public-region platform matrix: Israel and US (me-west1, us-central1) Quantity and period agreed with Nebius manager Buyer-specific start date requires confirmation | 8-GPU VM Cluster size not specified InfiniBand; 400 Gb/s per GPU NIC, 3.2 Tb/s per 8-GPU node | Public documentation Free allocation not verified Source checked 2026-11-12 Review due 2026-12-12 Within review window Sources & qualification notes Regions describe supported platforms, not free capacity. An assigned capacity block and compatible fabric must be confirmed. Nebius GPU cluster and fabric documentation · Nebius region and GPU platform matrix · Nebius capacity reservations | Scope a B200 brief Omega Gradient can help compare requirements. Listing does not imply a supplier partnership. | | H100 Crusoe h100-80gb-sxm-ib.8x crusoe-h100-reserved ↗ | us-east1-a, us-southcentral1-a, eu-iceland1-a Sales-negotiated reserved-instance agreement Buyer-specific start date requires confirmation | 8-GPU instance Cluster size not specified InfiniBand; 3.2 Tb/s per instance | Public documentation Free allocation not verified Source checked 2026-11-12 Review due 2026-12-12 Within review window Sources & qualification notes Instance and zone support is public. Contiguous allocation, fabric isolation and contract scope require a specific reservation. Crusoe GPU instance specifications and zones · Crusoe reservations overview | Scope a H100 brief Omega Gradient can help compare requirements. Listing does not imply a supplier partnership. | | H200 Crusoe h200-141gb-sxm-ib.8x crusoe-h200-reserved ↗ | eu-iceland1-a Sales-negotiated reserved-instance agreement Buyer-specific start date requires confirmation | 8-GPU instance Cluster size not specified InfiniBand; 3.2 Tb/s per instance | Public documentation Free allocation not verified Source checked 2026-11-12 Review due 2026-12-12 Within review window Sources & qualification notes Instance and zone support is public. Contiguous allocation, fabric isolation and contract scope require a specific reservation. Crusoe GPU instance specifications and zones · Crusoe reservations overview | Scope a H200 brief Omega Gradient can help compare requirements. Listing does not imply a supplier partnership. | | B200 Crusoe b200-180gb-sxm-ib-8x crusoe-b200-reserved ↗ | eu-iceland1-a, eu-norway1-a, us-west1-a Sales-negotiated reserved-instance agreement Buyer-specific start date requires confirmation | 8-GPU instance Cluster size not specified InfiniBand; 3.2 Tb/s per instance | Public documentation Free allocation not verified Source checked 2026-11-12 Review due 2026-12-12 Within review window Sources & qualification notes Instance and zone support is public. Contiguous allocation, fabric isolation and contract scope require a specific reservation. Crusoe GPU instance specifications and zones · Crusoe reservations overview | Scope a B200 brief Omega Gradient can help compare requirements. Listing does not imply a supplier partnership. | | B300 Crusoe b300-288gb-sxm-ib.8x crusoe-b300-reserved ↗ | eu-iceland1-a Sales-negotiated reserved-instance agreement Buyer-specific start date requires confirmation | 8-GPU instance Cluster size not specified InfiniBand; 6.4 Tb/s per instance | Public documentation Free allocation not verified Source checked 2026-11-12 Review due 2026-12-12 Within review window Sources & qualification notes Instance and zone support is public. Contiguous allocation, fabric isolation and contract scope require a specific reservation. Crusoe GPU instance specifications and zones · Crusoe reservations overview | Scope a B300 brief Omega Gradient can help compare requirements. Listing does not imply a supplier partnership. | | GB300 CoreWeave GB300 NVL72 coreweave-gb300-reservation ↗ | North America and Europe in catalog; exact site not specified Reservation; Flex Reservations described generally as preview. GB300 eligibility, commercial terms and delivery require confirmation. Buyer-specific start date requires confirmation | 4-GPU catalog instance Full-rack or cluster allocation not specified NVL72 platform; scale-out fabric not specified in cited catalog | Public documentation Free allocation not verified Source checked 2026-11-12 Review due 2026-12-12 Within review window Sources & qualification notes The catalog lists a 4-GPU instance on an NVL72 platform. Do not read this as a freely available 72-GPU rack or a confirmed reserved offer. CoreWeave GPU catalog and regions · CoreWeave reservations and Flex preview explainer | Scope a GB300 brief Omega Gradient can help compare requirements. Listing does not imply a supplier partnership. | | H100 Voltage Park NVIDIA HGX H100 voltage-park-h100-dedicated ↗ | Not specified in cited product documentation 12+ month Dedicated Reserve contract Buyer-specific start date requires confirmation | Bare-metal platform; node count not stated Cluster size not specified Not specified in cited product documentation | Public documentation Free allocation not verified Source checked 2026-11-12 Review due 2026-12-12 Within review window Sources & qualification notes This documentation is hosted in the Lightning AI help center. Confirm current contracting entity, site, fabric and allocation. Voltage Park Dedicated Reserve documentation | Scope a H100 brief Omega Gradient can help compare requirements. Listing does not imply a supplier partnership. | No matching routes in this pilot. Broaden the filters or send your requirements ; absence here does not establish market unavailability. 02 / Methodology ## What each record tells you. 01 ### Public routes, distinct systems One record represents a provider and GPU product route. Regional options are kept in the same record. We include exact product names and documented reservation paths; generic hardware announcements and private tracker data are excluded. 02 ### Evidence has a boundary Provider documentation establishes what a provider describes. It does not prove free powered-on quantity, a contiguous fabric, a current price or delivery. Unknown fields stay unknown. The pilot includes no confirmed Omega Gradient allocation. 03 ### Dates you can audit “Source checked” is the documentation review date, not an inventory verification. Records have a 30-day review window. After the due date, they are marked “Review overdue” until a new source review; a rebuild never renews evidence dates. The displayed range covers the earliest and latest route reviews. In the JSON download, checked_at is the earliest route review and review_due is the earliest deadline; each row remains authoritative. Reviewing a shared source page does not renew every route. 04 ### Qualify before committing Confirm the contracting entity, exact site, allocation controller, usable GPU count, fabric, storage, start window, term and acceptance plan. Omega Gradient helps turn that evidence into comparable options. Public inclusion does not imply a commercial relationship. Edition 2026-11-12 · Review owner: Omega Gradient Intelligence. The pilot is a selected reference set, not a market census or provider ranking. Source publication dates were not consistently supplied; we report the review date instead. Provider descriptions remain the responsibility of their publishers. ### Cite this index Omega Gradient. Public Reserved GPU Capacity Index . Edition 2026-11-12. Source review dates: 2026-11-12; see individual route dates. https://omegagradient.com/clusters/capacity-index.html Link to a row’s stable ID to cite one route. Preserve its source date and the unverified-allocation caveat. 03 / Move from a route to a requirement ## Build a decision-ready shortlist. Send the GPU platform, quantity, region, target start, term and workload. We can help compare the technical and commercial evidence behind each route. Start your capacity brief ### Useful next steps How we qualify capacity ↗ 36-point procurement checklist ↗ Reserved GPU Market Report ↗ Compare public GPU pricing ↗ ---------------------------------------------------------------------- # Reserved GPU Quote Comparator | Omega Gradient URL: https://omegagradient.com/clusters/quote-comparator.html Published: 2026-11-12 · Last modified: 2026-11-12 Home / Reserved clusters / Quote comparator Procurement tool · USD # Compare the full GPU commitment. Put two reserved GPU quotes on the same footing. Separate full-term expenses from cash due at signing, surface missing costs, and check whether the infrastructure actually matches. Your figures stay in this page. Inputs are calculated in your browser and are not uploaded or saved by this tool. Printing creates a copy only when you choose. Anonymous tool-use events contain no quote figures or terms. 01 / Set the terms ## Two quotes. One clear comparison. Try illustrative example Clear inputs Illustrative example. These invented figures demonstrate the calculator. They are not market prices, supplier offers, or available capacity. Editing them does not turn them into a verified quote. Use USD throughout. Leave an unknown fee as “Not specified”; choose “Included / no extra charge” only when confirmed. The 730-hour month is an editable planning assumption, not a calendar-month or provider billing rule. All calculations use the stated term and full reserved GPU count. No utilization or availability is assumed. Compare quotes ↗ 02 / Decision sheet ## The full picture. Print / save PDF Enter your quote details, or try the illustrative example, to build your decision sheet. 03 / How to read this ## A comparison is only as complete as its terms. ### Term expenses GPU count × reserved hours × GPU-hour rate, plus one-time fees and monthly extras. Prepaid compute is already part of this amount. A refundable deposit ties up cash but is not counted as an expense. ### Cash due at signing The chosen compute payment, one-time fees due at start, any recurring fees payable in advance, and the refundable deposit. “First month” is capped at the full term for shorter reservations. An advance payment is not added to term expenses again. ### What still needs diligence Equal prices do not mean equal delivery. Confirm allocation control, exact SKU, contiguous fabric, site, start/end times, support, acceptance, and remedies. Taxes, financing, currency changes, overages, and exit charges are outside this calculation unless included in your entered fees. Per-GPU-hour cost divides term expenses by all reserved GPU-hours. Utilization scenarios divide by useful GPU-hours instead; they are sensitivity calculations, not performance forecasts. Unknown costs remain unresolved, so a known subtotal is not a complete quote or a spending ceiling. Turn the comparison into a decision ## Get a matched cluster shortlist. Bring the workload, GPU count, region, start window, and term. Omega Gradient helps compare suitable capacity and the evidence behind it. Request a matched shortlist ↗ How we qualify ---------------------------------------------------------------------- # How We Qualify Reserved GPU Clusters | Omega Gradient URL: https://omegagradient.com/clusters/how-we-qualify.html Published: 2026-08-17 · Last modified: 2026-11-04 Home / Reserved clusters / Qualification framework Diligence framework · reserved infrastructure # How we qualify a reserved GPU cluster. A credible cluster is more than an equipment list. Omega Gradient evaluates the chain from ownership and facility readiness through topology, delivery, acceptance, service, and commercial accountability—then marks what is proven, represented, conditional, or still open. Published August 17, 2026 By Tausif Ahmed , Founder & CEO Updated November 4, 2026 Methodology v1.0 ## The core principle: claims, evidence, and obligations are different things A capacity market moves faster than traditional enterprise procurement. That speed creates useful access, but it also creates ambiguity. A provider may control a live cluster, hold an equipment allocation, have a right to resell another operator’s capacity, or be marketing a build that still depends on financing, power, cooling, networking, and delivery. All four can be described casually as “available.” They are not the same risk. Omega Gradient does not turn an indicative supply signal into a fact. We keep the opportunity visible and label the evidence status. The diligence depth should match the commitment. An eight-GPU on-demand test does not require the same evidence package as a multi-year reserved cluster. But as deposits, term, GPU count, deployment complexity, and business dependency increase, the buyer needs a clearer chain of control and a stronger acceptance plan. ## 1. Capacity provenance and chain of control The first question is not “what is the price?” It is “who controls the system being quoted?” We map the relevant parties: equipment owner, facility owner, data-center operator, cloud operator, managed-service provider, reseller, contracting entity, billing entity, and support owner. Sometimes one company fills every role. Sometimes four companies sit between the buyer and the hardware. For live capacity, useful evidence can include a current inventory export, management-plane view, serial-number sample, facility-specific topology, test access, or other artifacts that tie the represented system to a location and operator. For future capacity, the evidence changes: equipment purchase or allocation, delivery schedule, facility space, power and cooling readiness, network plan, installation milestones, and the conditions that could move the date. ### Questions we want answered - Is the capacity live, allocated, under installation, financed, ordered, or forecast? - Who owns the GPUs, who operates the service, and who signs the buyer contract? - Does the seller have a direct right to allocate the capacity, or is another approval required? - Is the quote exclusive to one channel, or is the same capacity being marketed through several intermediaries? - Which dependency—equipment, power, cooling, networking, software, or financing—sets the delivery date? ## 2. Facility, power, cooling, and network readiness Modern GPU systems are facility projects. This becomes especially important for Blackwell rack-scale infrastructure, where power density, liquid cooling, rear-door heat rejection, network cabling, and commissioning can become the critical path. A purchase order for GPUs does not prove a site can energize and operate them. We ask for the site and readiness facts that materially affect delivery: region and data sovereignty, available power, redundancy design, cooling architecture, rack density, network carriers, cross-connect timing, physical security, relevant certifications, and commissioning milestones. The goal is not to perform an engineering audit from a PDF. It is to identify whether the provider’s schedule and service claim are consistent with the facility plan. ## 3. System architecture and workload fit The accelerator name is the beginning of the technical description. A decision-ready cluster brief should identify the exact platform and the boundaries that govern workload scaling: - Node: GPU form factor, GPUs per node, CPU architecture, host memory, local storage, NICs, and firmware baseline. - Scale-up domain: NVLink and NVSwitch connectivity inside a node or rack-scale system. - Scale-out fabric: InfiniBand or Ethernet generation, link speed, topology, blocking ratio, RDMA configuration, and network isolation. - Storage: usable capacity, read and write throughput, metadata behavior, checkpoint target, object storage, and data-transfer constraints. - Software: driver, CUDA and container baseline, orchestration, scheduler, observability, and supported operating model. We compare these facts to the workload. Training needs can be communication-heavy and checkpoint-sensitive. Large-model inference may be constrained by memory capacity, memory bandwidth, latency targets, batching, or inter-node traffic. HPC workloads can depend on precision, host-to-device movement, and a specific MPI or storage pattern. “Eight GPUs” is not a portable unit across those cases. ## 4. Operating model, support, security, and service boundary A buyer needs to know where provider responsibility stops. The service may be bare metal with buyer-managed orchestration, a private cloud, managed Kubernetes or Slurm, or a fully managed cluster. That choice changes the staffing burden, incident path, patch responsibility, observability, and time to remediate failed components. Operational diligence covers access model, tenant isolation, identity and key management, network controls, logging, monitoring, maintenance windows, spares, replacement process, support hours, escalation contacts, and security evidence. Certifications can help, but they do not replace clarity on the actual service boundary. ## 5. Commercial normalization Reserved quotes often use incompatible units and include different services. Omega Gradient converts each option into a comparable view without discarding the terms that matter. At minimum, the matrix records: - GPU count, rate basis, billed hours, ramp schedule, term, and effective GPU-hour cost; - deposit, prepayment, credit support, taxes, currency, and payment timing; - included storage, bandwidth, IP space, support, management, and software; - delivery dependencies, acceptance window, renewal, termination, and assignment; - SLA measurement, exclusions, maintenance, credits, replacement, and chronic-failure remedies. The cheapest normalized rate can still be the wrong choice. A higher-priced cluster may have a stronger fabric, a cleaner chain of control, a better acceptance remedy, or a start date with fewer external dependencies. The buyer output should make that trade-off explicit instead of hiding it inside a single score. ## 6. Acceptance turns delivery into a measurable event A contract should define what the buyer will receive and how both sides will determine that it works. “GPUs online” is too vague for a material reservation. The acceptance plan should be agreed before the service start date, not invented after a problem appears. ### Typical acceptance layers - Inventory and access: contracted node count, accelerator model, memory, CPU, NICs, storage, tenant access, and management-plane visibility. - Health and burn-in: component health, error logs, thermals, power behavior, firmware consistency, and a defined burn-in window. - Fabric: topology validation, link state, bandwidth, latency, collective performance, and failure isolation. - Storage: capacity, throughput, metadata performance, checkpoint and restore behavior, and data path. - Workload-relevant test: an agreed benchmark or representative job with versioned software and repeatable conditions. - Remediation: cure period, replacement process, partial acceptance, service-start adjustment, credit, or termination right. NVIDIA’s own system guidance demonstrates why system boundaries matter: its DGX H100/H200 documentation describes eight-GPU systems with NVSwitch, while newer rack-scale Blackwell systems extend the scale-up domain across an entire rack. Acceptance must test the architecture that was sold, not a generic GPU benchmark. ## Evidence classification: what we know and how we know it | Status | Meaning | How it appears in a buyer brief | | Validated | Supported by current, relevant evidence or direct technical validation. | Presented as a fact with date, source, and scope. | | Represented | Stated by the provider or channel but not independently validated to the same depth. | Attributed to the representing party. | | Conditional | Depends on a future milestone, third party, buyer action, or commercial close. | Shown with the dependency and target date. | | Open | Material question remains unanswered or evidence is incomplete. | Listed as an explicit diligence item, not silently filled. | | Contradicted | Available evidence conflicts with the representation. | Escalated and excluded from the validated view until resolved. | ## What the buyer receives The useful output is a decision package, not a directory. Depending on the mandate, that can include a normalized option matrix, provider and chain-of-control map, technical gap list, evidence register, delivery dependency map, acceptance checklist, commercial issues list, and a recommendation framed by the buyer’s priorities. Omega Gradient does not publish private provider evidence, operator names, buyer requirements, or