ChatGPT 
Summary
Key takeaways
- The Uvik Software GPU Price Index tracks public on-demand rental prices for 13 GPU models across 19 providers.
- NVIDIA H100 SXM pricing varies significantly by provider, with a median on-demand price of $4.20 per GPU-hour and a range from $3.20 to $12.29 in October 2026.
- Specialized GPU clouds are generally cheaper than hyperscalers, with hyperscaler median prices running about 1.7 to 2.7 times higher for comparable GPUs.
- The NVIDIA H200 costs more per hour than the H100 but provides substantially more GPU memory, making it more economical for memory-intensive workloads.
- AMD MI300X offers particularly strong memory economics, with the lowest cost per GB-hour among GPUs with enough provider coverage for comparison.
- NVIDIA B200 and B300 command higher hourly prices, but their much higher throughput can make them competitive on cost per generated token.
- List prices and actual market prices can differ substantially, especially for large deployments, long commitments, and negotiated capacity.
- Three-year commitments with major cloud providers can cut H100 pricing by more than half compared with their public on-demand rates.
- GPU rental prices changed direction during 2026, with several verified increases following significant price reductions during 2025.
- GPU selection should be based on workload economics, memory requirements, throughput, networking, availability, and commitment terms rather than hourly price alone.
When this applies
This GPU price index applies when you are budgeting AI infrastructure, comparing GPU cloud providers, selecting hardware for model training or inference, or estimating the economics of running production AI workloads. It is especially useful when comparing NVIDIA H100, H200, B200, B300, A100, L40S, and AMD MI300X capacity across specialized GPU clouds and major hyperscalers. The data can also help engineering and finance teams decide whether on-demand capacity, committed contracts, or alternative GPU architectures provide the best value for a specific workload.
When this does not apply
This index should not be treated as a guaranteed quote for a specific deployment. Public on-demand prices do not include every discount, negotiated contract, spot instance, marketplace offer, tax, networking charge, storage cost, or data-transfer expense. It is also less useful when purchasing physical GPUs, because NVIDIA does not publish consistent public list prices for many data-center accelerators and hardware acquisition costs vary by system configuration, distributor, volume, and availability. For large or long-running workloads, direct provider quotations should supplement the public prices in the index.
Checklist
- Define whether the workload is training, fine-tuning, batch inference, or real-time inference.
- Estimate the amount of GPU memory required by the model and expected context size.
- Identify the GPU models that can run the workload without excessive model partitioning.
- Compare median on-demand prices rather than relying on a single provider quote.
- Check the full price range because the same GPU can vary significantly between providers.
- Compare specialized GPU clouds with AWS, Microsoft Azure, Google Cloud, and Oracle Cloud.
- Calculate cost per GB of GPU memory for memory-intensive workloads.
- Estimate cost per token or per completed workload when throughput benchmarks are available.
- Check the minimum number of GPUs that must be rented for each instance type.
- Review GPU-to-GPU networking when training or serving models across multiple accelerators.
- Confirm capacity availability in the regions where the application must operate.
- Include storage, data transfer, networking, and integration costs in the total infrastructure estimate.
- Compare on-demand pricing with committed-use and long-term contract options.
- Ask several providers for negotiated quotes before committing to a large deployment.
- Re-check current GPU prices immediately before procurement because market pricing can change quickly.
Common pitfalls
- Choosing the cheapest GPU-hour without considering actual model throughput.
- Comparing instance prices without dividing the price by the number of GPUs included in the instance.
- Assuming hyperscaler on-demand pricing represents the broader GPU rental market.
- Ignoring GPU memory capacity when comparing H100, H200, B200, and alternative accelerators.
- Using list prices as the expected cost for a large or long-term deployment without requesting negotiated quotes.
- Ignoring networking performance for multi-GPU training and distributed inference workloads.
- Comparing GPUs only by compute specifications instead of cost per useful unit such as tokens, memory, or completed jobs.
- Forgetting additional costs such as storage, data transfer, support, and cloud integration.
- Assuming prices consistently decline over time despite recent periods of rising GPU rental rates.
- Using an old GPU pricing comparison without verifying current provider prices and capacity before making a purchasing decision.
