Data center GPU price: published rates, dates and sources
Data center GPU prices with sources and dates: H100 and B200 rental rates from AWS, Lambda and the SemiAnalysis index, plus where no price is published.
In this article7

What data center GPUs cost right now
Three published figures answer the question as of 23 September 2026, each from a different price layer. AWS EC2 Capacity Blocks for ML lists the NVIDIA H100 at $5.191 per accelerator-hour in US regions and the B200 at $12.355 per accelerator-hour, with new reservation rates announced effective 7 October 2026[1]. Lambda lists an on-demand H100 SXM instance at $3.99 per GPU-hour and a B200 SXM6 at $6.69 per GPU-hour[2]. The SemiAnalysis H100 spot-contract composite index, a weighted mean of observed rental prices across the market, stood at $2.82 per GPU-hour in April 2026[3].
The spread between those numbers is not a contradiction. It reflects three distinct price layers, and a defensible figure names which layer it comes from. The accelerator itself moves the number too: the H100 is no longer the top of the stack, and the B200 and H200 carry different rates on every venue that rents them.
List price, street price and contract price
- List price: the on-demand rate a provider publishes on its own pricing page, such as Lambda's $3.99 per H100 GPU-hour[2].
- Street price: what market indices observe buyers actually paying. The SemiAnalysis H100 spot-contract composite index is built from hyperscale clouds, neoclouds, marketplaces and rental aggregators, weighted toward hyperscalers.
- Contract price: reserved or committed terms. AWS Capacity Blocks carry a reservation fee plus an operating system fee, and reservation prices are updated regularly with supply and demand.
Over the past year the contract layer moved sharply upward. SemiAnalysis reports H100 one-year rental contract pricing rising almost 40 percent, from a low of $1.70 per GPU-hour in October 2025 to $2.35 per GPU-hour by March 2026, with on-demand rental capacity sold out across all GPU types in early 2026[4]. Those are point figures from the launch article; the same firm's index history publishes the one-year term as a 25th-75th percentile range, which ran $1.45-2.00 in November 2025 and widened to $2.10-2.70 in April 2026, with the on-demand H100 row marked sold out from February through April 2026[3]. The two starting figures are not in conflict: one is a survey-based point estimate, the other a percentile band from the scraped-market index.
Prices per accelerator across named vendors
Most accelerator vendors publish no list price at all, so prices surface mainly through cloud rental venues: even the strongest GPU in the world is priced by the venue that rents it, not by its maker. In the table below, a dash means the vendor's page lists no value.
| Vendor | Accelerator named on its page | Published price or price basis | Retrieval date |
|---|---|---|---|
| AMD | Instinct accelerators | – | 23 September 2026 |
| Axelera AI | Edge AI inference accelerators | – | 23 September 2026 |
| Cerebras Systems | Wafer-scale accelerators, inference service | – | 23 September 2026 |
| Etched | Transformer-specific ASICs | – | 23 September 2026 |
| FuriosaAI | AI inference accelerators | – | 23 September 2026 |
| Graphcore | IPU accelerators | – | 23 September 2026 |
| Groq | LPU inference accelerators, inference service | – | 23 September 2026 |
| Huawei | Ascend AI accelerators | – | 23 September 2026 |
| Intel | Gaudi accelerators | – | 23 September 2026 |
| NVIDIA | GPUs and AI accelerators | – | 23 September 2026 |
Where no public price exists
Where a number does exist, it belongs to a venue rather than a vendor. Google Cloud prices GPUs on its accelerator-optimized machine-family page, which lists 8-GPU H100 machine types such as the a3-megagpu-8g[5], while the general VM pricing page carries the discount structure: Spot discounts and committed-use discounts on most machine types (retrieved 23 September 2026).
Budgeting a cluster from published figures
A defensible cluster number states its measurement basis explicitly: specification sheets, MLPerf results and shipping dates, never a vendor projection restated as an independent finding. It also names the workload behind any performance figure, because GPU throughput is limited by data movement rather than raw compute. The recurring error on this topic is a peak figure quoted without the workload that produced it: FlashAttention (arXiv, 2022) reports a 3x training speedup on GPT-2 at sequence length 1K and a 15% end-to-end wall-clock speedup on BERT-large against the MLPerf 1.1 training speed record[6].
Combine dated rental tiers with index ranges rather than single points. Lambda's 1-Click Clusters price an H100 at $6.16 per GPU-hour for 16 GPUs, $5.85 at 64 and $5.54 at 256, showing the scale discount a single spot figure hides[2]. The SemiAnalysis one-year H100 contract range of $2.10-2.70 per GPU-hour in April 2026 is a different layer from the same page's spot-contract composite index level of $2.82[3].
- State the unit beside every figure: per GPU-hour, per instance-hour or per accelerator-hour.
- State the retrieval date and the source page beside every figure.
- Name the price layer: list, street or contract.
- Name the workload behind any performance or price-performance claim.
- Use ranges from a published index for contract terms, and posted rates only for the layer they describe.
Frequently asked questions
How much does an NVIDIA H100 or B200 cost right now, and how have AI GPU prices moved over the past year?
On AWS EC2 Capacity Blocks an H100 rents at $5.191 per accelerator-hour in US regions and a B200 at $12.355, with new rates announced effective 7 October 2026. Lambda lists an on-demand H100 SXM at $3.99 and a B200 SXM6 at $6.69 per GPU-hour. Over the past year, H100 1-year rental contracts rose from $1.70 to $2.35 per GPU-hour between October 2025 and March 2026 per SemiAnalysis, and on-demand H100 capacity was marked sold out on the SemiAnalysis index from February to April 2026.
Why do GPU prices differ so much between providers?
Because the quoted figures are often different products: a per-accelerator reservation, a per-GPU-hour on-demand VM and per-token serverless inference are never combined into one number. Region also moves the price: the same AWS H100 Capacity Block is $5.191 per accelerator-hour in US regions and $4.720 in London, Stockholm, Tokyo and others.
How do I budget a GPU cluster from published figures?
Take a dated per-GPU-hour rate, multiply by cluster size and term, and state the source page and retrieval date beside the result. Lambda's 1-Click Clusters price an H100 at $6.16 per GPU-hour for 16 GPUs, $5.85 at 64 and $5.54 at 256, so cluster size changes the unit price. Cross-check against an index range: H100 1-year contracts traded at $2.10 to $2.70 per GPU-hour in April 2026 per SemiAnalysis.
Sources
Rankings and scores on Optinest cannot be bought — no commercial relationship changes how a provider is assessed. How we score
Related reading
All Accelerators →1 H100 GPU cost: rental rates by route, with dates
What 1 H100 GPU costs per hour on demand, in reserved blocks and in contract indexes, with every rate's source and date, read 26 October 2026.

GPU failure rate: data center statistics and node reliability
Discover the real-world failure rate of data center GPUs based on published cluster studies, from H100 hardware interruptions to the impact on training MTTF.

Batch OCR on serverless GPUs: cost and throughput
Price a 10,000-page batch OCR run two ways, per token and per GPU-hour, from published European rate cards, with the crossover shown.