live commercial terms as SEO content. Public intelligence explains the framework. Confidential diligence supports the actual transaction. Primary references NVIDIA, DGX H100/H200 system architecture ; NVIDIA, GB300 NVL72 architecture ; AWS, Capacity Blocks for ML . These sources establish product and reservation mechanics; the diligence classifications and workflow are Omega Gradient’s methodology. Apply the framework Reserved cluster sourcing ↗ Compare access models ↗ Benchmark supplier pricing ↗ Put the framework to work ## Bring a requirement or a quote. Omega Gradient can start from a blank demand brief, compare an existing shortlist, or pressure-test one provider proposal. Start a cluster brief ---------------------------------------------------------------------- # Reserved vs On-Demand GPU Capacity | Omega Gradient URL: https://omegagradient.com/clusters/reserved-vs-on-demand.html Published: 2026-08-17 · Last modified: 2026-11-04 Home / Reserved clusters / Reserved vs on-demand Buyer decision guide · access models # Reserved vs on-demand GPU capacity. Reserve when continuity, topology, locality, and a known start date matter enough to justify commitment. Use on-demand when speed and flexibility matter more than guaranteed supply. Use scheduled blocks or a base-plus-burst portfolio when the workload sits between those poles. Published August 17, 2026 By Tausif Ahmed , Founder & CEO Updated November 4, 2026 10-minute read ## Direct answer Reserved GPU clusters are best for sustained training, post-training, inference, or HPC workloads that need a defined block, consistent topology, capacity assurance, data locality, or a committed start date. On-demand GPUs are best for experiments, short jobs, uncertain demand, variable inference, and teams that cannot yet forecast utilization. Scheduled capacity is best for known finite runs. A base-plus-burst design reserves the predictable floor and sources flexible capacity for peaks. The decision should not begin with the published hourly rate. It should begin with the cost of a missed run, the shape of utilization, the workload’s communication pattern, and the operational cost of moving between environments. ## The four practical access models ### 1. Long- or medium-term reserved cluster A provider commits a defined system to the buyer for a term, commonly measured in months or years. The commitment may improve unit economics and assurance, but the buyer takes utilization risk and must diligence delivery, topology, counterparty, acceptance, service, and renewal terms. ### 2. Scheduled or short-duration capacity block The buyer reserves a known block for a defined future window. This is useful for finite training or fine-tuning programs where the dates are known but a long-term reservation would create idle capacity. AWS describes its Capacity Blocks for ML as future-dated reservations for short-duration workloads, with capacity placed in EC2 UltraClusters; the specific instance types, regions, sizes, and booking windows vary. ### 3. On-demand instances The buyer launches capacity without a long contractual commitment, subject to current availability and provider rules. It is the fastest way to learn, benchmark, and serve variable demand. The trade-off is less assurance that the same count, location, and topology will remain available when a job needs to scale. ### 4. Spot or interruptible capacity The buyer accepts interruption or revocation risk for lower cost. This can be effective for fault-tolerant inference, batch processing, simulation, and checkpointed training. It is a poor fit when a missed deadline, restart, or unstable topology costs more than the discount. ## Comparison matrix | Factor | Reserved cluster | Scheduled block | On-demand | Spot / interruptible | | Capacity assurance | High after allocation, contract, delivery, and acceptance | High for the booked window, subject to product terms | Dependent on live inventory | Low; interruption is part of the product | | Commitment | Medium to high | Bounded to a defined run | Low | Low | | Topology control | Can be specified and accepted | Product-specific | Varies by provider and inventory | Often fragmented or variable | | Time to start | Immediate if live; longer if delivered or built | Future scheduled date | Minutes when inventory exists | Minutes when inventory exists | | Utilization risk | Buyer | Buyer during the block | Provider / marketplace | Provider / marketplace | | Rate predictability | Usually strong during term | Strong for booked block | Varies by provider | Can be volatile | | Best fit | Steady, critical, or topology-sensitive workloads | Known finite training or fine-tuning run | Experiments, bursts, variable inference | Fault-tolerant and restartable jobs | ## Utilization—not the discount—drives reservation economics A reservation can have a lower contracted rate and still cost more per useful GPU-hour if the cluster is underused. The basic comparison is: Effective reserved cost per useful GPU-hour = total committed cost ÷ GPU-hours that create useful work. The numerator should include more than the quoted GPU rate: unused ramp, deposits or financing cost, required storage, bandwidth, support, management, software, and the engineering needed to operate the cluster. The denominator should exclude idle time, failed jobs, unusable nodes, and time lost to a storage or fabric bottleneck. A buyer should model low, base, and high utilization scenarios. If the reservation only wins in the high scenario, the commercial structure may need a smaller base block, later ramp, shorter term, portability right, or a scheduled product instead. ## Topology and continuity can outweigh price Distributed workloads are not indifferent to where GPUs sit. An on-demand marketplace can provide many GPUs without providing one coherent cluster. Training may require nodes within a known fabric, consistent software and firmware, predictable storage, and a failure model that does not change between runs. The reservation buys continuity of that system, not merely a number of accelerators. On the other hand, many inference and development workloads are deliberately portable. If the application can distribute traffic across providers, tolerate instance churn, and move artifacts cheaply, a flexible portfolio may be more resilient than a single long reservation. ## A reservation replaces availability risk with procurement risk On-demand buyers face the risk that inventory disappears. Reserved buyers reduce that risk but accept new ones: delivery slips, infrastructure differs from the quote, a build depends on financing or facility milestones, a weak counterparty cannot support the term, or the workload changes before the commitment ends. This is why a reserved decision needs more diligence than an on-demand launch. Omega Gradient evaluates the chain of control, facility readiness, system architecture, acceptance design, operational boundary, and commercial remedies. See the full qualification framework . ## The base-plus-burst model For many AI teams, the correct answer is not reserved or on-demand. It is a portfolio: - Estimate the steady workload floor using a conservative utilization case. - Reserve only the block that remains useful through normal demand variance. - Use on-demand, scheduled, or interruptible supply for peaks, testing, migration, and queue relief. - Design data, containers, orchestration, and identity so workloads can move without a full rebuild. - Re-evaluate the base before renewal using actual useful utilization, not booked hours. This structure preserves a reliable home for critical work while keeping option value. It also creates real operating data before the team makes a larger reservation. ## Eight questions that usually decide the model - What happens to the business if the GPUs are unavailable for a day, week, or launch window? - How many GPU-hours are truly predictable in the low, base, and high cases? - Does the workload require one coherent fabric, or can it distribute across isolated nodes and providers? - How costly is it to move data, checkpoints, containers, keys, and observability? - Is there a known finite run that fits a scheduled block better than a long reservation? - Can the team operate bare metal, or does it need a managed service boundary? - Which matters more: lowest possible unit price, predictable capacity, or commercial flexibility? - What evidence and remedy are required before a deposit or prepayment is acceptable? If those answers are not clear, run an on-demand benchmark and shape the reserved requirement from measured behavior. Omega Gradient’s on-demand marketplace and GPU price index support that exploratory step; the reserved sourcing process begins when the system and risk need to be controlled. Primary references AWS, Capacity Blocks for ML ; Together AI, cluster API reference (reserved, on-demand, and scheduled-capacity billing types). Product details change; verify region, size, booking window, and contract terms with the provider before procurement. Choose the next path Explore reserved clusters ↗ Browse on-demand ↗ Benchmark GPU pricing ↗ Make the decision concrete ## Model the base. Source the burst. Bring forecast utilization, a target start date, or an existing quote. Omega Gradient will turn it into a comparable reserved and flexible-capacity plan. Start a cluster brief Browse on-demand ---------------------------------------------------------------------- # H100 & H200 Reserved GPU Clusters | Omega Gradient URL: https://omegagradient.com/clusters/hopper-h100-h200.html Published: 2026-08-17 · Last modified: 2026-11-04 Home / Reserved clusters / Hopper clusters Hopper infrastructure · buyer guide # H100 and H200 reserved GPU clusters. Hopper is a mature foundation for distributed training, post-training, inference, and HPC—but two H100 or H200 quotes can describe materially different systems. Form factor, memory, scale-up domain, scale-out fabric, storage, service boundary, and delivery evidence must be compared alongside price. Published August 17, 2026 By Tausif Ahmed , Founder & CEO Updated November 4, 2026 Reserved-cluster sourcing ## Direct answer Choose an H100 or H200 cluster by useful workload performance, memory fit, delivery confidence, and total commitment—not by accelerator name or headline GPU-hour rate alone. H200’s 141 GB HBM3e footprint can reduce memory pressure for large models and memory-bound inference, while H100 remains a broadly deployed and often easier-to-source platform. In either case, node topology, fabric, storage, and operational consistency determine whether a distributed workload scales. Omega Gradient sources and compares reserved Hopper clusters for buyers that need a defined count, region, start date, topology, or term. On-demand options remain available for benchmarking, experiments, and flexible workloads. ## H100 vs H200: start with memory behavior and workload economics NVIDIA’s DGX documentation describes eight-GPU systems with 640 GB total H100 memory or 1,128 GB total H200 memory. At the accelerator level, NVIDIA lists H100 SXM with 80 GB HBM3 and H200-based systems with 141 GB per GPU. That larger memory footprint can reduce model partitioning, enable larger batches or contexts, or keep more of an inference workload resident. It does not guarantee that every H200 workload is faster or cheaper. | Decision area | H100 | H200 | | Memory | Common H100 SXM systems provide 80 GB per GPU; H100 NVL is a different PCIe product | 141 GB HBM3e per GPU in common H200 systems | | Best reason to choose | Broad deployment, mature operating experience, potentially wider supply | Workloads that benefit from larger memory and higher memory bandwidth | | Validation focus | Exact variant, node topology, fabric, and system consistency | Measured memory benefit, software readiness, node topology, and fabric | | Commercial question | Does availability or rate outweigh the value of more memory? | Does the workload save enough nodes, time, or engineering complexity to justify the premium? | A useful comparison runs the buyer’s model or a representative benchmark under the intended precision, parallelism, sequence length, batch behavior, and storage path. The economic unit is cost per completed training objective, fine-tuning run, inference token at the target latency, or scientific result—not the GPU-hour in isolation. ## “H100” is not a complete hardware description H100 appears in SXM systems and in H100 NVL/PCIe configurations. NVIDIA lists different memory and NVLink characteristics for those products. An eight-GPU HGX or DGX-style node with NVSwitch is not interchangeable with independent PCIe GPUs connected only through host PCIe and the scale-out network. The quote should state exact accelerator variant, GPUs per node, CPU, host memory, local NVMe, NICs, NVLink/NVSwitch topology, and the network between nodes. If the seller cannot provide that information, the buyer does not yet have a cluster design to compare. ## Fabric and storage determine distributed efficiency NVIDIA cites 900 GB/s GPU-to-GPU interconnect for H100 SXM through fourth-generation NVLink, while DGX H100/H200 systems use NVSwitch inside the eight-GPU node. Multi-node scaling still depends on the external fabric. The provider should specify InfiniBand or Ethernet, generation and link speed, topology, blocking ratio, RDMA, network isolation, and the expected behavior of NCCL or other collectives. Storage should be qualified against the buyer’s data and checkpoint pattern. Important facts include usable capacity, aggregate and per-node read/write throughput, metadata performance, local scratch, object-store path, checkpoint frequency, restore behavior, ingress, and egress. A cluster can pass a GPU microbenchmark and fail the actual training pipeline because the data path is undersized. ## Normalize the whole quote Hopper is available through hyperscalers, neoclouds, private-cloud operators, bare-metal providers, scheduled-capacity products, marketplaces, and resellers. Each channel packages the system differently. Omega Gradient compares: - effective GPU-hour economics at the contracted count and billed hours; - term, ramp, deposit, prepayment, taxes, currency, and minimums; - included storage, bandwidth, public IPs, support, management, and software; - exact node and fabric design, orchestration, access, and tenant isolation; - delivery state, dependencies, acceptance window, service start, and remedies; - renewal, price reset, termination, portability, and end-of-term data handling. Public on-demand pricing is useful as a market reference, not as a substitute for a reserved quote. Omega Gradient’s daily GPU price index and provider pricing report help establish the visible market; reserved pricing still depends on count, term, start date, system, region, and commercial structure. ## Availability needs a date and an evidence status A Hopper cluster may be live, allocated to another customer until a future date, assembled from inventory across sites, or planned. The diligence record should state what is available, where, from when, in what topology, and under whose control. A provider representation can be commercially useful before full validation, but it should remain attributed. For a material deposit or long commitment, the buyer may need current system evidence, a controlled test, chain-of-control documentation, facility confirmation, or milestone-based payment and termination protection. ## H100/H200 acceptance should test the contracted system - Inventory: count, exact variant, memory, node configuration, NICs, storage, firmware, and access. - Health: component diagnostics, error history, thermals, sustained burn-in, and spare/replacement process. - Scale-up: NVLink/NVSwitch health and topology inside each node. - Scale-out: link state, bandwidth, latency, RDMA, and collective performance across the reserved block. - Storage: read/write and metadata behavior under realistic concurrency, plus checkpoint and restore. - Workload: an agreed benchmark or representative job using versioned drivers, libraries, and configuration. The contract should define pass criteria, test duration, partial failure, cure period, replacement, service-start adjustment, and the buyer remedy if the system does not match the agreed design. ## Minimum Hopper cluster brief - H100, H200, or flexible; acceptable form factor and GPUs per node; - total GPU count, region, target start window, ramp, and term; - training, post-training, inference, or HPC workload and current benchmark data; - required scale-up topology and scale-out fabric; - dataset, storage throughput, checkpointing, ingress, and egress; - bare metal, private cloud, Slurm, Kubernetes, or managed-service preference; - security, compliance, support, budget, and commercial constraints. If the team does not yet know the right platform, begin with a measured on-demand benchmark and use the result to size the reservation. Primary references NVIDIA, H100 specifications ; NVIDIA, DGX H100/H200 system architecture ; NVIDIA, H200 NVL reference architecture ; AWS, H100/H200 Capacity Blocks documentation . Verify current product, region, and availability details with the provider. Continue the Hopper decision H100 market pricing ↗ H200 market pricing ↗ Compare access models ↗ Hopper mandate ## Make the H100/H200 comparison workload-specific. Send the GPU count, region, start window, term, workload, topology, storage, and any current quote. Omega Gradient will structure the comparison and identify the evidence gaps. Start a Hopper cluster brief ---------------------------------------------------------------------- # B200, B300 & NVL72 Reserved Clusters | Omega Gradient URL: https://omegagradient.com/clusters/blackwell-b200-b300.html Published: 2026-08-17 · Last modified: 2026-11-04 Home / Reserved clusters / Blackwell clusters Blackwell infrastructure · buyer guide # B200, B300 and NVL72 reserved clusters. Blackwell procurement is a system and facility decision. Buyers need to distinguish HGX B200 or B300 nodes from Grace Blackwell NVL72 rack-scale systems, then qualify power, liquid cooling, scale-up fabric, scale-out network, software readiness, delivery dependencies, and acceptance. Published August 17, 2026 By Tausif Ahmed , Founder & CEO Updated November 4, 2026 Architecture-first sourcing ## Direct answer A request for a “B200 cluster” is incomplete. It could mean eight-GPU HGX nodes connected over a scale-out fabric, a GB200 NVL72 rack-scale system with 72 Blackwell GPUs and 36 Grace CPUs, or capacity marketed before the facility and system are ready. A buyer should specify the workload and required system boundary, then compare only options that make the exact architecture, site, delivery state, and acceptance path explicit. Omega Gradient prioritizes reserved Blackwell mandates where the buyer needs a coherent cluster, a future delivery window, or a commercial comparison across different provider structures. We do not publish private operator identities or uncontracted capacity as “available.” ## Decode the names before comparing quotes ### B200 and B300 B200 is a Blackwell data-center GPU; B300 is the Blackwell Ultra successor. Providers may offer these accelerators in HGX systems or cloud instance products. The accelerator tells you the generation, but it does not by itself define GPU count per node, host architecture, NVLink/NVSwitch domain, network, cooling, or storage. ### GB200 and GB300 NVL72 GB200 and GB300 pair Grace CPUs with Blackwell GPUs in rack-scale NVL72 systems. NVIDIA describes GB300 NVL72 as a fully liquid-cooled rack-scale architecture with 72 Blackwell Ultra GPUs and 36 Grace CPUs . That 72-GPU scale-up domain is materially different from a collection of independent eight-GPU nodes. It also carries different facility, installation, commissioning, and operational requirements. ### Cloud product names Cloud providers wrap the underlying platform in their own instance, cluster, or capacity-block products. The product can add an orchestration layer, storage, network, and commercial rules. Compare the underlying system and service boundary, not a provider SKU alone. ## Choose HGX nodes or rack-scale NVL72 from the workload | Question | HGX B200 / B300 cluster | GB200 / GB300 NVL72 | | Scale-up domain | Typically centered on the GPUs