Short answer: how much does an NVIDIA H100 cost? To rent, an H100 SXM costs a median $4.20 per GPU-hour on demand across 16 providers (7 October 2026). Prices run from $3.20 to $12.29. To buy, NVIDIA publishes no list price, and published estimates do not agree, so this index tracks rental prices.
Key findings (Uvik Software GPU Price Index, October 2026)
- The NVIDIA H100 SXM rents for a median $4.20 per GPU-hour on demand across 16 providers. The lowest list price is $3.20 and the highest is $12.29, a 3.8x spread.
- Hyperscalers (AWS, Microsoft Azure, Google Cloud and Oracle Cloud) charge 1.7 to 2.7 times the GPU-cloud price per GPU-hour for the same GPU, median against median. The gap is largest for the H100.
- Median on-demand prices per GPU-hour: L40S $1.62, A100 80GB $2.75, AMD MI300X $3.45, H100 $4.20, H200 $5.70, B200 $8.19 and B300 $9.50.
- The H200 costs 36% more per hour than the H100 but has 76% more memory. Per GB of GPU memory, the H200 is 23% cheaper.
- Among the GPUs with 3 or more providers, the AMD MI300X has the cheapest GPU memory: $0.018 per GB-hour, about one third of the H100 price per GB ($0.053).
- With MLPerf v6.1 results, one million output tokens of Llama 2 70B costs about $0.18 on a rented B200 and $0.19 on a B300, at median prices and best-case throughput.
- Prices are rising in 2026. The SemiAnalysis H100 one-year contract index rose 38% from October 2025 to March 2026, and Nebius raised its on-demand H100 price from $3.85 to $4.50 per GPU-hour on 1 October 2026.
Get the index data. Download all 87 prices with the provider, product, GPU count and source link for each one, as CSV or JSON. You can reuse the data with credit to Uvik Software (CC BY 4.0).
Download the GPU Price Index (CSV) | JSON feed
GPU Price Index, October 2026: median prices
The Uvik Software GPU Price Index tracks on-demand list prices for 13 GPU models from 19 providers. On 7 October 2026, the median price per GPU-hour was $1.62 for the L40S, $2.75 for the A100 80GB, $4.20 for the H100 SXM, $5.70 for the H200, $8.19 for the B200 and $9.50 for the B300.
Figure 1. Uvik Software GPU Price Index, October 2026: median on-demand price per GPU-hour and the range from the lowest to the highest provider price.
| GPU | Median per GPU-hour | Lowest | Highest | Providers |
|---|---|---|---|---|
| NVIDIA L40S | $1.62 | $1.09 | $3.50 | 10 |
| NVIDIA A100 80GB | $2.75 | $1.09 | $5.07 | 14 |
| NVIDIA H100 PCIe | $2.89 | $2.50 | $2.99 | 3 |
| AMD Instinct MI300X | $3.45 | $2.59 | $6.00 | 5 |
| NVIDIA H100 SXM | $4.20 | $3.20 | $12.29 | 16 |
| NVIDIA H200 | $5.70 | $3.99 | $10.60 | 14 |
| NVIDIA B200 | $8.19 | $6.00 | $14.24 | 9 |
| NVIDIA B300 | $9.50 | $7.40 | $17.80 | 7 |
| NVIDIA GB200 (small sample) | $16.00 | $10.50 | $27.04 | 3 |
How to read the index: these are public on-demand list prices for one GPU-hour, in US dollars, without tax. When a provider sells only 8-GPU or 4-GPU instances, we divide the instance price by the number of GPUs. Committed contracts, spot capacity and marketplaces are cheaper. We do not publish a median for GPUs with fewer than 3 providers (GB300, H100 NVL, MI325X and MI355X), but their prices are in the CSV.
Where can you rent an H200 or a B200? 14 providers in the index offer the H200 on demand: AWS, Microsoft Azure, Google Cloud, Oracle Cloud, CoreWeave, Nebius, Together AI, RunPod, Jarvislabs, Hyperstack, Crusoe, DigitalOcean, OVHcloud and Verda. 9 offer the B200: AWS, Oracle Cloud, CoreWeave, Lambda, Nebius, Together AI, RunPod, Hyperstack and Verda. 7 offer the B300: AWS, Oracle Cloud, Nebius, Together AI, RunPod, Hyperstack and Verda.
NVIDIA H100 price per hour: 16 providers compared
On 7 October 2026, the on-demand price of one H100 SXM GPU-hour ran from $3.20 at Hyperstack and Thunder Compute to $12.29 at Microsoft Azure. The median across 16 providers was $4.20. The four hyperscalers had a median of $10.53, and the 12 GPU clouds a median of $3.95.
Figure 2. NVIDIA H100 price per hour: on-demand list price per GPU-hour at 16 providers, October 2026. Source: Uvik Software GPU Price Index.
| Provider | Product | Type | Per GPU-hour | Note |
|---|---|---|---|---|
| Hyperstack | H100 SXM | GPU cloud | $3.20 | |
| Thunder Compute | H100 SXM | GPU cloud | $3.20 | |
| RunPod | H100 SXM (Secure Cloud) | GPU cloud | $3.49 | |
| Jarvislabs | H100 SXM | GPU cloud | $3.49 | |
| Verda | H100 SXM5 | GPU cloud | $3.81 | |
| Crusoe | H100 HGX | GPU cloud | $3.90 | |
| Lambda | H100 SXM (8 GPUs) | GPU cloud | $3.99 | Plus tax |
| Together AI | HGX H100 | GPU cloud | $3.99 | |
| DigitalOcean | HGX H100 | GPU cloud | $4.41 | |
| Nebius | HGX H100 | GPU cloud | $4.50 | Price from 1 October 2026 (was $3.85) |
| Paperspace | H100 | GPU cloud | $5.95 | Labeled as a special promotion price |
| CoreWeave | HGX H100 | GPU cloud | $6.16 | $49.24 per 8-GPU instance |
| AWS | p5.48xlarge | Hyperscaler | $6.88 | $55.04 per 8-GPU instance |
| Oracle Cloud | BM.GPU.H100.8 | Hyperscaler | $10.00 | |
| Google Cloud | a3-highgpu-8g | Hyperscaler | $11.06 | $88.49 per 8-GPU instance |
| Microsoft Azure | ND96isr H100 v5 | Hyperscaler | $12.29 | $98.32 per 8-GPU instance |
Where can you rent an H100? All 16 providers in the table sell H100 capacity on demand. Before you choose, check four things: the minimum number of GPUs per order, the network between GPUs (for training), the region and the support level. For a full list of costs in an AI project, see our AI development cost guide.
Hyperscalers vs GPU clouds: the price gap
Hyperscalers charge 1.7 to 2.7 times the GPU-cloud price per GPU-hour for the same GPU, median against median. The gap is largest for the H100: $10.53 against $3.95. Three-year commitments at AWS, Google Cloud and Microsoft Azure cut H100 prices by 56% to 57%.
Figure 3. Hyperscaler GPUs cost 1.7 to 2.7 times the GPU-cloud price: median price per GPU-hour at hyperscalers vs GPU clouds, October 2026. Source: Uvik Software GPU Price Index.
| GPU | Hyperscaler median (providers) | GPU cloud median (providers) | Hyperscaler premium |
|---|---|---|---|
| NVIDIA L40S | $2.68 (2) | $1.57 (8) | 1.7x |
| NVIDIA A100 80GB | $4.10 (3) | $2.30 (11) | 1.8x |
| NVIDIA H100 SXM | $10.53 (4) | $3.95 (12) | 2.7x |
| NVIDIA H200 | $10.30 (4) | $4.78 (10) | 2.2x |
| NVIDIA B200 | $14.12 (2) | $7.20 (7) | 2.0x |
| NVIDIA B300 | $16.40 (2) | $9.06 (5) | 1.8x |
| AMD MI300X | $6.00 (2) | $2.99 (3) | 2.0x |
Hyperscalers give large discounts for commitments. The three-year price of an 8-GPU H100 instance is 57% below on demand at AWS, and 56% below at Google Cloud and Microsoft Azure. Per GPU-hour, these commitments cost $2.97 (AWS), $4.86 (Google Cloud) and $5.40 (Azure). The AWS three-year price is 25% below the GPU-cloud on-demand median, but you pay for three years. If your data and services already run on a hyperscaler, include data transfer and integration costs in the comparison.