within each HGX node, then scales out across the cluster fabric | Rack-scale 72-GPU NVLink domain | | Deployment unit | Node-based; can support more incremental cluster shapes | Integrated rack-scale system with Grace CPUs, compute trays, switch trays, and liquid cooling | | Facility burden | High-density GPU infrastructure; exact cooling and power depend on system | Facility integration and liquid-cooling readiness are central to delivery | | Workload fit | Broad training, fine-tuning, inference, and HPC where node-based scaling is suitable | Large AI training and reasoning workloads that benefit from the rack-scale scale-up domain | | Diligence focus | Node consistency, scale-out fabric, storage, orchestration, and available count | Rack configuration, liquid cooling, power, NVLink domain, scale-out network, commissioning, and system software | Neither architecture is automatically better. NVL72’s integrated scale-up domain can be decisive for some model architectures and parallelism strategies. Node-based clusters can be easier to source incrementally, fit existing facilities, or align with an established Slurm or Kubernetes operating model. Benchmark the intended software stack and communication pattern. Teams planning beyond Blackwell should also review the Vera Rubin buyer guide . Vera Rubin introduces HBM4, Vera CPUs, NVLink 6, and a new facility and delivery cycle, but an announced provider deployment should not be treated as contractable capacity without buyer-specific evidence. ## Facility readiness is part of the product Blackwell supply discussions often focus on manufacturer allocation. The actual delivery chain is longer: equipment, rack integration, facility space, power, cooling, network, installation, firmware, cluster software, burn-in, and buyer acceptance. Each dependency needs an owner and a date. ### Facility questions for a material reservation - Which site will host the system, and is the quoted power allocated to that hall and deployment? - What cooling architecture serves the exact rack configuration? If liquid-cooled, who owns the CDU and facility-water responsibilities? - What are the installation, energization, network, and commissioning milestones? - Is the equipment live, on site, in transit, allocated, ordered, or dependent on a future purchase? - Which party controls the equipment and which party controls the site? - What redundancy and maintenance design applies to the buyer’s service boundary? A provider can have a legitimate allocation and still miss the desired start date because the facility is the critical path. Omega Gradient records that distinction rather than treating all future capacity as equivalent. ## Scale-up fabric, scale-out fabric, and storage all need acceptance criteria NVLink and NVSwitch describe communication inside the scale-up domain. A cluster with multiple nodes or NVL72 racks still needs a scale-out network. NVIDIA positions Quantum-X800 InfiniBand and Spectrum-X Ethernet with its GB300 NVL72 platform; the provider must specify the actual network deployed, topology, link speed, RDMA design, isolation, and expected collective behavior. Storage can be just as limiting. The brief should define dataset ingress, aggregate read throughput, metadata behavior, checkpoint frequency and size, restore path, local scratch, durable storage, and the cost and time to move data out. A provider benchmark that runs entirely in GPU memory does not validate the buyer’s end-to-end training pipeline. ## Separate live, near-term, and development capacity Omega Gradient classifies Blackwell options by delivery state: - Live and testable: system is operating and can support a scoped technical validation. - Installed or commissioning: equipment is on site but acceptance dependencies remain. - Allocated and scheduled: equipment and site path are documented, with future milestones. - Conditional development: delivery depends on financing, power, equipment, customer anchor, or another unresolved condition. - Indicative only: market signal without sufficient evidence for a procurement decision. That classification does not automatically exclude a future project. It determines the evidence, contract structure, deposit protection, milestone payments, and remedies the buyer should require. ## Blackwell cluster acceptance A suitable plan is architecture-specific. It should cover inventory and firmware, component health, sustained burn-in, thermals and power, scale-up link health, scale-out bandwidth and collective tests, storage throughput, orchestration, observability, tenant access, and a representative workload. For NVL72, acceptance should confirm the expected rack-scale system boundary rather than treating each compute tray as an unrelated node. The contract should state the test versions and conditions, acceptance window, partial-delivery treatment, cure period, component replacement process, service-start adjustment, and the remedy if the delivered topology or performance materially differs from the agreed design. ## Minimum Blackwell buyer brief - Target platform: HGX B200, HGX B300, GB200 NVL72, GB300 NVL72, or architecture-flexible; - GPU or rack count, target start window, ramp, term, and acceptable regions; - Model, framework, precision, parallelism strategy, sequence/context characteristics, and expected utilization; - Required scale-up and scale-out topology, network preference, and orchestration model; - Dataset size, ingress, storage throughput, checkpoint pattern, and egress constraints; - Security, tenancy, compliance, support, and managed-service requirements; - Budget range, deposit or prepayment constraints, and acceptance requirements. With that brief, Omega Gradient can compare credible paths instead of collecting generic “B200 available” claims. Primary references NVIDIA, GB300 NVL72 overview and specifications ; NVIDIA, GB200/GB300 data-center architecture ; NVIDIA, HGX AI Factory components ; AWS, B200 and B300 Capacity Blocks documentation . Verify current product, region, and availability details with the provider. Continue the Blackwell decision B200 market pricing ↗ B300 market pricing ↗ Vera Rubin buyer guide ↗ Blackwell mandate ## Source the system, not the label. Send the target platform, GPU or rack count, region, start window, term, workload, fabric, and storage requirements. Unknowns can be worked through in the brief. Start a Blackwell cluster brief ---------------------------------------------------------------------- # Vera Rubin: What GPU Buyers Need to Know | Omega Gradient URL: https://omegagradient.com/clusters/vera-rubin-guide.html Published: 2026-08-18 · Last modified: 2026-11-04 Home / Reserved clusters / Vera Rubin buyer guide Next-generation infrastructure · buyer guide # Vera Rubin: what GPU buyers need to know. NVIDIA now lists Vera Rubin NVL72 as available and describes the platform as in full production. That does not establish unallocated capacity at a particular provider or region. Buyers should understand the system architecture, facility requirements, performance claims, delivery evidence, and contracting dependencies before treating a Vera Rubin quote as executable supply. Published August 18, 2026 Reviewed October 26, 2026 By Tausif Ahmed , Founder & CEO Updated November 4, 2026 NVIDIA product specifications checked ## Direct answer Vera Rubin is NVIDIA’s next rack-scale AI platform after Grace Blackwell. The flagship Vera Rubin NVL72 combines 72 Rubin GPUs, 36 Vera CPUs, sixth-generation NVLink, ConnectX-9 networking, BlueField-4 DPUs, and HBM4 memory in one liquid-cooled system. The current NVIDIA product page lists the system as available; its platform page describes full production. Provider allocation, service readiness, and the terms of a specific offer still need confirmation. For a buyer, the important distinction is between a product being available, a provider announcing deployment, a dated allocation, and capacity that can be contracted with a defined start date and acceptance plan. Those are four different evidence states. ## Vera Rubin is a platform, not just a new GPU The name combines the NVIDIA Vera CPU and NVIDIA Rubin GPU, but the commercial unit is a tightly codesigned rack-scale system. NVIDIA describes Vera Rubin as a set of compute, scale-up, scale-out, networking, storage, security, and management components engineered together rather than assembled from independent parts. The core platform includes: - Rubin GPU: the accelerator, with 288 GB of HBM4 and 19.2 TB/s of memory bandwidth in NVIDIA’s current published NVL72 specification; - Vera CPU: NVIDIA’s custom Arm-compatible CPU with 88 Olympus cores per superchip configuration; - NVLink 6: the scale-up fabric connecting the 72-GPU rack-scale domain; - ConnectX-9: high-bandwidth scale-out network interfaces; - BlueField-4: infrastructure processing for networking, storage, and security; - Spectrum-6 and Quantum-X800: Ethernet and InfiniBand paths for scaling beyond one rack; - Mission Control: NVIDIA’s management layer for configuration, facilities integration, and cluster operations. This matters because a quote for “Rubin GPUs” is not enough. The buyer needs to know whether the offer is Vera Rubin NVL72, Vera Rubin NVL4, another Rubin system configuration, or future cloud instances abstracting the underlying hardware. ## The headline Vera Rubin NVL72 specifications The table follows NVIDIA’s product page and linked datasheet as reviewed on October 26, 2026. These values describe the published NVL72 configuration; confirm the provider’s exact bill of materials, bandwidth definitions, and operating conditions before contracting. Performance projections have their own assumptions and qualifications. | System area | Vera Rubin NVL72 published specification | Buyer implication | | Compute | 72 Rubin GPUs and 36 Vera CPUs | The procurement unit is an integrated rack-scale system, not a loose pool of cards. | | GPU memory | 20.7 TB HBM4 total; 288 GB per Rubin GPU | Larger models, longer contexts, and memory-heavy inference may require less partitioning. | | Memory bandwidth | 1,400 TB/s across the rack; 19.2 TB/s per GPU | Potentially material for memory-bound training and inference, but workload measurement still matters. | | Scale-up | 216 TB/s NVLink 6 bandwidth across the NVL72 system | Parallelism strategy should be designed around the 72-GPU scale-up domain. | | CPU | 3,168 Olympus cores and up to 54 TB LPDDR5X across 36 Vera CPUs | Agent orchestration, data preparation, simulation, and CPU-heavy control paths are first-order parts of the system. | | Scale-out | 32.4 TB/s system networking bandwidth, bidirectional; ConnectX-9 and BlueField-4 interfaces | Multi-rack design still requires a specified fabric, topology, congestion plan, and acceptance test. | | Operations | Mission Control, NVIDIA AI Enterprise, and DGX OS in the DGX configuration | The provider’s actual software and managed-service boundary must match the buyer’s operating model. | Specification review · October 26, 2026. The current NVIDIA product specification and linked Vera Rubin datasheet, page 8 are the source version for this table. This review replaces the earlier memory-bandwidth figures of 1,580 TB/s per rack and 22 TB/s per GPU, 260 TB/s NVLink, and 28.8 TB/s scale-out bandwidth. NVIDIA’s older launch material may retain different figures. ## What changes from Grace Blackwell NVL72 The clearest architectural changes are HBM4 memory, the Vera CPU, sixth-generation NVLink, new network interfaces, and a platform designed around agentic AI, mixture-of-experts models, long-context reasoning, and more CPU-intensive orchestration. NVIDIA and CoreWeave report that an early DeepSeek-R1 benchmark delivered ten times more throughput per megawatt than Grace Blackwell NVL72. That is an important launch benchmark, but it is not a universal workload guarantee. It was produced on a specific model, software stack, configuration, and early provider system. A buyer should ask what portion of the improvement comes from accelerator throughput, memory, precision, CPU orchestration, software maturity, batching, or power-envelope differences—and then benchmark the intended workload. Vera Rubin does not make Blackwell obsolete on arrival. An available and accepted B200, B300, GB200, or GB300 cluster can be commercially superior to a future Rubin system if it starts earlier, fits the workload, has a proven operating stack, or avoids the risk and premium of a launch-generation deployment. ## Product availability is not a buyer allocation NVIDIA’s current availability statement concerns its product and supply ecosystem. It does not establish that a specific provider has unallocated capacity in a buyer’s preferred region, count, date, term, or service model. Confirm the actual block and delivery state rather than treating a product announcement as a reservation. Omega Gradient would classify a Vera Rubin opportunity using the same evidence ladder applied to other reserved clusters: - Live and testable: the provider can identify the system, site, available block, and a scoped validation path. - Installed or commissioning: hardware is on site, but network, software, burn-in, or acceptance remains. - Allocated and scheduled: equipment and facility milestones support a documented future date. - Conditional deployment: capacity depends on another customer, financing, equipment delivery, power, cooling, or facility work. - Announcement only: a roadmap or partnership exists, but no buyer-specific allocation has been evidenced. A provider announcement can be credible and still sit in the fifth category for a particular buyer. ## Who should evaluate Vera Rubin first? Vera Rubin is most relevant when the workload can exploit more than a simple generational speed increase: - large mixture-of-experts training or inference with demanding communication patterns; - long-context and reasoning workloads constrained by memory capacity or bandwidth; - agentic systems with substantial CPU orchestration, tool execution, sandboxing, or state management; - power-constrained deployments optimizing useful tokens per megawatt; - large scientific and HPC programs needing native FP64 paths or tightly integrated AI and simulation; - multi-rack AI factories willing to adopt a new system and software generation at scale. Smaller fine-tuning jobs, conventional inference, early-stage experiments, and teams without stable utilization may be better served by mature Hopper or Blackwell capacity. The operational and commercial premium of a launch platform has to create measurable workload value. ## The facility is part of the Vera Rubin product Vera Rubin NVL72 is a liquid-cooled rack-scale system. A credible delivery plan needs more than a manufacturer allocation. It needs a site with appropriate power, cooling, water-loop responsibilities, rack integration, network, security, commissioning, spares, software, and staff. Before accepting a future start date, buyers should ask: - Which facility and hall will host the system? - Is power committed to the quoted racks, and what milestone makes it usable? - Which party owns the CDU, facility-water loop, and cooling remediation obligations? - Is the scale-out network Quantum-X800 InfiniBand, Spectrum-X Ethernet, or another design? - What is the blocking ratio, oversubscription boundary, and multi-rack topology? - When do installation, energization, networking, commissioning, burn-in, and acceptance occur? - Who controls the equipment, who operates the cloud, and who signs the buyer contract? ## A Vera Rubin procurement checklist - Name the exact system. NVL72, NVL4, cloud instance, or another configuration; final GPU, CPU, memory, NIC, DPU, storage, and software specification. - Define the capacity. Rack count, GPU count, region, start window, ramp, term, tenancy, and whether the block is exclusive. - Map the delivery chain. Manufacturer or OEM, equipment owner, facility operator, cloud operator, reseller, contracting entity, and support owner. - Verify the facility path. Power, liquid cooling, network, installation, commissioning, and unresolved dependencies. - Confirm software readiness. Drivers, CUDA, frameworks, orchestration, Mission Control responsibility, observability, and workload compatibility. - Normalize the economics. GPU-hour or rack economics, deposit, prepayment, storage, bandwidth, support, software, ramp, renewal, and termination. - Write the acceptance plan. Inventory, firmware, burn-in, NVLink, scale-out fabric, storage, representative workload, pass criteria, cure period, and remedies. - Protect the delivery date. Milestone payments, refund or termination rights, partial delivery, substitution rules, and what happens if final specifications change. ## Should buyers reserve Vera Rubin now or wait? Begin sourcing now if the program has a known 2026–27 start window, the workload can plausibly benefit from the architecture, and early access is strategically valuable. The purpose of an early process is to learn the real provider, facility, allocation, and commercial landscape—not to treat every announcement as supply. Wait or benchmark first if utilization is uncertain, software compatibility is untested, a mature Blackwell system meets the requirement, or the commercial case depends on unverified performance assumptions. An H200 or Blackwell benchmark can establish the workload baseline that a future Rubin proposal must beat. The procurement question is not “Is Vera Rubin faster?” It is “Will this specific Vera Rubin system deliver more useful work per dollar, megawatt, and month of deployment risk for our workload?” Omega Gradient can structure a Vera Rubin market scan while keeping Hopper, Blackwell, and scheduled or on-demand alternatives in the same decision frame. Primary references NVIDIA, Vera Rubin NVL72 architecture and published specifications ; NVIDIA, DGX Vera Rubin NVL72 ; NVIDIA, current Vera Rubin platform and production status ; NVIDIA, July 2026 production-ramp and partner update ; NVIDIA, DGX SuperPOD and Rubin platform announcement ; NVIDIA Newsroom, Vera Rubin for scientific computing . Source review: October 26, 2026. The current NVL72 product page and its linked datasheet supersede the earlier bandwidth figures in this guide. Confirm the provider configuration and preserve any benchmark-specific qualifications before procurement. Build the comparison set B200 & B300 guide ↗ How we qualify capacity ↗ Compare access models ↗ Vera Rubin mandate ## Separate roadmap access from contractable capacity. Send the target rack or GPU count, region, start window, term, workload, software baseline, and any provider announcement or quote already under review. Start a Vera Rubin market scan ---------------------------------------------------------------------- # GPU Cluster Procurement Checklist | Omega Gradient URL: https://omegagradient.com/clusters/gpu-cluster-procurement-checklist.html Published: 2026-08-21 · Last modified: 2026-11-04 Home / Reserved clusters / Procurement checklist Buyer field guide · 36 diligence questions # The reserved GPU cluster procurement checklist. A quote is not a cluster. Use this checklist to turn a GPU count and headline rate into a decision-ready view of the system, the delivery chain, the total obligation, and the evidence supporting it. Published August 21, 2026 By Tausif Ahmed , Founder & CEO Updated November 4, 2026 Buyer-side framework v1.0 ## Direct answer Before signing a reserved GPU cluster agreement, a buyer should be able to answer four questions without relying on inference: What exact system is being delivered? Who controls the equipment and facility? What evidence supports the start date and performance? What contractual remedy applies when delivery, acceptance, uptime, or support misses the promise? A defensible procurement process covers nine connected areas: workload requirements, capacity provenance, facility readiness, node and fabric architecture, storage and software, operations, normalized economics, acceptance, and the final decision record. Missing facts do not automatically disqualify a provider, but they must be labeled as open, conditional, or represented—not silently upgraded into validated capacity. Working method: print or save this page, mark every item as validated, represented, conditional, or open, and attach the evidence location. Re-score