List price vs market price: what other GPU indexes show
List prices are where a negotiation starts. Market indexes show lower H100 prices: $2.81 per GPU-hour (Silicon Data, 7 October 2026) and $2.86 (Ornn, an index built from executed trades, 6 October 2026), against the Uvik Software on-demand list median of $4.20.
| Index | What it measures | H100 value | Date |
|---|---|---|---|
| Uvik Software GPU Price Index | Median on-demand list price, 16 providers | $4.20 | 7 October 2026 |
| Silicon Data SDH100RT | H100 rental price index | $2.81 | 7 October 2026 |
| Ornn OCPI | Index built from executed trades (H200 $5.04, B200 $7.56, A100 $0.90) | $2.86 | 6 October 2026 settlement |
| SemiAnalysis | H100 composite rental price | $2.82 | April 2026 (latest public value) |
Use list prices for budgets and for short jobs. For large or long jobs, ask several providers for committed-use quotes. Each index value above comes from the publisher’s own site.
GPU rental price trend: 2025 to 2026
GPU rental prices fell in 2025 and rose in 2026. AWS cut its on-demand H100 (P5) price by 44% in June 2025. Then the SemiAnalysis H100 one-year contract index rose from $1.70 to $2.35 per GPU-hour (October 2025 to March 2026), and Nebius raised its on-demand prices by 17% to 21% on 1 October 2026.
Figure 4. GPU rental prices fell in 2025 and rose in 2026: verified price events, June 2025 to October 2026.
| Date | Event | Source |
|---|---|---|
| 1 June 2025 (announced 5 June) | AWS cuts on-demand prices: P5 (H100) by 44%, P5en (H200) by 25%, P4d and P4de (A100) by 33% | AWS News Blog |
| October 2025 to March 2026 | SemiAnalysis H100 one-year contract index rises from $1.70 to $2.35 per GPU-hour (+38%) | SemiAnalysis |
| 1 October 2026 | Nebius raises on-demand prices: H100 from $3.85 to $4.50 (+17%), H200 from $4.50 to $5.40 (+20%) | Nebius prices |
| 7 October 2026 | AWS raises EC2 Capacity Blocks prices by about 15%. On-demand prices do not change. | AWS EC2 Capacity Blocks pricing |
The verified events show a change in direction: prices fell in 2025 and rose in 2026. Re-check prices before each large order: price changes of 17% to 44% happened in the last 16 months.
H100 vs H200 vs B200: specs and price per GB of memory
The H200 has the same compute as the H100 but 141 GB of memory instead of 80 GB. The B200 has 180 GB and 2.3x the dense FP8 compute of the H100. Per GB of GPU memory, the AMD MI300X is the cheapest GPU in this table at $0.018 per GB-hour, and the H100 is the most expensive at $0.053.
| GPU | Memory | Memory bandwidth | Dense FP8 | Maximum power | Median per GPU-hour | Per GB-hour |
|---|---|---|---|---|---|---|
| NVIDIA L40S | 48 GB GDDR6 | 864 GB/s | 733 TFLOPS | 350 W | $1.62 | $0.034 |
| NVIDIA A100 80GB SXM | 80 GB HBM2e | 2,039 GB/s | No FP8 (BF16: 312 TFLOPS) | 400 W | $2.75 | $0.034 |
| NVIDIA H100 SXM | 80 GB | 3.35 TB/s | 1,979 TFLOPS | Up to 700 W | $4.20 | $0.053 |
| NVIDIA H200 SXM | 141 GB | 4.8 TB/s | 1,979 TFLOPS | Up to 700 W | $5.70 | $0.040 |
| NVIDIA B200 (HGX, per GPU) | 180 GB HBM3e | 8 TB/s | 4.5 PFLOPS | Not on the product page | $8.19 | $0.046 |
| AMD Instinct MI300X | 192 GB HBM3 | 5.3 TB/s | 2.61 PFLOPS | 750 W | $3.45 | $0.018 |
NVIDIA publishes Tensor Core figures with sparsity. The dense values in this table are half of those figures, as NVIDIA’s footnotes state. B200 values per GPU are calculated from the 8-GPU HGX and DGX totals.
- H100 vs H200. The H200 costs 36% more per hour ($5.70 against $4.20). It has 76% more memory and 43% more memory bandwidth, with the same compute. Choose the H200 for large models and long contexts that do not fit in 80 GB.
- H100 vs B200. The B200 costs 95% more per hour ($8.19 against $4.20). It has 2.3x the dense FP8 compute, 2.25x the memory and 2.4x the memory bandwidth.
- B200 vs H200. The B200 costs 44% more per hour, with 28% more memory, 67% more bandwidth and 2.3x the dense FP8 compute.
- B200 vs B300. The B300 costs 16% more per hour ($9.50 against $8.19). In MLPerf v6.1 (Llama 2 70B, Offline), it produced 11% more tokens per second per GPU, so the cost per token is almost the same. NVIDIA lists the HGX B300 at 108 PFLOPS of dense FP4 per 8 GPUs, against 72 for the HGX B200, but with much lower INT8 and FP64 performance.
- H100 vs A100. The H100 costs 53% more per hour ($4.20 against $2.75). It adds FP8 support and has 64% more memory bandwidth.
What a million tokens costs on Blackwell GPUs
At median rental prices and best-case MLPerf v6.1 throughput, one million output tokens of Llama 2 70B costs about $0.18 on a B200, $0.19 on a B300 and $0.36 on a GB200. For DeepSeek-R1, it costs about $0.31 on a B200 or a B300 and $0.61 on a GB200 (Uvik Software calculation).
Figure 5. What a million tokens costs on Blackwell GPUs: median GPU-hour price divided by MLPerf v6.1 Offline throughput per GPU. Sources: Uvik Software GPU Price Index; MLPerf Inference v6.1.
Formula: cost per 1 million tokens = (price per GPU-hour x 1,000,000) / (tokens per second per GPU x 3,600)
| Model (MLPerf v6.1, Offline) | GPU (submitted system) | Tokens per second per GPU | Median per GPU-hour | Per 1M output tokens |
|---|---|---|---|---|
| Llama 2 70B | B200 (CoreWeave, 8 GPUs) | 12,838 | $8.19 | $0.18 |
| Llama 2 70B | B300 (Cisco, 16 GPUs) | 14,223 | $9.50 | $0.19 |
| Llama 2 70B | GB200 (CoreWeave NVL72, 72 GPUs) | 12,417 | $16.00 | $0.36 |
| DeepSeek-R1 | B200 (CoreWeave, 8 GPUs) | 7,364 | $8.19 | $0.31 |
| DeepSeek-R1 | B300 (Cisco, 16 GPUs) | 8,567 | $9.50 | $0.31 |
| DeepSeek-R1 | GB200 (Azure NVL72, 288 GPUs) | 7,261 | $16.00 | $0.61 |
Read these numbers as a floor. MLPerf Offline is a best-case batch test, and a real service does not keep every GPU busy all the time. At 50% utilization, the cost per token doubles. The GB200 median uses only 3 providers: at the CoreWeave GB200 price ($10.50 per GPU-hour), Llama 2 70B costs about $0.23 per million tokens. Use the calculator below with your own throughput and utilization.
GPU cost per million tokens
Estimate GPU rental cost per one million output tokens using your throughput and utilization assumptions.
Loading GPU price index…
Manual price for selected GPU.
The estimate divides the GPU-hour price by output tokens produced in an hour at the entered throughput and utilization. It excludes other serving costs.
Estimated cost per million output tokens
Enter valid inputs to calculate.
Source: Uvik Software GPU Price Index.
MLPerf Inference v6.1 Offline reference throughput
Output tokens per second per GPU. Results depend on workload and system and are reference assumptions, not a production-performance guarantee. GB200 is shown for context only and is not an available priced calculator option.
| Model | GPU | Submitted system | Output tokens/sec/GPU |
|---|---|---|---|
| Llama 2 70B | B200 | CoreWeave, 8 GPUs | 12,838 |
| Llama 2 70B | B300 | Cisco, 16 GPUs | 14,223 |
| Llama 2 70B | GB200 | CoreWeave NVL72, 72 GPUs | 12,417 |
| DeepSeek-R1 | B200 | CoreWeave, 8 GPUs | 7,364 |
| DeepSeek-R1 | B300 | Cisco, 16 GPUs | 8,567 |
| DeepSeek-R1 | GB200 | Azure NVL72, 288 GPUs | 7,261 |
How much does it cost to buy an NVIDIA GPU?