whenever the provider, facility, delivery date, or bill of materials changes. Print / save PDF ## Use a four-state evidence model Binary “yes/no” diligence hides too much. A provider may give a credible answer that is not yet evidenced, or show an allocation that still depends on facility work. The following status model preserves that difference throughout sourcing and negotiation. Validated Primary evidence supports the claim and matches the offered system. Represented The accountable party states it, but documentary or technical proof is pending. Conditional The outcome depends on a named milestone, third party, or unresolved dependency. Open The answer, owner, evidence, or remedy has not been established. ### 1. Workload and system definition Start with the useful work. A provider cannot design or price the right cluster when the brief stops at a GPU model and count. - What workload must the cluster run? Record training, post-training, inference, rendering, simulation, or HPC; model scale; frameworks; precision; expected concurrency; and data behavior. - What result defines success? Establish target time-to-train, tokens per second, latency, throughput, checkpoint time, or another workload-level measure. - Which requirements are hard constraints? Separate mandatory GPU, memory, region, security, topology, start date, and term from preferences that can trade against price or availability. - What utilization will actually be sustained? Model ramp, idle periods, maintenance, data preparation, experimentation, and demand variability before comparing reservation economics with on-demand or blended access . ### 2. Capacity control and delivery evidence The commercial seller, equipment owner, facility operator, cloud operator, and support team may be different entities. Map the chain before relying on an availability claim. - Who is the contracting entity? Confirm the legal name, jurisdiction, authority to sell the service, and the balance sheet standing behind refunds, credits, and damages. - Who owns or controls the equipment? Ask whether the seller owns it, leases it, has a binding allocation, resells another operator, or is still assembling supply. - What is the capacity state today? Classify the block as live and testable, installed, commissioning, allocated, ordered, financed, forecast, or dependent on another customer. - What evidence matches this exact block? Request redacted serial inventory, rack or cluster identifiers, purchase or allocation support, facility confirmation, commissioning records, or a scoped test path—not a generic screenshot. ### 3. Facility and delivery readiness An equipment allocation is not usable capacity until the site can power, cool, network, secure, commission, and operate the promised system. - Where will the cluster run? Identify country, metro, facility, hall or deployment boundary, data-sovereignty implications, and whether relocation is permitted. - Is power committed to these racks? Distinguish contracted utility capacity, energized capacity, reserved distribution, and power that still depends on construction or interconnection. - Is the cooling design ready? For liquid-cooled systems, assign responsibility for CDUs, facility-water loops, water quality, leak response, monitoring, and remediation. - What milestones lead to service? Put equipment arrival, installation, energization, network turn-up, burn-in, customer access, acceptance, and production start on one dated critical path. ### 4. Node, scale-up, and scale-out architecture “H100,” “B200,” or “Rubin” is not a system specification. The useful cluster is defined by the entire communication path around the accelerator. - What is the exact node or rack-scale bill of materials? Record accelerator form factor, GPUs per node or rack, GPU memory, CPU, host memory, local NVMe, NICs, DPUs, firmware, and OEM. - What is the scale-up domain? Confirm NVLink and NVSwitch generation, bandwidth, GPU boundaries, and whether the offered partition preserves the topology assumed by the workload. - What is the scale-out fabric? Specify InfiniBand or Ethernet generation, link speed per GPU or node, topology, rail design, RDMA configuration, adaptive routing, and congestion controls. - Where is the oversubscription boundary? Ask for blocking ratio and topology diagrams across compute, storage, management, and external connectivity; “non-blocking” should name the measured boundary. ### 5. Storage, security, and software platform GPU utilization often fails outside the GPU. Storage, identity, orchestration, observability, and image support need measurable service boundaries. - Can storage feed the workload? Define usable capacity, aggregate and per-client throughput, metadata performance, checkpoint behavior, local scratch, durability, backup, and restore. - How does data enter and leave? Document network paths, private connectivity, transfer windows, bandwidth caps, egress pricing, encryption, deletion, and end-of-term export. - What isolation and access model applies? Confirm single tenancy or partitioning, identity provider integration, roles, audit logs, key management, vulnerability process, remote-hands controls, and compliance evidence. - Who owns software compatibility? Freeze or bound driver, firmware, CUDA, framework, scheduler, container, image, observability, and upgrade responsibilities; define who cures regressions. ### 6. Operations, support, and service level A large cluster is an operating relationship. The buyer needs named response paths and measurable responsibilities, not only an uptime percentage. - What does the provider operate? Draw the boundary across facility, hardware, network, storage, base software, scheduler, Kubernetes or Slurm, images, security, and application support. - How are failed components handled? Set spares strategy, GPU and node replacement targets, drain procedures, maintenance coordination, and the treatment of chronically degraded hardware. - What is measured for the SLA? Define availability denominator, excluded events, planned maintenance, degraded performance, network and storage failures, reporting source, and dispute process. - Who responds at 2:00 AM? Establish severity definitions, 24/7 channels, acknowledgement and restoration targets, escalation names, incident reports, and executive review cadence. ### 7. Commercial normalization and total obligation Normalize the whole commitment. Two identical GPU-hour rates can produce different effective costs once minimums, deposits, ramp, support, and data movement are included. - What exactly is the billing unit? Confirm whether the quote is per GPU, node, rack, cluster, reserved block, or consumed hour; define calendar hours, partial availability, and rounding. - What cash moves before acceptance? Record deposit, prepayment, milestone payment, letter of credit, security, refund conditions, and the party holding buyer funds. - Which costs sit outside the headline rate? Include storage, bandwidth, egress, IP addresses, support, licenses, remote hands, setup, taxes, power adjustment, and cross-connects. - What changes during the term? Model ramp, take-or-pay, usage minimums, renewal, indexation, hardware substitution, expansion, contraction, assignment, and early termination. ### 8. Acceptance, cure, and remedies Acceptance is where sales language becomes an operational obligation. The test plan should be agreed before delivery and attached to the contract. - What converts delivery into acceptance? Require inventory and configuration verification, burn-in, GPU health, NVLink, fabric, storage, security, scheduler, and access tests. - Does the test reflect the workload? Add representative collectives, framework tests, checkpointing, or application benchmarks with dataset, configuration, duration, and reproducible pass criteria. - What happens when only part passes? Define partial acceptance, replacement, retest, cure period, billing start, schedule relief, and whether a smaller topology can be forced on the buyer. - What remedy matches each failure? Separate late delivery, failed acceptance, chronic degradation, downtime, security events, specification changes, and abandonment; pair each with refund, credit, replacement, or termination rights. ### 9. Decision record and red flags The final memo should make the trade-off legible to technical, finance, security, and legal reviewers—and preserve what was known when the commitment was approved. - Which facts are still unresolved? Carry every open and conditional item into the approval record with an owner, evidence request, deadline, and consequence. - What is the credible alternative? Compare the preferred reservation with a second provider, a mature GPU generation, delayed start, smaller base plus burst, and on-demand capacity . - What would cause a no-go? Predefine red lines such as unclear equipment control, unverifiable site, payment before evidence, refusal to attach specifications, no acceptance rights, or remedies that cannot be collected. - Who signs each risk? Require explicit approval from the owners of technical architecture, security, finance, legal, operations, and the business workload—not a blended “team approved” status. ## How to compare shortlisted clusters Do not reduce the scorecard to one total. A high aggregate score can hide a fatal dependency. First apply hard gates: legal seller identified, capacity path evidenced, site and architecture defined, economics normalized, acceptance attached, and remedies enforceable. Then compare the surviving options across workload performance, delivery confidence, operating model, total cost, flexibility, and counterparty exposure. Red flags deserve asymmetric weight. A provider may have an excellent technical design but still be unbuyable if it cannot demonstrate control of the capacity. Another may control live equipment but fail the workload because the fabric or storage boundary is wrong. The purpose of diligence is not to make every option look comparable; it is to reveal when they are not. A reserved cluster is ready to buy when the system, evidence, obligations, and remedies describe the same capacity. ## What should be attached to the contract? At minimum, attach the final bill of materials, topology, facility region, capacity and tenancy definition, delivery milestones, acceptance plan, service description, support matrix, SLA measurement method, pricing schedule, excluded charges, data return or deletion procedure, security responsibilities, substitution rules, and remedies. If a material diligence answer lives only in email or a sales deck, convert it into the agreement or explicitly accept that it is not an obligation. Omega Gradient uses this framework to structure demand briefs, qualify providers, normalize offers, and produce an evidence-marked shortlist. For the underlying method, read how Omega Gradient qualifies reserved GPU clusters . For platform-specific considerations, see the H100 and H200 guide , Blackwell guide , and Vera Rubin buyer guide . Scope note This is a buyer-side operational framework, not legal, engineering, security, tax, or financial advice. Adapt the questions, evidence threshold, tests, and contract language to the workload, jurisdiction, risk level, and provider structure. Procurement mandate ## Turn the checklist into a live market scan. Send the GPU platform, count, region, start window, term, workload, and any quotes already under review. Omega Gradient can shape the brief, identify relevant capacity paths, and mark the evidence and open risks behind each option. Start a reserved cluster brief ---------------------------------------------------------------------- # Reserved GPU Market Report — Q3 2026 | Omega Gradient URL: https://omegagradient.com/research/reserved-gpu-market-report-2026-q3.html Published: 2026-08-21 · Last modified: 2026-08-21 Home / Research / Market report Q3 2026 early read · 50 eligible observations # Reserved GPU market report. A sample-safe view of represented reserved supply: which platforms are showing up, how large the offered blocks are, when capacity is expected, and where the evidence record is still incomplete. Published August 21, 2026 Data cut August 21, 2026 Fixed edition · v1.0 By Tausif Ahmed , Founder & CEO ## Executive read Omega Gradient’s Q3 2026 early read shows a forward-heavy, Blackwell-heavy represented supply market in which evidence maturity often trails the headline availability claim. Across 50 eligible observations, 70% referenced Blackwell-family platforms, 54% pointed to Q4 2026 start windows, 18% were represented as available now, and only 36% explicitly recorded a network or fabric field. The practical implication is simple: a GPU model, count, price, and target date are opening metadata—not a decision-ready cluster. Buyers should refresh old representations, require the fabric and service boundary, normalize the full commitment, and tie forward delivery to dated evidence and acceptance milestones. 70% Blackwell-family 35 of 50 eligible observations 54% Q4 start window 27 represented paths 18% Available now 9 represented paths; not independently audited 36% Fabric stated 18 observations recorded a network field Read this first An observation is a tracker row, not a unique provider, facility, physical system, or audited unit of available capacity. Multiple observations can refer to alternative configurations or the same underlying supply path. For that reason, this report does not sum GPUs or present the sample as market share. ## Method and sample The source contained 128 internal tracker rows. We included 50 supply-side observations with a numeric GPU count, a valid verification date, and a status of active, represented available now, build-to-order, or outside a near-term buyer window. We excluded demand-side records, sold or dropped paths, watch and unconfirmed rows, and records that could not support the required fields. July 8, 2026 Earliest eligible verification August 21, 2026 Report cutoff n ≥ 5 Minimum public segment 0 names Providers and buyers disclosed Every percentage below uses row-level observations as its denominator. Categories smaller than the public threshold are suppressed or combined. Price fields were reviewed only for completeness because their units, inclusions, terms, fabric assumptions, and payment structures were not sufficiently normalized for a responsible public median. ## Blackwell leads the observed platform mix Blackwell-family systems appeared in 35 of 50 eligible observations (70%). Hopper-family systems appeared in 11 (22%). Smaller platform groups were suppressed to avoid over-reading thin segments. Blackwell 35 observations · 70% Hopper 11 observations · 22% Other / mixed Suppressed (n < 5) | Observed segment | Observations | Share | | Blackwell | 35 | 70% | | Hopper | 11 | 22% | | Other / mixed | n < 5 | Suppressed | | Total eligible sample | 50 | 100% | This is a mix of represented sourcing paths, not a shipment forecast or installed-base estimate. The signal is directionally useful: buyers comparing new reservations are increasingly likely to encounter Blackwell and Blackwell Ultra proposals, which makes rack, power, cooling, scale-up domain, and scale-out fabric part of the commercial comparison—not technical details to defer until after award. ## Most represented starts sit in Q4 27 observations (54%) pointed to Q4 2026. Only 9 (18%) were represented as available now, while the remainder sat in nearer forward windows or lacked a dated start. A target date should therefore be decomposed into equipment, facility, commissioning, access, and acceptance milestones. Represented available now 9 observations · 18% By September 2026 7 observations · 14% Q4 2026 27 observations · 54% 2027 or later 7 observations · 14% | Observed segment | Observations | Share | | Represented available now | 9 | 18% | | By September 2026 | 7 | 14% | | Q4 2026 | 27 | 54% | | 2027 or later | 7 | 14% | | Total eligible sample | 50 | 100% | “Available now” is the status recorded in the sourcing tracker, not independent certification. Buyers should still verify the exact block, tenancy boundary, software state, test path, and contracting chain. ## The observation set skews toward large blocks The largest band—1,024 GPUs or more—accounted for 22 observations (44%). That concentration does not mean those GPUs should be added together or treated as simultaneously available. It shows that many represented paths are being discussed as major programs rather than small, immediately testable allocations. 1–255 GPUs 13 observations · 26% 256–1,023 GPUs 15 observations · 30% 1,024+ GPUs 22 observations · 44% | Observed segment | Observations | Share | | 1–255 GPUs | 13 | 26% | | 256–1,023 GPUs | 15 | 30% | | 1,024+ GPUs | 22 | 44% | | Total eligible sample | 50 | 100% | For a buyer, a larger represented block increases the importance of chain-of-control evidence, phased acceptance, topology preservation, facility milestones, replacement capacity, and remedies that remain collectible if the full program does not arrive on schedule. ## Evidence freshness and quote completeness are uneven 22 observations (44%) were verified within 14 days of the cutoff. 17 (34%) were more than 30 days old. A stale representation may still be useful as a lead, but it should not survive into a shortlist without a dated refresh. Verified 0–14 days before cutoff 22 observations · 44% Verified 15–30 days before cutoff 11 observations · 22% Verified 31+ days before cutoff 17 observations · 34% | Observed segment | Observations | Share | | Verified 0–14 days before cutoff | 22 | 44% | | Verified 15–30 days before cutoff | 11 | 22% | | Verified 31+ days before cutoff | 17 | 34% | | Total eligible sample | 50 | 100% | ### Field completeness exposes the comparison risk Headline pricing was present in almost every eligible row, but the terms needed to interpret that number were not. A term field appeared in 70% of observations, an upfront-payment or deposit field in 52%, and network or fabric detail in only 36%. Headline price field stated 49 observations · 98% Term field stated 35 observations · 70% Upfront / deposit field stated 26 observations · 52% Network / fabric field stated 18 observations · 36% Dated availability field 50 observations · 100% This is why a public “average reserved GPU price” would be misleading. Rates can differ in GPU generation, topology, networking, storage, support, start state, term, prepayment, and what happens before acceptance. The unit must be normalized before the price can be compared. ## What buyers should do with this signal - Refresh the representation. Reconfirm SKU, count, location, start, block state, and accountable seller whenever the supporting evidence is more than 14 days old or a milestone has moved. - Make fabric a gate. Do not shortlist a large training or post-training cluster without the scale-up domain, scale-out fabric, link speeds, topology, oversubscription boundary, and storage path. - Normalize the full obligation. Compare rate, term, deposit, payment timing, included services, data movement, support, maintenance, ramp, and remedies on one sheet. - Turn forward dates into milestones. Separate equipment allocation, site readiness, installation, commissioning, buyer access, acceptance, and production start. - Accept in phases. For large blocks, attach reproducible health, collective, storage, and workload tests; define partial acceptance, cure, billing start, and termination rights before signing. Use the 36-point procurement checklist to structure the evidence request, and read how Omega Gradient qualifies reserved clusters for the full delivery and counterparty framework. ## Limitations - This is a partial-quarter early read covering eligible verifications from July 8, 2026 through August 21, 2026; it is not a complete census of the reserved GPU market. - The sample is shaped by Omega Gradient’s sourcing workflow and may overrepresent platforms, regions, block sizes, or counterparties relevant to active mandates. - One tracker observation is not necessarily one unique provider, one facility, or one unique physical cluster. - Statuses and dates are represented states as of the recorded verification date. They can change after the cutoff. - No raw price distribution is published because the commercial units and service inclusions were not consistently normalized. - Provider identities, buyer identities, private quotes, contacts, and row-level capacity details are intentionally withheld. ## Download the aggregate The machine-readable report includes the denominators, inclusion rules, minimum cell threshold, and the exact aggregate values used on this page. It contains no provider, buyer, contact, quote, or row-level capacity data. Download aggregate JSON Reference register Platform-family labeling was checked against NVIDIA’s DGX H100/H200 guide , HGX B200 product summary , and GB300 NVL72 reference architecture . These sources establish platform context; all market percentages are Omega Gradient tracker aggregates. Corrections policy This is a fixed edition. Material corrections will be dated and described here; later market changes will be published as a new edition rather than silently rewriting this sample. Reserved cluster mandate ## Turn the market signal into a qualified shortlist. Send the platform, count, region, start window, term, workload, topology, and any quotes already under review. Omega Gradient can scan relevant capacity paths, normalize the offers, and mark the evidence and open risks behind each option. Start a reserved cluster brief ---------------------------------------------------------------------- # Cloud GPU Price Index 2026: Sampled Public Rates | Omega Gradient URL: https://omegagradient.com/gpus/ Report assembled: 2026-11-21T17:02:24.332Z Omega Gradient · Market Data # Cloud GPU Price Index (2026): sampled public rates Normalized USD per GPU-hour pricing across 15 GPU families, computed from 261 evidence-verified public rate rows spanning a registry of 224 monitored providers . Flagship sample: H100 has a min included rate of $1.68 to a max of $11.12 per GPU-hour, median provider average $3.09 across 9 providers. 136 listings were listed in the dated catalog sample, from $0.07/GPU-hr. Report assembled 2026-11-21T17:02:24.332Z Source observations: 2026-11-02T00:00:00.000Z to 2026-11-21T17:02:03.771Z. Report assembled 2026-11-21T17:02:24.332Z; assembly does not refresh source evidence. Catalog snapshot 2026-11-21T17:00:36.649Z Dated catalog sample. Availability and prices may have changed. Check the current marketplace before relying on a listing. GPU families tracked 15 with verified public rates Providers monitored 224 in the provider registry Verified price rows 261 55 provider-GPU pairs Snapshot listings 136 sampled from $0.07/GPU-hr ## Price index — USD per GPU-hour | GPU | Min | Median provider avg | Max | Median $/mo | Providers | Price rows | Snapshot listings | | A100 | $0.48 | $2.18 | $7.12 | $1570 | 8 | 62 | 22 | | A4000 | $0.76 | $0.76 | $0.76 | $547 | 1 | 1 | 4 | | A6000 | $0.44 | $0.70 | $0.75 | $504 | 3 | 3 | 7 | | B200 | $6.56 | $8.44 | $10.00 | $6078 | 1 | 9 | 6 | | B300 | $6.64 | $9.74 | $17.80 | $7016 | 10 | 15 | 2 | | GB200 | $9.38 | $9.38 | $9.39 | $6754 | 1 | 3 | 3 | | GB300 | $15.01 | $16.50 | $18.00 | $11884 | 2 | 2 | 0 | | H100 | $1.68 | $3.09 | $11.12 | $2225 | 9 | 19 | 11 | | H200 | $2.60 | $4.23 | $6.83 | $3045 | 4 | 50 | 16 | | L40 | $0.70 | $0.95 | $1.25 | $684 | 3 | 3 | 3 | | L40S | $0.47 | $1.20 | $3.51 | $864 | 5 | 17 | 7 | | RTX 4090 | $0.35 | $0.43 | $0.74 | $307 | 4 | 8 | 2 | | RTX 5090 | $0.42 | $0.73 | $1.16 | $528 | 2 | 63 | 1 | | RTX 6000 | $0.87 | $1.14 | $1.47 | $821 | 1 | 5 | — | | T4 | $0.81 | $0.81 | $0.81 | $583 | 1 | 1 | 0 | Min and max are the lowest and highest individual included rates. The median is the median of provider averages per GPU family. Median $/mo is the monthly cost math for one GPU running a full 720-hour month at the median rate (multiply by GPU count for multi-GPU nodes). Catalog counts are anonymous listings observed in the snapshot; confirm current availability on /compute.html . Snapshot 2026-11-21. ## Methodology USD per GPU-hour; instance-hour prices are divided by gpu_count when gpu_count is known. PRICING BASIS Provider-normalized average: average SKUs/offers within each provider and GPU family first, then average provider means so one marketplace does not dominate. AVERAGING METHOD Every included price row must include an evidence URL, pass the evidence-confidence threshold, use a supported GPU-hour normalization path, and stay within a GPU-family sanity envelope before it is included in averages. VERIFICATION METHOD Candidate prices outside GPU-family USD/GPU-hour sanity floors or ceilings are quarantined from averages and the public feed until manually verified. PRICE SANITY METHOD Choose the next path Reserved clusters ↗ How we qualify ↗ On-demand marketplace ↗ ## Frequently asked questions ### How often is the Omega Gradient GPU price index updated? Collection is scheduled daily, but publication and individual source observations can differ. This report was assembled 2026-11-21. Use the source observation dates and check the marketplace for current listings. ### How many GPU providers does Omega Gradient monitor? 224 providers are in the monitored registry. This index shows the 261 price rows across 55 provider-GPU pairs that passed evidence verification and sanity checks on 2026-11-21. ### What was the cheapest GPU in this catalog snapshot? In the 2026-11-21 catalog snapshot, the lowest-priced listing is a GTX 1070 at $0.07/GPU-hour, out of 136 sampled listings. ### How are GPU prices normalized? USD per GPU-hour; instance-hour prices are divided by gpu_count when gpu_count is known. ---------------------------------------------------------------------- # H100 Cloud Pricing 2026: $1.68–$11.12/hr Across 9 Providers | Omega Gradient URL: https://omegagradient.com/gpus/h100.html Report assembled: 2026-11-21T17:02:24.332Z Cloud GPU Price Index · H100 # H100 Cloud Pricing 2026: $1.68–$11.12/hr Across 9 Providers The sampled NVIDIA H100 rates have a minimum included rate of $1.68 , a median provider-average rate of $3.09 , and a maximum included rate of $11.12 per GPU-hour across 9 cloud providers with verified public rates (19 price rows, report assembled 2026-11-21). 11 H100 listings observed in the catalog from $2.87/GPU-hr, updated 2026-11-21. Report assembled 2026-11-21T17:02:24.332Z Source observations: 2026-11-21T17:02:03.771Z. Report assembled 2026-11-21T17:02:24.332Z; assembly does not refresh source evidence. Catalog snapshot 2026-11-21T17:00:36.649Z Dated catalog sample. Availability and prices may have changed. Check the current marketplace before relying on a listing. Min included rate $1.68 per GPU-hour Median provider average $3.09 per GPU-hour Max included rate $11.12 per GPU-hour Snapshot listings 11 from $2.87/GPU-hr ## Named-provider market rates — H100 Publicly listed on-demand rates, normalized to USD per GPU-hour. These are market-rate context from provider price pages, not Omega Gradient marketplace listings. | Provider | Listed SKU | Min $/GPU-hr | Avg $/GPU-hr | Max $/GPU-hr | Source / observation dates | | Latitude.sh | 1 x NVIDIA H100 80GB | $1.68 | $1.68 | $1.68 | latitude.sh Source observations: 2026-11-21T17:02:03.771Z | | TensorDock | H100 SXM5 | $2.25 | $2.25 | $2.25 | tensordock.com Source observations: 2026-11-21T17:02:03.771Z | | Polaris Cloud | H100 | $2.28 | $2.28 | $2.28 | polaris.computer Source observations: 2026-11-21T17:02:03.771Z | | GMI Cloud | H100 | $2.60 | $2.60 | $2.60 | gmicloud.ai Source observations: 2026-11-21T17:02:03.771Z | | Paperspace | NVIDIA H100 | $3.09 | $3.09 | $3.09 | paperspace.com Source observations: 2026-11-21T17:02:03.771Z | | Vast.ai | H100 PCIE, H100 SXM, H100 NVL | $2.55 | $3.58 | $4.77 | console.vast.ai Source observations: 2026-11-21T17:02:03.771Z | | Replicate | Nvidia H100 | $5.49 | $5.49 | $5.49 | replicate.com Source observations: 2026-11-21T17:02:03.771Z | | CoreWeave | NVIDIA HGX H100 | $6.16 | $6.16 | $6.16 | coreweave.com Source observations: 2026-11-21T17:02:03.771Z | | CivoCloud | H100 | $11.12 | $11.12 | $11.12 | civo.com Source observations: 2026-11-21T17:02:03.771Z | Public list prices reflect the source observation dates above; full report at /gpu-supplier-pricing.html . ## Marketplace catalog snapshot 11 H100 listings observed in the catalog from $2.87/GPU-hr, updated 2026-11-21. Marketplace listings are provider-anonymous by design. | GPU | Count | vRAM | Region | $/GPU/hr | Instance $/hr | | | H100 | 1× | 80 GB | Africa | $2.87 | $2.87 | Check availability | | H100 | 1× | 80 GB | Europe · West | $3.32 | $3.32 | Check availability | | H100 | 2× | 80 GB | Europe · West | $3.63 | $7.27 | Check availability | | H100 | 1× | 80 GB | Europe · East | $3.64 | $3.64 | Check availability | | H100 | 4× | 80 GB | Asia · Northeast | $3.67 | $14.69 | Check availability | | H100 | 2× | 80 GB | Asia · Northeast | $3.68 | $7.36 | Check availability | | H100 | 1× | 94 GB | US · Central | $3.76 | $3.76 | Check availability | | H100 | 1× | 94 GB | Asia · Northeast | $4.02 | $4.02 | Check availability | | H100 | 2× | 80 GB | Europe · East | $5.04 | $10.09 | Check availability | | H100 | 1× | 80 GB | Global · flexible | $5.97 | $5.97 | Check availability | | H100 | 1× | 80 GB | US · East | $6.59 | $6.59 | Check availability | Snapshot 2026-11-21; current view on /compute.html . ## Monthly cost math (720 hours) | Scenario | $/GPU-hr | × 720 hrs | | Lowest tracked public rate | $1.68 | $1210 | | Median provider-average rate | $3.09 | $2225 | | Highest tracked public rate | $11.12 | $8006 | | Cheapest sampled marketplace listing | $2.87 | $2066 | 720 hours ≈ one month of continuous single-GPU runtime. Multiply by GPU count for multi-GPU nodes. Put the data to work H100 & H200 ↗ Reserved clusters ↗ On-demand marketplace ↗ ## Frequently asked questions ### How much does a H100 cost per hour in 2026? Public on-demand rates tracked by Omega Gradient range from $1.68 to $11.12 per GPU-hour across 9 providers, with a median provider-average rate of $3.09 (report assembled 2026-11-21). ### Was a H100 listed in the catalog snapshot? 11 H100 listings observed in the catalog from $2.87/GPU-hr, updated 2026-11-21. Check current availability at omegagradient.com/compute.html. ### What does a H100 cost per month? At the median tracked rate of $3.09/GPU-hr, one GPU running a full 720-hour month costs about $2225. At the lowest tracked rate of $1.68/GPU-hr, about $1210. ### Where do these prices come from? Included public rate observations collected by Omega Gradient; dates can vary by source. Provider-normalized average: average SKUs/offers within each provider and GPU family first, then average provider means so one marketplace does not dominate. ---------------------------------------------------------------------- # H200 Cloud Pricing 2026: $2.60–$6.83/hr Across 4 Providers | Omega Gradient URL: https://omegagradient.com/gpus/h200.html Report assembled: 2026-11-21T17:02:24.332Z Cloud GPU Price Index · H200 # H200 Cloud Pricing 2026: $2.60–$6.83/hr Across 4 Providers The sampled NVIDIA H200 rates have a minimum included rate of $2.60 , a median provider-average rate of $4.23 , and a maximum included rate of $6.83 per GPU-hour across 4 cloud providers with verified public rates (50 price rows, report assembled 2026-11-21). 16 H200 listings observed in the catalog from $3.50/GPU-hr, updated 2026-11-21. Report assembled 2026-11-21T17:02:24.332Z Source observations: 2026-11-21T17:02:03.771Z. Report assembled 2026-11-21T17:02:24.332Z; assembly does not refresh source evidence. Catalog snapshot 2026-11-21T17:00:36.649Z Dated catalog sample. Availability and prices may have changed. Check the current marketplace before relying on a listing. Min included rate $2.60 per GPU-hour Median provider average $4.23 per GPU-hour Max included rate $6.83 per GPU-hour Snapshot listings 16 from $3.50/GPU-hr ## Named-provider market rates — H200 Publicly listed on-demand rates, normalized to USD per GPU-hour. These are market-rate context from provider price pages, not Omega Gradient marketplace listings. | Provider | Listed SKU | Min $/GPU-hr | Avg $/GPU-hr | Max $/GPU-hr | Source / observation dates | | GMI Cloud | H200 | $2.60 | $2.60 | $2.60 | gmicloud.ai Source observations: 2026-11-21T17:02:03.771Z | | Oblivus Cloud | H200 | $3.99 | $3.99 | $3.99 | oblivus.com Source observations: 2026-11-21T17:02:03.771Z | | Vast.ai | H200 NVL, H200 | $3.18 | $4.47 | $6.83 | console.vast.ai Source observations: 2026-11-21T17:02:03.771Z | | CoreWeave | NVIDIA HGX H200 | $6.30 | $6.30 | $6.30 | coreweave.com Source observations: 2026-11-21T17:02:03.771Z | Public list prices reflect the source observation dates above; full report at /gpu-supplier-pricing.html . ## Marketplace catalog snapshot 16 H200 listings observed in the catalog from $3.50/GPU-hr, updated 2026-11-21. Marketplace listings are provider-anonymous by design. | GPU | Count | vRAM | Region | $/GPU/hr | Instance $/hr | | | H200 | 2× | 140 GB | US · Central | $3.50 | $7.01 | Check availability | | H200 | 4× | 140 GB | Europe · East | $3.92 | $15.70 | Check availability | | H200 | 2× | 140 GB | Europe · East | $3.94 | $7.89 | Check availability | | H200 | 1× | 140 GB | US · Central | $3.98 | $3.98 | Check availability | | H200 | 1× | 140 GB | Europe · East | $3.98 | $3.98 | Check availability | | H200 | 2× | 140 GB | Europe · South | $3.98 | $7.97 | Check availability | | H200 | 1× | 140 GB | Europe · South | $4.01 | $4.01 | Check availability | | H200 | 4× | 140 GB | US · Central | $4.33 | $17.33 | Check availability | | H200 | 8× | 141 GB | Global · flexible | $4.40 | $35.21 | Check availability | | H200 | 2× | 140 GB | Europe · West | $4.41 | $8.82 | Check availability | | H200 | 1× | 140 GB | Asia · Northeast | $4.42 | $4.42 | Check availability | | H200 | 1× | 140 GB | Global · flexible | $4.73 | $4.73 | Check availability | | H200 | 8× | 140 GB | Asia · Northeast | $5.43 | $43.44 | Check availability | | H200 | 4× | 140 GB | Asia · Northeast | $5.43 | $21.73 | Check availability | | H200 | 8× | 140 GB | Europe · South | $5.88 | $47.00 | Check availability | | H200 | 8× | 140 GB | Middle East | $7.51 | $60.09 | Check availability | Snapshot 2026-11-21; current view on /compute.html . ## Monthly cost math (720 hours) | Scenario | $/GPU-hr | × 720 hrs | | Lowest tracked public rate | $2.60 | $1872 | | Median provider-average rate | $4.23 | $3045 | | Highest tracked public rate | $6.83 | $4915 | | Cheapest sampled marketplace listing | $3.50 | $2524 | 720 hours ≈ one month of continuous single-GPU runtime. Multiply by GPU count for multi-GPU nodes. Put the data to work H100 & H200 ↗ Reserved clusters ↗ On-demand marketplace ↗ ## Frequently asked questions ### How much does a H200 cost per hour in 2026? Public on-demand rates tracked by Omega Gradient range from $2.60 to $6.83 per GPU-hour across 4 providers, with a median provider-average rate of $4.23 (report assembled 2026-11-21). ### Was a H200 listed in the catalog snapshot? 16 H200 listings observed in the catalog from $3.50/GPU-hr, updated 2026-11-21. Check current availability at omegagradient.com/compute.html. ### What does a H200 cost per month? At the median tracked rate of $4.23/GPU-hr, one GPU running a full 720-hour month costs about $3045. At the lowest tracked rate of $2.60/GPU-hr, about $1872. ### Where do these prices come from? Included public rate observations collected by Omega Gradient; dates can vary by source. Provider-normalized average: average SKUs/offers within each provider and GPU family first, then average provider means so one marketplace does not dominate. ---------------------------------------------------------------------- # B200 Cloud Pricing 2026: $6.56–$10.00/hr Across 1 Provider | Omega Gradient URL: https://omegagradient.com/gpus/b200.html Report assembled: 2026-11-21T17:02:24.332Z Cloud GPU Price Index · B200 # B200 Cloud Pricing 2026: $6.56–$10.00/hr Across 1 Provider The sampled NVIDIA B200 rates have a minimum included rate of $6.56 , a median provider-average rate of $8.44 , and a maximum included rate of $10.00 per GPU-hour across 1 cloud provider with verified public rates (9 price rows, report assembled 2026-11-21). 6 B200 listings observed in the catalog from $7.24/GPU-hr, updated 2026-11-21. Report assembled 2026-11-21T17:02:24.332Z Source observations: 2026-11-21T17:02:03.771Z. Report assembled 2026-11-21T17:02:24.332Z; assembly does not refresh source evidence. Catalog snapshot 2026-11-21T17:00:36.649Z Dated catalog sample. Availability and prices may have changed. Check the current marketplace before relying on a listing. Min included rate $6.56 per GPU-hour Median provider average $8.44 per GPU-hour Max included rate $10.00 per GPU-hour Snapshot listings 6 from $7.24/GPU-hr ## Named-provider market rates — B200 Publicly listed on-demand rates, normalized to USD per GPU-hour. These are market-rate context from provider price pages, not Omega Gradient marketplace listings. | Provider | Listed SKU | Min $/GPU-hr | Avg $/GPU-hr | Max $/GPU-hr | Source / observation dates | | Vast.ai | B200 | $6.56 | $8.44 | $10.00 | console.vast.ai Source observations: 2026-11-21T17:02:03.771Z | Public list prices reflect the source observation dates above; full report at /gpu-supplier-pricing.html . ## Marketplace catalog snapshot 6 B200 listings observed in the catalog from $7.24/GPU-hr, updated 2026-11-21. Marketplace listings are provider-anonymous by design. | GPU | Count | vRAM | Region | $/GPU/hr | Instance $/hr | | | B200 | 2× | 179 GB | US · West | $7.24 | $14.47 | Check availability | | B200 | 1× | 179 GB | US · West | $8.27 | $8.27 | Check availability | | B200 | 8× | 179 GB | US · Central | $9.63 | $77.02 | Check availability | | B200 | 4× | 179 GB | US · Central | $9.63 | $38.52 | Check availability | | B200 | 4× | 179 GB | US · West | $9.63 | $38.54 | Check availability | | B200 | 2× | 179 GB | US · Central | $11.01 | $22.02 | Check availability | Snapshot 2026-11-21; current view on /compute.html . ## Monthly cost math (720 hours) | Scenario | $/GPU-hr | × 720 hrs | | Lowest tracked public rate | $6.56 | $4726 | | Median provider-average rate | $8.44 | $6078 | | Highest tracked public rate | $10.00 | $7201 | | Cheapest sampled marketplace listing | $7.24 | $5209 | 720 hours ≈ one month of continuous single-GPU runtime. Multiply by GPU count for multi-GPU nodes. Put the data to work B200, B300 & GB300 ↗ Reserved clusters ↗ On-demand marketplace ↗ ## Frequently asked questions ### How much does a B200 cost per hour in 2026? Public on-demand rates tracked by Omega Gradient range from $6.56 to $10.00 per GPU-hour across 1 provider, with a median provider-average rate of $8.44 (report assembled 2026-11-21). ### Was a B200 listed in the catalog snapshot? 6 B200 listings observed in the catalog from $7.24/GPU-hr, updated 2026-11-21. Check current availability at omegagradient.com/compute.html. ### What does a B200 cost per month? At the median tracked rate of $8.44/GPU-hr, one GPU running a full 720-hour month costs about $6078. At the lowest tracked rate of $6.56/GPU-hr, about $4726. ### Where do these prices come from? Included public rate observations collected by Omega Gradient; dates can vary by source. Provider-normalized average: average SKUs/offers within each provider and GPU family first, then average provider means so one marketplace does not dominate. ---------------------------------------------------------------------- # B300 Cloud Pricing 2026: $6.64–$17.80/hr Across 10 Providers | Omega Gradient URL: https://omegagradient.com/gpus/b300.html Report assembled: 2026-11-21T17:02:24.332Z Cloud GPU Price Index · B300 # B300 Cloud Pricing 2026: $6.64–$17.80/hr Across 10 Providers The sampled NVIDIA B300 rates have a minimum included rate of $6.64 , a median provider-average rate of $9.74 , and a maximum included rate of $17.80 per GPU-hour across 10 cloud providers with verified public rates (15 price rows, report assembled 2026-11-21). 2 B300 listings observed in the catalog from $10.84/GPU-hr, updated 2026-11-21. Report assembled 2026-11-21T17:02:24.332Z Source observations: 2026-11-02T00:00:00.000Z to 2026-11-21T17:02:03.771Z. Report assembled 2026-11-21T17:02:24.332Z; assembly does not refresh source evidence. Catalog snapshot 2026-11-21T17:00:36.649Z Dated catalog sample. Availability and prices may have changed. Check the current marketplace before relying on a listing. Min included rate $6.64 per GPU-hour Median provider average $9.74 per GPU-hour Max included rate $17.80 per GPU-hour Snapshot listings 2 from $10.84/GPU-hr ## Named-provider market rates — B300 Publicly listed on-demand rates, normalized to USD per GPU-hour. These are market-rate context from provider price pages, not Omega Gradient marketplace listings. | Provider | Listed SKU | Min $/GPU-hr | Avg $/GPU-hr | Max $/GPU-hr | Source / observation dates | | Prime Intellect | B300 | $6.64 | $6.64 | $6.64 | primeintellect.ai Source observations: 2026-11-04T00:00:00.000Z | | Modal | B300 | $7.10 | $7.10 | $7.10 | modal.com Source observations: 2026-11-04T00:00:00.000Z | | CivoCloud | 8 x NVIDIA B300 | $7.49 | $7.49 | $7.49 | civo.com Source observations: 2026-11-04T00:00:00.000Z | | RunPod | B300 | $8.99 | $8.99 | $8.99 | runpod.io Source observations: 2026-11-04T00:00:00.000Z | | Nebius | NVIDIA HGX B300 | $9.50 | $9.50 | $9.50 | nebius.com Source observations: 2026-11-04T00:00:00.000Z | | Together AI | HGX B300 | $9.99 | $9.99 | $9.99 | together.ai Source observations: 2026-11-04T00:00:00.000Z | | Vast.ai | B300 | $9.85 | $11.12 | $11.88 | console.vast.ai Source observations: 2026-11-21T17:02:03.771Z | | Verda/DataCrunch | 1x B300 SXM6 268GB | $13.05 | $13.05 | $13.05 | verda.com Source observations: 2026-11-20T00:00:00.000Z | | Oracle Cloud Infrastructure | B300 | $15.00 | $15.00 | $15.00 | oracle.com Source observations: 2026-11-02T00:00:00.000Z | | AWS | p6-b300.48xlarge (8x B300) | $17.80 | $17.80 | $17.80 | aws.amazon.com Source observations: 2026-11-03T00:00:00.000Z | Public list prices reflect the source observation dates above; full report at /gpu-supplier-pricing.html . ## Marketplace catalog snapshot 2 B300 listings observed in the catalog from $10.84/GPU-hr, updated 2026-11-21. Marketplace listings are provider-anonymous by design. | GPU | Count | vRAM | Region | $/GPU/hr | Instance $/hr | | | B300 | 8× | 269 GB | US · Central | $10.84 | $86.70 | Check availability | | B300 | 2× | 269 GB | US · West | $11.70 | $23.40 | Check availability | Snapshot 2026-11-21; current view on /compute.html . ## Monthly cost math (720 hours) | Scenario | $/GPU-hr | × 720 hrs | | Lowest tracked public rate | $6.64 | $4781 | | Median provider-average rate | $9.74 | $7016 | | Highest tracked public rate | $17.80 | $12817 | | Cheapest sampled marketplace listing | $10.84 | $7803 | 720 hours ≈ one month of continuous single-GPU runtime. Multiply by GPU count for multi-GPU nodes. Put the data to work B200, B300 & GB300 ↗ Reserved clusters ↗ On-demand marketplace ↗ ## Frequently asked questions ### How much does a B300 cost per hour in 2026? Public on-demand rates tracked by Omega Gradient range from $6.64 to $17.80 per GPU-hour across 10 providers, with a median provider-average rate of $9.74 (report assembled 2026-11-21). ### Was a B300 listed in the catalog snapshot? 2 B300 listings observed in the catalog from $10.84/GPU-hr, updated 2026-11-21. Check current availability at omegagradient.com/compute.html. ### What does a B300 cost per month? At the median tracked rate of $9.74/GPU-hr, one GPU running a full 720-hour month costs about $7016. At the lowest tracked rate of $6.64/GPU-hr, about $4781. ### Where do these prices come from? Included public rate observations collected by Omega Gradient; dates can vary by source. Provider-normalized average: average SKUs/offers within each provider and GPU family first, then average provider means so one marketplace does not dominate. ---------------------------------------------------------------------- # A100 (80GB) Cloud Pricing 2026: $0.48–$7.12/hr Across 8 Providers | Omega Gradient URL: https://omegagradient.com/gpus/a100-80gb.html Report assembled: 2026-11-21T17:02:24.332Z Cloud GPU Price Index · A100 # A100 (80GB) Cloud Pricing 2026: $0.48–$7.12/hr Across 8 Providers The sampled NVIDIA A100 (80GB) rates have a minimum included rate of $0.48 , a median provider-average rate of $2.18 , and a maximum included rate of $7.12 per GPU-hour across 8 cloud providers with verified public rates (62 price rows, report assembled 2026-11-21). 22 A100 listings observed in the catalog from $0.60/GPU-hr, updated 2026-11-21. Report assembled 2026-11-21T17:02:24.332Z Source observations: 2026-11-21T17:02:03.771Z. Report assembled 2026-11-21T17:02:24.332Z; assembly does not refresh source evidence. Catalog snapshot 2026-11-21T17:00:36.649Z Dated catalog sample. Availability and prices may have changed. Check the current marketplace before relying on a listing. Min included rate $0.48 per GPU-hour Median provider average $2.18 per GPU-hour Max included rate $7.12 per GPU-hour Snapshot listings 22 from $0.60/GPU-hr ## Named-provider market rates — A100 Publicly listed on-demand rates, normalized to USD per GPU-hour. These are market-rate context from provider price pages, not Omega Gradient marketplace listings. Rates cover the A100 family; where a provider lists both 40GB and 80GB SKUs, the exact SKU is shown in the table below. Explicit 80GB SKUs in the dataset: 3 public rate row(s) from $1.15 to $5.04 per GPU-hour. | Provider | Listed SKU | Min $/GPU-hr | Avg $/GPU-hr | Max $/GPU-hr | Source / observation dates | | Vast.ai | A100 SXM4, A100 PCIE | $0.48 | $0.94 | $1.72 | console.vast.ai Source observations: 2026-11-21T17:02:03.771Z | | Koyeb | A100 | $1.60 | $1.60 | $1.60 | koyeb.com Source observations: 2026-11-21T17:02:03.771Z | | Paperspace | NVIDIA A100-80G, NVIDIA A100 | $1.15 | $1.70 | $2.24 | paperspace.com Source observations: 2026-11-21T17:02:03.771Z | | TensorDock | A100 SXM4 | $1.80 | $1.80 | $1.80 | tensordock.com Source observations: 2026-11-21T17:02:03.771Z | | Oblivus Cloud | A100 | $2.56 | $2.56 | $2.56 | oblivus.com Source observations: 2026-11-21T17:02:03.771Z | | CoreWeave | NVIDIA A100 | $2.70 | $2.70 | $2.70 | coreweave.com Source observations: 2026-11-21T17:02:03.771Z | | Replicate | Nvidia A100 80GB, 2x Nvidia A100 80GB | $5.04 | $5.04 | $5.04 | replicate.com Source observations: 2026-11-21T17:02:03.771Z | | CivoCloud | A100 | $7.12 | $7.12 | $7.12 | civo.com Source observations: 2026-11-21T17:02:03.771Z | Public list prices reflect the source observation dates above; full report at /gpu-supplier-pricing.html . ## Marketplace catalog snapshot 22 A100 listings observed in the catalog from $0.60/GPU-hr, updated 2026-11-21. Marketplace listings are provider-anonymous by design. | GPU | Count | vRAM | Region | $/GPU/hr | Instance $/hr | | | A100 | 1× | 40 GB | US · Central | $0.60 | $0.60 | Check availability | | A100 | 2× | 40 GB | Asia · Northeast | $0.60 | $1.21 | Check availability | | A100 | 2× | 40 GB | US · West | $0.61 | $1.22 | Check availability | | A100 | 2× | 40 GB | US · Central | $0.74 | $1.49 | Check availability | | A100 | 1× | 40 GB | Asia · Northeast | $0.77 | $0.77 | Check availability | | A100 | 4× | 80 GB | Europe · West | $0.96 | $3.84 | Check availability | | A100 | 2× | 80 GB | Europe · East | $1.32 | $2.64 | Check availability | | A100 | 1× | 80 GB | Europe · East | $1.33 | $1.33 | Check availability | | A100-80G | 1× | 80 GB | US · Central | $1.52 | $1.52 | Check availability | | A100 | 1× | 40 GB | Global · flexible | $1.54 | $1.54 | Check availability | | A100 | 8× | 40 GB | Global · flexible | $1.54 | $12.32 | Check availability | | A100 | 1× | 80 GB | US · East | $1.71 | $1.71 | Check availability | | A100-80G | 1× | 80 GB | Global · flexible | $2.04 | $2.04 | Check availability | | A100-80G | 2× | 80 GB | Global · flexible | $2.04 | $4.09 | Check availability | | A100 | 1× | 40 GB | India | $2.19 | $2.19 | Check availability | | A100 | 1× | 40 GB | US · West | $2.19 | $2.19 | Check availability | | A100 | 8× | 40 GB | US · West | $2.19 | $17.52 | Check availability | | A100-80G | 4× | 80 GB | US · Central | $2.20 | $8.81 | Check availability | | A100-80G | 2× | 80 GB | US · Central | $2.21 | $4.41 | Check availability | | A100-80G | 8× | 80 GB | US · Central | $3.07 | $24.56 | Check availability | | A100-80G | 8× | 80 GB | US · East | $3.07 | $24.56 | Check availability | | A100-80G | 1× | 80 GB | US · East | $3.61 | $3.61 | Check availability | Snapshot 2026-11-21; current view on /compute.html . ## Monthly cost math (720 hours) | Scenario | $/GPU-hr | × 720 hrs | | Lowest tracked public rate | $0.48 | $343 | | Median provider-average rate | $2.18 | $1570 | | Highest tracked public rate | $7.12 | $5126 | | Cheapest sampled marketplace listing | $0.60 | $432 | 720 hours ≈ one month of continuous single-GPU runtime. Multiply by GPU count for multi-GPU nodes. Put the data to work Reserved vs. on-demand ↗ Reserved clusters ↗ On-demand marketplace ↗ ## Frequently asked questions ### How much does a A100 (80GB) cost per hour in 2026? Public on-demand rates tracked by Omega Gradient range from $0.48 to $7.12 per GPU-hour across 8 providers, with a median provider-average rate of $2.18 (report assembled 2026-11-21). ### Was a A100 (80GB) listed in the catalog snapshot? 22 A100 listings observed in the catalog from $0.60/GPU-hr, updated 2026-11-21. Check current availability at omegagradient.com/compute.html. ### What does a A100 (80GB) cost per month? At the median tracked rate of $2.18/GPU-hr, one GPU running a full 720-hour month costs about $1570. At the lowest tracked rate of $0.48/GPU-hr, about $343. ### Where do these prices come from? Included public rate observations collected by Omega Gradient; dates can vary by source. Provider-normalized average: average SKUs/offers within each provider and GPU family first, then average provider means so one marketplace does not dominate. ---------------------------------------------------------------------- # L40S Cloud Pricing 2026: $0.47–$3.51/hr Across 5 Providers | Omega Gradient URL: https://omegagradient.com/gpus/l40s.html Report assembled: 2026-11-21T17:02:24.332Z Cloud GPU Price Index · L40S # L40S Cloud Pricing 2026: $0.47–$3.51/hr Across 5 Providers The sampled NVIDIA L40S rates have a minimum included rate of $0.47 , a median provider-average rate of $1.20 , and a maximum included rate of $3.51 per GPU-hour across 5 cloud providers with verified public rates (17 price rows, report assembled 2026-11-21). 7 L40S listings observed in the catalog from $0.54/GPU-hr, updated 2026-11-21. Report assembled 2026-11-21T17:02:24.332Z Source observations: 2026-11-21T17:02:03.771Z. Report assembled 2026-11-21T17:02:24.332Z; assembly does not refresh source evidence. Catalog snapshot 2026-11-21T17:00:36.649Z Dated catalog sample. Availability and prices may have changed. Check the current marketplace before relying on a listing. Min included rate $0.47 per GPU-hour Median provider average $1.20 per GPU-hour Max included rate $3.51 per GPU-hour Snapshot listings 7 from $0.54/GPU-hr ## Named-provider market rates — L40S Publicly listed on-demand rates, normalized to USD per GPU-hour. These are market-rate context from provider price pages, not Omega Gradient marketplace listings. | Provider | Listed SKU | Min $/GPU-hr | Avg $/GPU-hr | Max $/GPU-hr | Source / observation dates | | HydraHost | L40S | $0.70 | $0.70 | $0.70 | hydrahost.com Source observations: 2026-11-21T17:02:03.771Z | | Vast.ai | L40S | $0.47 | $0.81 | $1.08 | console.vast.ai Source observations: 2026-11-21T17:02:03.771Z | | Koyeb | L40S | $1.20 | $1.20 | $1.20 | koyeb.com Source observations: 2026-11-21T17:02:03.771Z | | CoreWeave | NVIDIA L40S | $2.25 | $2.25 | $2.25 | coreweave.com Source observations: 2026-11-21T17:02:03.771Z | | Replicate | Nvidia L40S, 2x Nvidia L40S | $3.51 | $3.51 | $3.51 | replicate.com Source observations: 2026-11-21T17:02:03.771Z | Public list prices reflect the source observation dates above; full report at /gpu-supplier-pricing.html . ## Marketplace catalog snapshot 7 L40S listings observed in the catalog from $0.54/GPU-hr, updated 2026-11-21. Marketplace listings are provider-anonymous by design. | GPU | Count | vRAM | Region | $/GPU/hr | Instance $/hr | | | L40S | 1× | 45 GB | US · East | $0.54 | $0.54 | Check availability | | L40S | 1× | 48 GB | US · Central | $0.97 | $0.97 | Check availability | | L40S | 2× | 48 GB | US · Central | $0.97 | $1.94 | Check availability | | L40S | 4× | 48 GB | US · Central | $0.97 | $3.88 | Check availability | | L40S | 8× | 48 GB | US · Central | $1.65 | $13.20 | Check availability | | L40S | 10× | 48 GB | US · Central | $1.65 | $16.51 | Check availability | | L40S | 1× | 48 GB | Europe · East | $1.88 | $1.88 | Check availability | Snapshot 2026-11-21; current view on /compute.html . ## Monthly cost math (720 hours) | Scenario | $/GPU-hr | × 720 hrs | | Lowest tracked public rate | $0.47 | $338 | | Median provider-average rate | $1.20 | $864 | | Highest tracked public rate | $3.51 | $2527 | | Cheapest sampled marketplace listing | $0.54 | $389 | 720 hours ≈ one month of continuous single-GPU runtime. Multiply by GPU count for multi-GPU nodes. Put the data to work Reserved vs. on-demand ↗ Reserved clusters ↗ On-demand marketplace ↗ ## Frequently asked questions ### How much does a L40S cost per hour in 2026? Public on-demand rates tracked by Omega Gradient range from $0.47 to $3.51 per GPU-hour across 5 providers, with a median provider-average rate of $1.20 (report assembled 2026-11-21). ### Was a L40S listed in the catalog snapshot? 7 L40S listings observed in the catalog from $0.54/GPU-hr, updated 2026-11-21. Check current availability at omegagradient.com/compute.html. ### What does a L40S cost per month? At the median tracked rate of $1.20/GPU-hr, one GPU running a full 720-hour month costs about $864. At the lowest tracked rate of $0.47/GPU-hr, about $338. ### Where do these prices come from? Included public rate observations collected by Omega Gradient; dates can vary by source. Provider-normalized average: average SKUs/offers within each provider and GPU family first, then average provider means so one marketplace does not dominate. ---------------------------------------------------------------------- # L40 Cloud Pricing 2026: $0.70–$1.25/hr Across 3 Providers | Omega Gradient URL: https://omegagradient.com/gpus/l40.html Report assembled: 2026-11-21T17:02:24.332Z Cloud GPU Price Index · L40 # L40 Cloud Pricing 2026: $0.70–$1.25/hr Across 3 Providers The sampled NVIDIA L40 rates have a minimum included rate of $0.70 , a median provider-average rate of $0.95 , and a maximum included rate of $1.25 per GPU-hour across 3 cloud providers with verified public rates (3 price rows, report assembled 2026-11-21). 3 L40 listings observed in the catalog from $0.97/GPU-hr, updated 2026-11-21. Report assembled 2026-11-21T17:02:24.332Z Source observations: 2026-11-21T17:02:03.771Z. Report assembled 2026-11-21T17:02:24.332Z; assembly does not refresh source evidence. Catalog snapshot 2026-11-21T17:00:36.649Z Dated catalog sample. Availability and prices may have changed. Check the current marketplace before relying on a listing. Min included rate $0.70 per GPU-hour Median provider average $0.95 per GPU-hour Max included rate $1.25 per GPU-hour Snapshot listings 3 from $0.97/GPU-hr ## Named-provider market rates — L40 Publicly listed on-demand rates, normalized to USD per GPU-hour. These are market-rate context from provider price pages, not Omega Gradient marketplace listings. | Provider | Listed SKU | Min $/GPU-hr | Avg $/GPU-hr | Max $/GPU-hr | Source / observation dates | | HydraHost | L40 | $0.70 | $0.70 | $0.70 | hydrahost.com Source observations: 2026-11-21T17:02:03.771Z | | Oblivus Cloud | L40 | $0.95 | $0.95 | $0.95 | oblivus.com Source observations: 2026-11-21T17:02:03.771Z | | CoreWeave | NVIDIA L40 | $1.25 | $1.25 | $1.25 | coreweave.com Source observations: 2026-11-21T17:02:03.771Z | Public list prices reflect the source observation dates above; full report at /gpu-supplier-pricing.html . ## Marketplace catalog snapshot 3 L40 listings observed in the catalog from $0.97/GPU-hr, updated 2026-11-21. Marketplace listings are provider-anonymous by design. | GPU | Count | vRAM | Region | $/GPU/hr | Instance $/hr | | | L40 | 1× | 48 GB | US · Central | $0.97 | $0.97 | Check availability | | L40 | 2× | 48 GB | US · Central | $1.09 | $2.18 | Check availability | | L40 | 1× | 48 GB | Canada | $1.11 | $1.11 | Check availability | Snapshot 2026-11-21; current view on /compute.html . ## Monthly cost math (720 hours) | Scenario | $/GPU-hr | × 720 hrs | | Lowest tracked public rate | $0.70 | $504 | | Median provider-average rate | $0.95 | $684 | | Highest tracked public rate | $1.25 | $900 | | Cheapest sampled marketplace listing | $0.97 | $698 | 720 hours ≈ one month of continuous single-GPU runtime. Multiply by GPU count for multi-GPU nodes. Put the data to work Reserved vs. on-demand ↗ Reserved clusters ↗ On-demand marketplace ↗ ## Frequently asked questions ### How much does a L40 cost per hour in 2026? Public on-demand rates tracked by Omega Gradient range from $0.70 to $1.25 per GPU-hour across 3 providers, with a median provider-average rate of $0.95 (report assembled 2026-11-21). ### Was a L40 listed in the catalog snapshot? 3 L40 listings observed in the catalog from $0.97/GPU-hr, updated 2026-11-21. Check current availability at omegagradient.com/compute.html. ### What does a L40 cost per month? At the median tracked rate of $0.95/GPU-hr, one GPU running a full 720-hour month costs about $684. At the lowest tracked rate of $0.70/GPU-hr, about $504. ### Where do these prices come from? Included public rate observations collected by Omega Gradient; dates can vary by source. Provider-normalized average: average SKUs/offers within each provider and GPU family first, then average provider means so one marketplace does not dominate. ---------------------------------------------------------------------- # RTX 4090 Cloud Pricing 2026: $0.35–$0.74/hr Across 4 Providers | Omega Gradient URL: https://omegagradient.com/gpus/rtx-4090.html Report assembled: 2026-11-21T17:02:24.332Z Cloud GPU Price Index · RTX 4090 # RTX 4090 Cloud Pricing 2026: $0.35–$0.74/hr Across 4 Providers The sampled NVIDIA RTX 4090 rates have a minimum included rate of $0.35 , a median provider-average rate of $0.43 , and a maximum included rate of $0.74 per GPU-hour across 4 cloud providers with verified public rates (8 price rows, report assembled 2026-11-21). 2 RTX 4090 listings observed in the catalog from $0.44/GPU-hr, updated 2026-11-21. Report assembled 2026-11-21T17:02:24.332Z Source observations: 2026-11-21T17:02:03.771Z. Report assembled 2026-11-21T17:02:24.332Z; assembly does not refresh source evidence. Catalog snapshot 2026-11-21T17:00:36.649Z Dated catalog sample. Availability and prices may have changed. Check the current marketplace before relying on a listing. Min included rate $0.35 per GPU-hour Median provider average $0.43 per GPU-hour Max included rate $0.74 per GPU-hour Snapshot listings 2 from $0.44/GPU-hr ## Named-provider market rates — RTX 4090 Publicly listed on-demand rates, normalized to USD per GPU-hour. These are market-rate context from provider price pages, not Omega Gradient marketplace listings. | Provider | Listed SKU | Min $/GPU-hr | Avg $/GPU-hr | Max $/GPU-hr | Source / observation dates | | TensorDock | RTX 4090 | $0.35 | $0.35 | $0.35 | tensordock.com Source observations: 2026-11-21T17:02:03.771Z | | HydraHost | RTX 4090 | $0.40 | $0.40 | $0.40 | hydrahost.com Source observations: 2026-11-21T17:02:03.771Z | | Vast.ai | RTX 4090 | $0.37 | $0.45 | $0.54 | console.vast.ai Source observations: 2026-11-21T17:02:03.771Z | | Oblivus Cloud | RTX 4090 | $0.74 | $0.74 | $0.74 | oblivus.com Source observations: 2026-11-21T17:02:03.771Z | Public list prices reflect the source observation dates above; full report at /gpu-supplier-pricing.html . ## Marketplace catalog snapshot 2 RTX 4090 listings observed in the catalog from $0.44/GPU-hr, updated 2026-11-21. Marketplace listings are provider-anonymous by design. | GPU | Count | vRAM | Region | $/GPU/hr | Instance $/hr | | | RTX 4090 | 8× | 24 GB | US · Central | $0.44 | $3.52 | Check availability | | RTX 4090 | 1× | 24 GB | Global · flexible | $0.66 | $0.66 | Check availability | Snapshot 2026-11-21; current view on /compute.html . ## Monthly cost math (720 hours) | Scenario | $/GPU-hr | × 720 hrs | | Lowest tracked public rate | $0.35 | $252 | | Median provider-average rate | $0.43 | $307 | | Highest tracked public rate | $0.74 | $533 | | Cheapest sampled marketplace listing | $0.44 | $317 | 720 hours ≈ one month of continuous single-GPU runtime. Multiply by GPU count for multi-GPU nodes. Put the data to work Reserved vs. on-demand ↗ Reserved clusters ↗ On-demand marketplace ↗ ## Frequently asked questions ### How much does a RTX 4090 cost per hour in 2026? Public on-demand rates tracked by Omega Gradient range from $0.35 to $0.74 per GPU-hour across 4 providers, with a median provider-average rate of $0.43 (report assembled 2026-11-21). ### Was a RTX 4090 listed in the catalog snapshot? 2 RTX 4090 listings observed in the catalog from $0.44/GPU-hr, updated 2026-11-21. Check current availability at omegagradient.com/compute.html. ### What does a RTX 4090 cost per month? At the median tracked rate of $0.43/GPU-hr, one GPU running a full 720-hour month costs about $307. At the lowest tracked rate of $0.35/GPU-hr, about $252. ### Where do these prices come from? Included public rate observations collected by Omega Gradient; dates can vary by source. Provider-normalized average: average SKUs/offers within each provider and GPU family first, then average provider means so one marketplace does not dominate. ---------------------------------------------------------------------- # RTX 5090 Cloud Pricing 2026: $0.42–$1.16/hr Across 2 Providers | Omega Gradient URL: https://omegagradient.com/gpus/rtx-5090.html Report assembled: 2026-11-21T17:02:24.332Z Cloud GPU Price Index · RTX 5090 # RTX 5090 Cloud Pricing 2026: $0.42–$1.16/hr Across 2 Providers The sampled NVIDIA RTX 5090 rates have a minimum included rate of $0.42 , a median provider-average rate of $0.73 , and a maximum included rate of $1.16 per GPU-hour across 2 cloud providers with verified public rates (63 price rows, report assembled 2026-11-21). 1 RTX 5090 listing observed in the catalog from $0.71/GPU-hr, updated 2026-11-21. Report assembled 2026-11-21T17:02:24.332Z Source observations: 2026-11-21T17:02:03.771Z. Report assembled 2026-11-21T17:02:24.332Z; assembly does not refresh source evidence. Catalog snapshot 2026-11-21T17:00:36.649Z Dated catalog sample. Availability and prices may have changed. Check the current marketplace before relying on a listing. Min included rate $0.42 per GPU-hour Median provider average $0.73 per GPU-hour Max included rate $1.16 per GPU-hour Snapshot listings 1 from $0.71/GPU-hr ## Named-provider market rates — RTX 5090 Publicly listed on-demand rates, normalized to USD per GPU-hour. These are market-rate context from provider price pages, not Omega Gradient marketplace listings. | Provider | Listed SKU | Min $/GPU-hr | Avg $/GPU-hr | Max $/GPU-hr | Source / observation dates | | HydraHost | RTX 5090 | $0.70 | $0.70 | $0.70 | hydrahost.com Source observations: 2026-11-21T17:02:03.771Z | | Vast.ai | RTX 5090 | $0.42 | $0.77 | $1.16 | console.vast.ai Source observations: 2026-11-21T17:02:03.771Z | Public list prices reflect the source observation dates above; full report at /gpu-supplier-pricing.html . ## Marketplace catalog snapshot 1 RTX 5090 listing observed in the catalog from $0.71/GPU-hr, updated 2026-11-21. Marketplace listings are provider-anonymous by design. | GPU | Count | vRAM | Region | $/GPU/hr | Instance $/hr | | | RTX 5090 | 8× | 32 GB | Global · flexible | $0.71 | $5.72 | Check availability | Snapshot 2026-11-21; current view on /compute.html . ## Monthly cost math (720 hours) | Scenario | $/GPU-hr | × 720 hrs | | Lowest tracked public rate | $0.42 | $299 | | Median provider-average rate | $0.73 | $528 | | Highest tracked public rate | $1.16 | $838 | | Cheapest sampled marketplace listing | $0.71 | $515 | 720 hours ≈ one month of continuous single-GPU runtime. Multiply by GPU count for multi-GPU nodes. Put the data to work Reserved vs. on-demand ↗ Reserved clusters ↗ On-demand marketplace ↗ ## Frequently asked questions ### How much does a RTX 5090 cost per hour in 2026? Public on-demand rates tracked by Omega Gradient range from $0.42 to $1.16 per GPU-hour across 2 providers, with a median provider-average rate of $0.73 (report assembled 2026-11-21). ### Was a RTX 5090 listed in the catalog snapshot? 1 RTX 5090 listing observed in the catalog from $0.71/GPU-hr, updated 2026-11-21. Check current availability at omegagradient.com/compute.html. ### What does a RTX 5090 cost per month? At the median tracked rate of $0.73/GPU-hr, one GPU running a full 720-hour month costs about $528. At the lowest tracked rate of $0.42/GPU-hr, about $299. ### Where do these prices come from? Included public rate observations collected by Omega Gradient; dates can vary by source. Provider-normalized average: average SKUs/offers within each provider and GPU family first, then average provider means so one marketplace does not dominate. ---------------------------------------------------------------------- # RTX A6000 Cloud Pricing 2026: $0.44–$0.75/hr Across 3 Providers | Omega Gradient URL: https://omegagradient.com/gpus/a6000.html Report assembled: 2026-11-21T17:02:24.332Z Cloud GPU Price Index · A6000 # RTX A6000 Cloud Pricing 2026: $0.44–$0.75/hr Across 3 Providers The sampled NVIDIA RTX A6000 rates have a minimum included rate of $0.44 , a median provider-average rate of $0.70 , and a maximum included rate of $0.75 per GPU-hour across 3 cloud providers with verified public rates (3 price rows, report assembled 2026-11-21). 7 A6000 listings observed in the catalog from $0.56/GPU-hr, updated 2026-11-21. Report assembled 2026-11-21T17:02:24.332Z Source observations: 2026-11-21T17:02:03.771Z. Report assembled 2026-11-21T17:02:24.332Z; assembly does not refresh source evidence. Catalog snapshot 2026-11-21T17:00:36.649Z Dated catalog sample. Availability and prices may have changed. Check the current marketplace before relying on a listing. Min included rate $0.44 per GPU-hour Median provider average $0.70 per GPU-hour Max included rate $0.75 per GPU-hour Snapshot listings 7 from $0.56/GPU-hr ## Named-provider market rates — A6000 Publicly listed on-demand rates, normalized to USD per GPU-hour. These are market-rate context from provider price pages, not Omega Gradient marketplace listings. | Provider | Listed SKU | Min $/GPU-hr | Avg $/GPU-hr | Max $/GPU-hr | Source / observation dates | | Oblivus Cloud | RTX A6000 | $0.44 | $0.44 | $0.44 | oblivus.com Source observations: 2026-11-21T17:02:03.771Z | | HydraHost | RTX A6000 | $0.70 | $0.70 | $0.70 | hydrahost.com Source observations: 2026-11-21T17:02:03.771Z | | Koyeb | RTX-A6000 | $0.75 | $0.75 | $0.75 | koyeb.com Source observations: 2026-11-21T17:02:03.771Z | Public list prices reflect the source observation dates above; full report at /gpu-supplier-pricing.html . ## Marketplace catalog snapshot 7 A6000 listings observed in the catalog from $0.56/GPU-hr, updated 2026-11-21. Marketplace listings are provider-anonymous by design. | GPU | Count | vRAM | Region | $/GPU/hr | Instance $/hr | | | RTX A6000 | 1× | 48 GB | Canada | $0.56 | $0.56 | Check availability | | RTX A6000 | 4× | 48 GB | US · Central | $0.63 | $2.51 | Check availability | | RTX A6000 | 1× | 48 GB | US · Central | $0.63 | $0.63 | Check availability | | RTX A6000 | 2× | 48 GB | US · Central | $0.63 | $1.26 | Check availability | | RTX A6000 | 2× | 48 GB | US · East | $2.13 | $4.25 | Check availability | | RTX A6000 | 4× | 48 GB | US · East | $2.13 | $8.50 | Check availability | | RTX A6000 | 1× | 48 GB | US · East | $2.13 | $2.13 | Check availability | Snapshot 2026-11-21; current view on /compute.html . ## Monthly cost math (720 hours) | Scenario | $/GPU-hr | × 720 hrs | | Lowest tracked public rate | $0.44 | $317 | | Median provider-average rate | $0.70 | $504 | | Highest tracked public rate | $0.75 | $540 | | Cheapest sampled marketplace listing | $0.56 | $403 | 720 hours ≈ one month of continuous single-GPU runtime. Multiply by GPU count for multi-GPU nodes. Put the data to work Reserved vs. on-demand ↗ Reserved clusters ↗ On-demand marketplace ↗ ## Frequently asked questions ### How much does a RTX A6000 cost per hour in 2026? Public on-demand rates tracked by Omega Gradient range from $0.44 to $0.75 per GPU-hour across 3 providers, with a median provider-average rate of $0.70 (report assembled 2026-11-21). ### Was a RTX A6000 listed in the catalog snapshot? 7 A6000 listings observed in the catalog from $0.56/GPU-hr, updated 2026-11-21. Check current availability at omegagradient.com/compute.html. ### What does a RTX A6000 cost per month? At the median tracked rate of $0.70/GPU-hr, one GPU running a full 720-hour month costs about $504. At the lowest tracked rate of $0.44/GPU-hr, about $317. ### Where do these prices come from? Included public rate observations collected by Omega Gradient; dates can vary by source. Provider-normalized average: average SKUs/offers within each provider and GPU family first, then average provider means so one marketplace does not dominate. ---------------------------------------------------------------------- # RTX A4000 Cloud Pricing 2026: $0.76/hr Across 1 Provider | Omega Gradient URL: https://omegagradient.com/gpus/a4000.html Report assembled: 2026-11-21T17:02:24.332Z Cloud GPU Price Index · A4000 # RTX A4000 Cloud Pricing 2026: $0.76/hr Across 1 Provider The sampled NVIDIA RTX A4000 rates have a minimum included rate of $0.76 , a median provider-average rate of $0.76 , and a maximum included rate of $0.76 per GPU-hour across 1 cloud provider with verified public rates (1 price rows, report assembled 2026-11-21). 4 A4000 listings observed in the catalog from $0.23/GPU-hr, updated 2026-11-21. Report assembled 2026-11-21T17:02:24.332Z Source observations: 2026-11-21T17:02:03.771Z. Report assembled 2026-11-21T17:02:24.332Z; assembly does not refresh source evidence. Catalog snapshot 2026-11-21T17:00:36.649Z Dated catalog sample. Availability and prices may have changed. Check the current marketplace before relying on a listing. Min included rate $0.76 per GPU-hour Median provider average $0.76 per GPU-hour Max included rate $0.76 per GPU-hour Snapshot listings 4 from $0.23/GPU-hr ## Named-provider market rates — A4000 Publicly listed on-demand rates, normalized to USD per GPU-hour. These are market-rate context from provider price pages, not Omega Gradient marketplace listings. | Provider | Listed SKU | Min $/GPU-hr | Avg $/GPU-hr | Max $/GPU-hr | Source / observation dates | | Paperspace | NVIDIA A4000 | $0.76 | $0.76 | $0.76 | paperspace.com Source observations: 2026-11-21T17:02:03.771Z | Public list prices reflect the source observation dates above; full report at /gpu-supplier-pricing.html . ## Marketplace catalog snapshot 4 A4000 listings observed in the catalog from $0.23/GPU-hr, updated 2026-11-21. Marketplace listings are provider-anonymous by design. | GPU | Count | vRAM | Region | $/GPU/hr | Instance $/hr | | | RTX A4000 | 1× | 16 GB | Europe · North | $0.23 | $0.23 | Check availability | | RTX A4000 | 1× | 16 GB | US · East | $0.88 | $0.88 | Check availability | | RTX A4000 | 2× | 16 GB | US · East | $0.88 | $1.76 | Check availability | | RTX A4000 | 4× | 16 GB | US · East | $0.88 | $3.52 | Check availability | Snapshot 2026-11-21; current view on /compute.html . ## Monthly cost math (720 hours) | Scenario | $/GPU-hr | × 720 hrs | | Lowest tracked public rate | $0.76 | $547 | | Median provider-average rate | $0.76 | $547 | | Highest tracked public rate | $0.76 | $547 | | Cheapest sampled marketplace listing | $0.23 | $166 | 720 hours ≈ one month of continuous single-GPU runtime. Multiply by GPU count for multi-GPU nodes. Put the data to work Reserved vs. on-demand ↗ Reserved clusters ↗ On-demand marketplace ↗ ## Frequently asked questions ### How much does a RTX A4000 cost per hour in 2026? Omega Gradient's dated report includes one public on-demand rate for this GPU: $0.76 per GPU-hour from 1 provider (report assembled 2026-11-21). A single observation is a point of reference, not a market range. ### Was a RTX A4000 listed in the catalog snapshot? 4 A4000 listings observed in the catalog from $0.23/GPU-hr, updated 2026-11-21. Check current availability at omegagradient.com/compute.html. ### What does a RTX A4000 cost per month? At the median tracked rate of $0.76/GPU-hr, one GPU running a full 720-hour month costs about $547. At the lowest tracked rate of $0.76/GPU-hr, about $547. ### Where do these prices come from? Included public rate observations collected by Omega Gradient; dates can vary by source. Provider-normalized average: average SKUs/offers within each provider and GPU family first, then average provider means so one marketplace does not dominate. ---------------------------------------------------------------------- # T4 Cloud Pricing 2026: $0.81/hr Across 1 Provider | Omega Gradient URL: https://omegagradient.com/gpus/t4.html Report assembled: 2026-11-21T17:02:24.332Z Cloud GPU Price Index · T4 # T4 Cloud Pricing 2026: $0.81/hr Across 1 Provider The sampled NVIDIA T4 rates have a minimum included rate of $0.81 , a median provider-average rate of $0.81 , and a maximum included rate of $0.81 per GPU-hour across 1 cloud provider with verified public rates (1 price rows, report assembled 2026-11-21). No anonymous T4 marketplace listings in the dated snapshot (2026-11-21); the sourcing desk secures T4 capacity on request via /clusters/#request. Report assembled 2026-11-21T17:02:24.332Z Source observations: 2026-11-21T17:02:03.771Z. Report assembled 2026-11-21T17:02:24.332Z; assembly does not refresh source evidence. Catalog snapshot 2026-11-21T17:00:36.649Z Dated catalog sample. Availability and prices may have changed. Check the current marketplace before relying on a listing. Min included rate $0.81 per GPU-hour Median provider average $0.81 per GPU-hour Max included rate $0.81 per GPU-hour Snapshot listings 0 desk sourcing available ## Named-provider market rates — T4 Publicly listed on-demand rates, normalized to USD per GPU-hour. These are market-rate context from provider price pages, not Omega Gradient marketplace listings. | Provider | Listed SKU | Min $/GPU-hr | Avg $/GPU-hr | Max $/GPU-hr | Source / observation dates | | Replicate | Nvidia T4 | $0.81 | $0.81 | $0.81 | replicate.com Source observations: 2026-11-21T17:02:03.771Z | Public list prices reflect the source observation dates above; full report at /gpu-supplier-pricing.html . ## Marketplace catalog snapshot No anonymous T4 marketplace listings in the dated snapshot (2026-11-21); the sourcing desk secures T4 capacity on request via /clusters/#request. Marketplace listings are provider-anonymous by design. Request reserved T4 capacity through the cluster sourcing desk or hello@omegagradient.com . ## Monthly cost math (720 hours) | Scenario | $/GPU-hr | × 720 hrs | | Lowest tracked public rate | $0.81 | $583 | | Median provider-average rate | $0.81 | $583 | | Highest tracked public rate | $0.81 | $583 | 720 hours ≈ one month of continuous single-GPU runtime. Multiply by GPU count for multi-GPU nodes. Put the data to work Reserved vs. on-demand ↗ Reserved clusters ↗ On-demand marketplace ↗ ## Frequently asked questions ### How much does a T4 cost per hour in 2026? Omega Gradient's dated report includes one public on-demand rate for this GPU: $0.81 per GPU-hour from 1 provider (report assembled 2026-11-21). A single observation is a point of reference, not a market range. ### Was a T4 listed in the catalog snapshot? No anonymous T4 marketplace listings in the dated snapshot (2026-11-21); the sourcing desk secures T4 capacity on request via /clusters/#request. ### What does a T4 cost per month? At the median tracked rate of $0.81/GPU-hr, one GPU running a full 720-hour month costs about $583. At the lowest tracked rate of $0.81/GPU-hr, about $583. ### Where do these prices come from? Included public rate observations collected by Omega Gradient; dates can vary by source. Provider-normalized average: average SKUs/offers within each provider and GPU family first, then average provider means so one marketplace does not dominate. ---------------------------------------------------------------------- # GB200 NVL72 Cloud Pricing 2026: $9.38–$9.39/hr Across 1 Provider | Omega Gradient URL: https://omegagradient.com/gpus/gb200.html Report assembled: 2026-11-21T17:02:24.332Z Cloud GPU Price Index · GB200 # GB200 NVL72 Cloud Pricing 2026: $9.38–$9.39/hr Across 1 Provider The sampled NVIDIA GB200 NVL72 rates have a minimum included rate of $9.38 , a median provider-average rate of $9.38 , and a maximum included rate of $9.39 per GPU-hour across 1 cloud provider with verified public rates (3 price rows, report assembled 2026-11-21). 3 GB200 listings observed in the catalog from $10.34/GPU-hr, updated 2026-11-21. Report assembled 2026-11-21T17:02:24.332Z Source observations: 2026-11-21T17:02:03.771Z. Report assembled 2026-11-21T17:02:24.332Z; assembly does not refresh source evidence. Catalog snapshot 2026-11-21T17:00:36.649Z Dated catalog sample. Availability and prices may have changed. Check the current marketplace before relying on a listing. Min included rate $9.38 per GPU-hour Median provider average $9.38 per GPU-hour Max included rate $9.39 per GPU-hour Snapshot listings 3 from $10.34/GPU-hr ## Named-provider market rates — GB200 Publicly listed on-demand rates, normalized to USD per GPU-hour. These are market-rate context from provider price pages, not Omega Gradient marketplace listings. | Provider | Listed SKU | Min $/GPU-hr | Avg $/GPU-hr | Max $/GPU-hr | Source / observation dates | | Vast.ai | GB200 | $9.38 | $9.38 | $9.39 | console.vast.ai Source observations: 2026-11-21T17:02:03.771Z | Public list prices reflect the source observation dates above; full report at /gpu-supplier-pricing.html . ## Marketplace catalog snapshot 3 GB200 listings observed in the catalog from $10.34/GPU-hr, updated 2026-11-21. Marketplace listings are provider-anonymous by design. | GPU | Count | vRAM | Region | $/GPU/hr | Instance $/hr | | | GB200 | 4× | 185 GB | Asia · Southeast | $10.34 | $41.35 | Check availability | | GB200 | 2× | 185 GB | Asia · Southeast | $10.36 | $20.72 | Check availability | | GB200 | 1× | 185 GB | Asia · Southeast | $10.41 | $10.41 | Check availability | Snapshot 2026-11-21; current view on /compute.html . ## Monthly cost math (720 hours) | Scenario | $/GPU-hr | × 720 hrs | | Lowest tracked public rate | $9.38 | $6752 | | Median provider-average rate | $9.38 | $6754 | | Highest tracked public rate | $9.39 | $6757 | | Cheapest sampled marketplace listing | $10.34 | $7443 | 720 hours ≈ one month of continuous single-GPU runtime. Multiply by GPU count for multi-GPU nodes. Put the data to work B200, B300 & GB300 ↗ Reserved clusters ↗ On-demand marketplace ↗ ## Frequently asked questions ### How much does a GB200 NVL72 cost per hour in 2026? Public on-demand rates tracked by Omega Gradient range from $9.38 to $9.39 per GPU-hour across 1 provider, with a median provider-average rate of $9.38 (report assembled 2026-11-21). ### Was a GB200 NVL72 listed in the catalog snapshot? 3 GB200 listings observed in the catalog from $10.34/GPU-hr, updated 2026-11-21. Check current availability at omegagradient.com/compute.html. ### What does a GB200 NVL72 cost per month? At the median tracked rate of $9.38/GPU-hr, one GPU running a full 720-hour month costs about $6754. At the lowest tracked rate of $9.38/GPU-hr, about $6752. ### Where do these prices come from? Included public rate observations collected by Omega Gradient; dates can vary by source. Provider-normalized average: average SKUs/offers within each provider and GPU family first, then average provider means so one marketplace does not dominate. ---------------------------------------------------------------------- # GB300 NVL72 Cloud Pricing 2026: $15.01–$18.00/hr Across 2 Providers | Omega Gradient URL: https://omegagradient.com/gpus/gb300.html Report assembled: 2026-11-21T17:02:24.332Z Cloud GPU Price Index · GB300 # GB300 NVL72 Cloud Pricing 2026: $15.01–$18.00/hr Across 2 Providers The sampled NVIDIA GB300 NVL72 rates have a minimum included rate of $15.01 , a median provider-average rate of $16.50 , and a maximum included rate of $18.00 per GPU-hour across 2 cloud providers with verified public rates (2 price rows, report assembled 2026-11-21). No anonymous GB300 marketplace listings in the dated snapshot (2026-11-21); the sourcing desk secures GB300 capacity on request via /clusters/#request. Report assembled 2026-11-21T17:02:24.332Z Source observations: 2026-11-02T00:00:00.000Z to 2026-11-20T00:00:00.000Z. Report assembled 2026-11-21T17:02:24.332Z; assembly does not refresh source evidence. Catalog snapshot 2026-11-21T17:00:36.649Z Dated catalog sample. Availability and prices may have changed. Check the current marketplace before relying on a listing. Min included rate $15.01 per GPU-hour Median provider average $16.50 per GPU-hour Max included rate $18.00 per GPU-hour Snapshot listings 0 desk sourcing available ## Named-provider market rates — GB300 Publicly listed on-demand rates, normalized to USD per GPU-hour. These are market-rate context from provider price pages, not Omega Gradient marketplace listings. | Provider | Listed SKU | Min $/GPU-hr | Avg $/GPU-hr | Max $/GPU-hr | Source / observation dates | | Verda/DataCrunch | 1x GB300 SXM6 288GB | $15.01 | $15.01 | $15.01 | verda.com Source observations: 2026-11-20T00:00:00.000Z | | Oracle Cloud Infrastructure | GB300 NVL72 | $18.00 | $18.00 | $18.00 | oracle.com Source observations: 2026-11-02T00:00:00.000Z | Public list prices reflect the source observation dates above; full report at /gpu-supplier-pricing.html . ## Marketplace catalog snapshot No anonymous GB300 marketplace listings in the dated snapshot (2026-11-21); the sourcing desk secures GB300 capacity on request via /clusters/#request. Marketplace listings are provider-anonymous by design. Request reserved GB300 capacity through the cluster sourcing desk or hello@omegagradient.com . ## Monthly cost math (720 hours) | Scenario | $/GPU-hr | × 720 hrs | | Lowest tracked public rate | $15.01 | $10807 | | Median provider-average rate | $16.50 | $11884 | | Highest tracked public rate | $18.00 | $12960 | 720 hours ≈ one month of continuous single-GPU runtime. Multiply by GPU count for multi-GPU nodes. Put the data to work Reserved vs. on-demand ↗ Reserved clusters ↗ On-demand marketplace ↗ ## Frequently asked questions ### How much does a GB300 NVL72 cost per hour in 2026? Public on-demand rates tracked by Omega Gradient range from $15.01 to $18.00 per GPU-hour across 2 providers, with a median provider-average rate of $16.50 (report assembled 2026-11-21). ### Was a GB300 NVL72 listed in the catalog snapshot? No anonymous GB300 marketplace listings in the dated snapshot (2026-11-21); the sourcing desk secures GB300 capacity on request via /clusters/#request. ### What does a GB300 NVL72 cost per month? At the median tracked rate of $16.50/GPU-hr, one GPU running a full 720-hour month costs about $11884. At the lowest tracked rate of $15.01/GPU-hr, about $10807. ### Where do these prices come from? Included public rate observations collected by Omega Gradient; dates can vary by source. Provider-normalized average: average SKUs/offers within each provider and GPU family first, then average provider means so one marketplace does not dominate. ---------------------------------------------------------------------- # GPU Availability Snapshot (2026): 136 Sampled Listings From $0.07/hr | Omega Gradient URL: https://omegagradient.com/gpus/availability.html Report assembled: 2026-11-21T17:02:24.332Z Cloud GPU Price Index · Catalog Snapshot # GPU availability snapshot 136 GPU listings across 34 models were listed in the 2026-11-21 catalog snapshot, from $0.07/GPU-hr (a GTX 1070). Listings are provider-anonymous; confirm availability in the marketplace before deploying. Catalog snapshot 2026-11-21T17:00:36.649Z Dated catalog sample. Availability and prices may have changed. Check the current marketplace before relying on a listing. Snapshot taken 2026-11-21T17:00:36.649Z; current view on /compute.html . ## B300 — 2 in snapshot, from $10.84/GPU-hr | GPU | Count | vRAM | Region | Fabric | $/GPU/hr | Instance $/hr | | | B300 | 8× | 269 GB | US · Central | PCIe | $10.84 | $86.70 | Check availability | | B300 | 2× | 269 GB | US · West | PCIe | $11.70 | $23.40 | Check availability | ## B200 — 6 in snapshot, from $7.24/GPU-hr | GPU | Count | vRAM | Region | Fabric | $/GPU/hr | Instance $/hr | | | B200 | 2× | 179 GB | US · West | NVLink | $7.24 | $14.47 | Check availability | | B200 | 1× | 179 GB | US · West | NVLink | $8.27 | $8.27 | Check availability | | B200 | 8× | 179 GB | US · Central | NVLink | $9.63 | $77.02 | Check availability | | B200 | 4× | 179 GB | US · Central | NVLink | $9.63 | $38.52 | Check availability | | B200 | 4× | 179 GB | US · West | NVLink | $9.63 | $38.54 | Check availability | | B200 | 2× | 179 GB | US · Central | NVLink | $11.01 | $22.02 | Check availability | ## GB200 — 3 in snapshot, from $10.34/GPU-hr | GPU | Count | vRAM | Region | Fabric | $/GPU/hr | Instance $/hr | | | GB200 | 4× | 185 GB | Asia · Southeast | NVLink | $10.34 | $41.35 | Check availability | | GB200 | 2× | 185 GB | Asia · Southeast | NVLink | $10.36 | $20.72 | Check availability | | GB200 | 1× | 185 GB | Asia · Southeast | NVLink | $10.41 | $10.41 | Check availability | ## H200 — 16 in snapshot, from $3.50/GPU-hr | GPU | Count | vRAM | Region | Fabric | $/GPU/hr | Instance $/hr | | | H200 | 2× | 140 GB | US · Central | NVLink | $3.50 | $7.01 | Check availability | | H200 | 4× | 140 GB | Europe · East | NVLink | $3.92 | $15.70 | Check availability | | H200 | 2× | 140 GB | Europe · East | NVLink | $3.94 | $7.89 | Check availability | | H200 | 1× | 140 GB | US · Central | PCIe | $3.98 | $3.98 | Check availability | | H200 | 1× | 140 GB | Europe · East | NVLink | $3.98 | $3.98 | Check availability | | H200 | 2× | 140 GB | Europe · South | NVLink | $3.98 | $7.97 | Check availability | | H200 | 1× | 140 GB | Europe · South | NVLink | $4.01 | $4.01 | Check availability | | H200 | 4× | 140 GB | US · Central | NVLink | $4.33 | $17.33 | Check availability | | H200 | 8× | 141 GB | Global · flexible | SXM | $4.40 | $35.21 | Check availability | | H200 | 2× | 140 GB | Europe · West | PCIe | $4.41 | $8.82 | Check availability | | H200 | 1× | 140 GB | Asia · Northeast | PCIe | $4.42 | $4.42 | Check availability | | H200 | 1× | 140 GB | Global · flexible | PCIe | $4.73 | $4.73 | Check availability | | H200 | 8× | 140 GB | Asia · Northeast | PCIe | $5.43 | $43.44 | Check availability | | H200 | 4× | 140 GB | Asia · Northeast | PCIe | $5.43 | $21.73 | Check availability | | H200 | 8× | 140 GB | Europe · South | NVLink | $5.88 | $47.00 | Check availability | | H200 | 8× | 140 GB | Middle East | NVLink | $7.51 | $60.09 | Check availability | ## H100 — 11 in snapshot, from $2.87/GPU-hr | GPU | Count | vRAM | Region | Fabric | $/GPU/hr | Instance $/hr | | | H100 | 1× | 80 GB | Africa | PCIe | $2.87 | $2.87 | Check availability | | H100 | 1× | 80 GB | Europe · West | PCIe | $3.32 | $3.32 | Check availability | | H100 | 2× | 80 GB | Europe · West | PCIe | $3.63 | $7.27 | Check availability | | H100 | 1× | 80 GB | Europe · East | PCIe | $3.64 | $3.64 | Check availability | | H100 | 4× | 80 GB | Asia · Northeast | PCIe | $3.67 | $14.69 | Check availability | | H100 | 2× | 80 GB | Asia · Northeast | PCIe | $3.68 | $7.36 | Check availability | | H100 | 1× | 94 GB | US · Central | PCIe | $3.76 | $3.76 | Check availability | | H100 | 1× | 94 GB | Asia · Northeast | PCIe | $4.02 | $4.02 | Check availability | | H100 | 2× | 80 GB | Europe · East | NVLink | $5.04 | $10.09 | Check availability | | H100 | 1× | 80 GB | Global · flexible | SXM | $5.97 | $5.97 | Check availability | | H100 | 1× | 80 GB | US · East | PCIe | $6.59 | $6.59 | Check availability | ## A100-80G — 8 in snapshot, from $1.52/GPU-hr | GPU | Count | vRAM | Region | Fabric | $/GPU/hr | Instance $/hr | | | A100-80G | 1× | 80 GB | US · Central | SXM | $1.52 | $1.52 | Check availability | | A100-80G | 1× | 80 GB | Global · flexible | SXM | $2.04 | $2.04 | Check availability | | A100-80G | 2× | 80 GB | Global · flexible | SXM | $2.04 | $4.09 | Check availability | | A100-80G | 4× | 80 GB | US · Central | PCIe | $2.20 | $8.81 | Check availability | | A100-80G | 2× | 80 GB | US · Central | PCIe | $2.21 | $4.41 | Check availability | | A100-80G | 8× | 80 GB | US · Central | SXM | $3.07 | $24.56 | Check availability | | A100-80G | 8× | 80 GB | US · East | SXM | $3.07 | $24.56 | Check availability | | A100-80G | 1× | 80 GB | US · East | SXM | $3.61 | $3.61 | Check availability | ## A100 — 14 in snapshot, from $0.60/GPU-hr | GPU | Count | vRAM | Region | Fabric | $/GPU/hr | Instance $/hr | | | A100 | 1× | 40 GB | US · Central | PCIe | $0.60 | $0.60 | Check availability | | A100 | 2× | 40 GB | Asia · Northeast | PCIe | $0.60 | $1.21 | Check availability | | A100 | 2× | 40 GB | US · West | NVLink | $0.61 | $1.22 | Check availability | | A100 | 2× | 40 GB | US · Central | PCIe | $0.74 | $1.49 | Check availability | | A100 | 1× | 40 GB | Asia · Northeast | PCIe | $0.77 | $0.77 | Check availability | | A100 | 4× | 80 GB | Europe · West | PCIe | $0.96 | $3.84 | Check availability | | A100 | 2× | 80 GB | Europe · East | NVLink | $1.32 | $2.64 | Check availability | | A100 | 1× | 80 GB | Europe · East | NVLink | $1.33 | $1.33 | Check availability | | A100 | 1× | 40 GB | Global · flexible | SXM | $1.54 | $1.54 | Check availability | | A100 | 8× | 40 GB | Global · flexible | SXM | $1.54 | $12.32 | Check availability | | A100 | 1× | 80 GB | US · East | NVLink | $1.71 | $1.71 | Check availability | | A100 | 1× | 40 GB | India | SXM | $2.19 | $2.19 | Check availability | | A100 | 1× | 40 GB | US · West | SXM | $2.19 | $2.19 | Check availability | | A100 | 8× | 40 GB | US · West | PCIe | $2.19 | $17.52 | Check availability | ## RTX PRO 6000 — 13 in snapshot, from $1.39/GPU-hr | GPU | Count | vRAM | Region | Fabric | $/GPU/hr | Instance $/hr | | | RTX PRO 6000 | 1× | 96 GB | Europe · South | PCIe | $1.39 | $1.39 | Check availability | | RTX PRO 6000 | 1× | 96 GB | Asia · Northeast | PCIe | $1.42 | $1.42 | Check availability | | RTX PRO 6000 | 2× | 96 GB | US · West | PCIe | $1.47 | $2.94 | Check availability | | RTX PRO 6000 | 1× | 96 GB | US · West | PCIe | $1.51 | $1.51 | Check availability | | RTX PRO 6000 | 1× | 96 GB | US · Central | PCIe | $1.54 | $1.54 | Check availability | | RTX PRO 6000 | 2× | 96 GB | Global · flexible | PCIe | $1.59 | $3.18 | Check availability | | RTX PRO 6000 | 1× | 96 GB | Global · flexible | PCIe | $1.64 | $1.64 | Check availability | | RTX PRO 6000 | 1× | 96 GB | Asia · Southeast | PCIe | $1.64 | $1.64 | Check availability | | RTX PRO 6000 | 1× | 96 GB | Europe · West | PCIe | $1.75 | $1.75 | Check availability | | RTX PRO 6000 | 1× | 96 GB | Europe · North | PCIe | $1.83 | $1.83 | Check availability | | RTX PRO 6000 | 1× | 96 GB | US · East | PCIe | $1.97 | $1.97 | Check availability | | RTX PRO 6000 | 1× | 96 GB | Europe · East | PCIe | $2.06 | $2.06 | Check availability | | RTX PRO 6000 | 2× | 96 GB | US · Central | PCIe | $2.41 | $4.82 | Check availability | ## RTX 6000 Ada — 8 in snapshot, from $0.71/GPU-hr | GPU | Count | vRAM | Region | Fabric | $/GPU/hr | Instance $/hr | | | RTX 6000 Ada | 1× | 48 GB | Asia · Northeast | PCIe | $0.71 | $0.71 | Check availability | | RTX 6000 Ada | 2× | 48 GB | Europe · East | PCIe | $0.72 | $1.45 | Check availability | | RTX 6000 Ada | 2× | 48 GB | Asia · Southeast | PCIe | $0.74 | $1.48 | Check availability | | RTX 6000 Ada | 1× | 48 GB | Oceania | PCIe | $0.76 | $0.76 | Check availability | | RTX 6000 Ada | 2× | 48 GB | US · Central | PCIe | $0.78 | $1.55 | Check availability | | RTX 6000 Ada | 1× | 48 GB | US · Central | PCIe | $0.78 | $0.78 | Check availability | | RTX 6000 Ada | 1× | 48 GB | Europe · North | PCIe | $0.92 | $0.92 | Check availability | | RTX 6000 Ada | 4× | 48 GB | US · Central | PCIe | $1.07 | $4.27 | Check availability | ## L40S — 7 in snapshot, from $0.54/GPU-hr | GPU | Count | vRAM | Region | Fabric | $/GPU/hr | Instance $/hr | | | L40S | 1× | 45 GB | US · East | PCIe | $0.54 | $0.54 | Check availability | | L40S | 1× | 48 GB | US · Central | PCIe | $0.97 | $0.97 | Check availability | | L40S | 2× | 48 GB | US · Central | PCIe | $0.97 | $1.94 | Check availability | | L40S | 4× | 48 GB | US · Central | PCIe | $0.97 | $3.88 | Check availability | | L40S | 8× | 48 GB | US · Central | PCIe | $1.65 | $13.20 | Check availability | | L40S | 10× | 48 GB | US · Central | PCIe | $1.65 | $16.51 | Check availability | | L40S | 1× | 48 GB | Europe · East | PCIe | $1.88 | $1.88 | Check availability | ## RTX 5090 — 1 in snapshot, from $0.71/GPU-hr | GPU | Count | vRAM | Region | Fabric | $/GPU/hr | Instance $/hr | | | RTX 5090 | 8× | 32 GB | Global · flexible | PCIe | $0.71 | $5.72 | Check availability | ## RTX 4090 — 2 in snapshot, from $0.44/GPU-hr | GPU | Count | vRAM | Region | Fabric | $/GPU/hr | Instance $/hr | | | RTX 4090 | 8× | 24 GB | US · Central | — | $0.44 | $3.52 | Check availability | | RTX 4090 | 1× | 24 GB | Global · flexible | PCIe | $0.66 | $0.66 | Check availability | ## L40 — 3 in snapshot, from $0.97/GPU-hr | GPU | Count | vRAM | Region | Fabric | $/GPU/hr | Instance $/hr | | | L40 | 1× | 48 GB | US · Central | PCIe | $0.97 | $0.97 | Check availability | | L40 | 2× | 48 GB | US · Central | PCIe | $1.09 | $2.18 | Check availability | | L40 | 1× | 48 GB | Canada | PCIe | $1.11 | $1.11 | Check availability | ## RTX A6000 — 7 in snapshot, from $0.56/GPU-hr | GPU | Count | vRAM | Region | Fabric | $/GPU/hr | Instance $/hr | | | RTX A6000 | 1× | 48 GB | Canada | PCIe | $0.56 | $0.56 | Check availability | | RTX A6000 | 4× | 48 GB | US · Central | PCIe | $0.63 | $2.51 | Check availability | | RTX A6000 | 1× | 48 GB | US · Central | PCIe | $0.63 | $0.63 | Check availability | | RTX A6000 | 2× | 48 GB | US · Central | PCIe | $0.63 | $1.26 | Check availability | | RTX A6000 | 2× | 48 GB | US · East | PCIe | $2.13 | $4.25 | Check availability | | RTX A6000 | 4× | 48 GB | US · East | PCIe | $2.13 | $8.50 | Check availability | | RTX A6000 | 1× | 48 GB | US · East | PCIe | $2.13 | $2.13 | Check availability | ## A10 — 2 in snapshot, from $1.42/GPU-hr | GPU | Count | vRAM | Region | Fabric | $/GPU/hr | Instance $/hr | | | A10 | 1× | 24 GB | US · East | PCIe | $1.42 | $1.42 | Check availability | | A10 | 1× | 24 GB | US · West | PCIe | $1.42 | $1.42 | Check availability | ## L4 — 3 in snapshot, from $1.05/GPU-hr | GPU | Count | vRAM | Region | Fabric | $/GPU/hr | Instance $/hr | | | L4 | 1× | 24 GB | Europe · West | PCIe | $1.05 | $1.05 | Check availability | | L4 | 1× | 24 GB | Europe · East | PCIe | $1.05 | $1.05 | Check availability | | L4 | 2× | 24 GB | Europe · East | PCIe | $1.05 | $2.10 | Check availability | ## RTX A5000 — 3 in snapshot, from $1.56/GPU-hr | GPU | Count | vRAM | Region | Fabric | $/GPU/hr | Instance $/hr | | | RTX A5000 | 4× | 24 GB | US · East | PCIe | $1.56 | $6.25 | Check availability | | RTX A5000 | 2× | 24 GB | US · East | PCIe | $1.56 | $3.13 | Check availability | | RTX A5000 | 1× | 24 GB | US · East | PCIe | $1.57 | $1.57 | Check availability | ## RTX A4000 — 4 in snapshot, from $0.23/GPU-hr | GPU | Count | vRAM | Region | Fabric | $/GPU/hr | Instance $/hr | | | RTX A4000 | 1× | 16 GB | Europe · North | PCIe | $0.23 | $0.23 | Check availability | | RTX A4000 | 1× | 16 GB | US · East | PCIe | $0.88 | $0.88 | Check availability | | RTX A4000 | 2× | 16 GB | US · East | PCIe | $0.88 | $1.76 | Check availability | | RTX A4000 | 4× | 16 GB | US · East | PCIe | $0.88 | $3.52 | Check availability | ## V100 — 4 in snapshot, from $0.13/GPU-hr | GPU | Count | vRAM | Region | Fabric | $/GPU/hr | Instance $/hr | | | V100 | 2× | 32 GB | US · West | PCIe | $0.13 | $0.26 | Check availability | | V100 | 2× | 32 GB | US · East | PCIe | $2.58 | $5.15 | Check availability | | V100 | 4× | 32 GB | US · East | PCIe | $2.58 | $10.30 | Check availability | | V100 | 1× | 32 GB | US · East | PCIe | $2.58 | $2.58 | Check availability | ## A40 — 3 in snapshot, from $2.05/GPU-hr | GPU | Count | vRAM | Region | Fabric | $/GPU/hr | Instance $/hr | | | A40 | 4× | 48 GB | Europe · West | PCIe | $2.05 | $8.19 | Check availability | | A40 | 1× | 48 GB | Europe · West | PCIe | $2.05 | $2.05 | Check availability | | A40 | 1× | 48 GB | US · West | PCIe | $2.05 | $2.05 | Check availability | ## GAUDI2 — 1 in snapshot, from $1.21/GPU-hr | GPU | Count | vRAM | Region | Fabric | $/GPU/hr | Instance $/hr | | | GAUDI2 | 8× | 96 GB | US · Central | PCIe | $1.21 | $9.69 | Check availability | ## GTX 1070 — 1 in snapshot, from $0.07/GPU-hr | GPU | Count | vRAM | Region | Fabric | $/GPU/hr | Instance $/hr | | | GTX 1070 | 2× | 8 GB | Asia · Northeast | PCIe | $0.07 | $0.13 | Check availability | ## GTX 1080 TI — 2 in snapshot, from $0.09/GPU-hr | GPU | Count | vRAM | Region | Fabric | $/GPU/hr | Instance $/hr | | | GTX 1080 TI | 2× | 11 GB | Asia · Northeast | PCIe | $0.09 | $0.17 | Check availability | | GTX 1080 TI | 1× | 11 GB | Asia · Northeast | PCIe | $0.10 | $0.10 | Check availability | ## P100 — 1 in snapshot, from $0.08/GPU-hr | GPU | Count | vRAM | Region | Fabric | $/GPU/hr | Instance $/hr | | | P100 | 4× | 12 GB | Global · flexible | PCIe | $0.08 | $0.33 | Check availability | ## Q RTX 8000 — 1 in snapshot, from $0.30/GPU-hr | GPU | Count | vRAM | Region | Fabric | $/GPU/hr | Instance $/hr | | | Q RTX 8000 | 1× | 48 GB | US · East | PCIe | $0.30 | $0.30 | Check availability | ## QUADRO P4000 — 2 in snapshot, from $0.09/GPU-hr | GPU | Count | vRAM | Region | Fabric | $/GPU/hr | Instance $/hr | | | QUADRO P4000 | 2× | 8 GB | Europe · East | PCIe | $0.09 | $0.17 | Check availability | | QUADRO P4000 | 1× | 8 GB | Europe · East | PCIe | $0.10 | $0.10 | Check availability | ## RTX 2060 — 1 in snapshot, from $0.08/GPU-hr | GPU | Count | vRAM | Region | Fabric | $/GPU/hr | Instance $/hr | | | RTX 2060 | 1× | 12 GB | Asia · Northeast | PCIe | $0.08 | $0.08 | Check availability | ## RTX 2070S — 1 in snapshot, from $0.09/GPU-hr | GPU | Count | vRAM | Region | Fabric | $/GPU/hr | Instance $/hr | | | RTX 2070S | 1× | 8 GB | Asia · Northeast | PCIe | $0.09 | $0.09 | Check availability | ## RTX 3080 — 1 in snapshot, from $0.15/GPU-hr | GPU | Count | vRAM | Region | Fabric | $/GPU/hr | Instance $/hr | | | RTX 3080 | 1× | 10 GB | Europe · North | PCIe | $0.15 | $0.15 | Check availability | ## RTX 3080 TI — 1 in snapshot, from $0.32/GPU-hr | GPU | Count | vRAM | Region | Fabric | $/GPU/hr | Instance $/hr | | | RTX 3080 TI | 1× | 12 GB | US · Central | PCIe | $0.32 | $0.32 | Check availability | ## RTX 5060 TI — 1 in snapshot, from $0.24/GPU-hr | GPU | Count | vRAM | Region | Fabric | $/GPU/hr | Instance $/hr | | | RTX 5060 TI | 1× | 16 GB | Europe · West | PCIe | $0.24 | $0.24 | Check availability | ## RTX 5070 — 1 in snapshot, from $0.23/GPU-hr | GPU | Count | vRAM | Region | Fabric | $/GPU/hr | Instance $/hr | | | RTX 5070 | 1× | 12 GB | US · Central | PCIe | $0.23 | $0.23 | Check availability | ## RTX 5070 TI — 1 in snapshot, from $0.24/GPU-hr | GPU | Count | vRAM | Region | Fabric | $/GPU/hr | Instance $/hr | | | RTX 5070 TI | 1× | 16 GB | Asia · Southeast | PCIe | $0.24 | $0.24 | Check availability | ## RTX6000 — 1 in snapshot, from $0.76/GPU-hr | GPU | Count | vRAM | Region | Fabric | $/GPU/hr | Instance $/hr | | | RTX6000 | 1× | 24 GB | US · Central | PCIe | $0.76 | $0.76 | Check availability | ## Monthly cost math (720 hours) | Scenario | $/GPU-hr | × 720 hrs | | Cheapest sampled listing (GTX 1070) | $0.07 | $47 | | Median sampled listing rate | $1.61 | $1163 | | Highest sampled listing rate | $11.70 | $8424 | 720 hours ≈ one month of continuous single-GPU runtime. Multiply by GPU count for multi-GPU nodes. Choose the right access model Reserved clusters ↗ Reserved vs. on-demand ↗ On-demand marketplace ↗ ## Frequently asked questions ### How many listings were in this catalog snapshot? 136 GPU listings across 34 models were observed in the 2026-11-21 snapshot, from $0.07 to $11.70 per GPU-hour. This page is a fixed snapshot. Check the marketplace for current stock and pricing. ### What was the cheapest GPU in this catalog snapshot? The lowest-priced listing in this dated snapshot is a GTX 1070 at $0.07/GPU-hour. The median rate across all 136 sampled listings is $1.61/GPU-hour. ### What does a sampled listing cost per month? At the cheapest sampled rate of $0.07/GPU-hr, one GPU running a full 720-hour month costs about $47. At the median sampled rate of $1.61/GPU-hr, about $1163. ### Why are the listings provider-anonymous? Omega Gradient aggregates capacity from many underlying clouds and data centers into one marketplace with one bill and one console. Listings show verified specs, region, and price without naming the operator; provider identities are withheld until qualification on reserved cluster deals.