NVIDIA and AMD do not publish list prices for data center GPUs. NVIDIA CEO Jensen Huang told CNBC in March 2024 that a Blackwell GPU, the family that includes the B200, would cost $30,000 to $40,000. Published estimates for the H100 do not agree with each other, so this report does not quote one.
| GPU | Reported purchase price | Source | Status |
|---|---|---|---|
| NVIDIA Blackwell GPU (B200 generation) | $30,000 to $40,000 per GPU | Jensen Huang to CNBC, 19 March 2024; also reported by Tom’s Hardware and Club386 | Verified for “Blackwell”. NVIDIA later said prices vary because it sells complete systems. |
| NVIDIA H100 | No single verified figure | Estimates differ by source and date | Not quoted |
| H200, B300, GB200 NVL72, GB300 NVL72, AMD MI300X, MI355X | No verified figure | Fewer than three matching sources | Not quoted |
In practice, few buyers purchase single GPUs. NVIDIA sells HGX boards and DGX systems, and server makers sell complete 8-GPU servers. A DGX B200 system with 8 GPUs draws up to about 14.3 kW, according to NVIDIA. Power, cooling, networking and staff add to the cost of each GPU.
Rent or buy? The break-even point
At the October 2026 median B200 rental price of $8.19 per GPU-hour, a Blackwell GPU that costs $30,000 to $40,000 pays for itself after 3,663 to 4,884 GPU-hours: 5.0 to 6.7 months of full use. This is a GPU-only lower bound. A server, networking, power, cooling and staff make ownership more expensive (Uvik Software calculation).
Buy GPUs only if all three statements are true:
- You can keep the GPUs busy most of the time for several years.
- You have data center space with enough power and cooling. One 8-GPU B200 system draws up to about 14.3 kW.
- You have engineers who can run and repair a GPU cluster.
If one statement is false, rent the capacity and negotiate a commitment. Use the calculator below with your own quote.
GPU rent vs buy
Compare ownership cost per used GPU-hour with GPU rental using your quoted costs and utilization.
Loading GPU price index…
Manual price for selected GPU.
Energy is charged during used hours and hosting throughout ownership. The rental comparison assumes billing only for used GPU-hours. Staff, financing, networking (unless included in your quote), and resale value are excluded.
Ownership vs rental cost
Enter valid inputs to calculate.
Source: Uvik Software GPU Price Index.
How to cut your GPU bill: six rules
- Choose the GPU by memory first. For inference of large models, memory is often the limit. Compare the price per GB-hour, not only the price per hour.
- Compare GPU clouds and hyperscalers. The same GPU costs 1.7 to 2.7 times as much at a hyperscaler.
- Commit when the load is steady. Three-year commitments at AWS, Google Cloud and Microsoft Azure cut H100 prices by 56% to 57%.
- Keep the GPUs busy. Batch your requests and use queues. At 50% utilization, the cost per token doubles.
- Measure the cost per token or per task, not the cost per hour. A faster GPU at a higher price can be cheaper per token.
- Check prices every month. Price changes of 17% to 44% happened in the last 16 months.
If you do not want to own a GPU bill at all, Uvik Software AI delivery pods are priced by accepted deliverables, and Uvik Software absorbs the inference cost. For build costs, see AI development cost. For model quality, see our AI hallucination rates report, because each extra evaluation run also costs GPU time.
Get AI results without the GPU bill. Uvik Software AI delivery pods are priced by accepted deliverables, and Uvik Software absorbs the inference cost. If you run your own models, our engineers size your GPU needs, compare rent and commit options and tune inference to cut your cost per token.
See AI delivery pods | Talk to an LLM engineer
Methodology
- What we track. Public on-demand list prices in US dollars for Linux instances, without tax. We captured the prices on 7 October 2026 from the price pages and price APIs of 19 providers.
- Price per GPU-hour. We use the per-GPU price when the provider shows one. Otherwise, we divide the instance price by the number of GPUs.
- One price per provider and GPU. For RunPod, we use Secure Cloud prices. For Nebius, we use the price that took effect on 1 October 2026. For hyperscalers, we use a US region: AWS us-east-1, Azure East US (East US 2 for H200, H100 NVL and MI300X) and Google Cloud us-central1.
- What we exclude. Spot and preemptible prices, marketplaces, serverless per-second prices, contract-only prices, reserved blocks and “starting at” prices that do not name a GPU.
- Medians. We publish a median only for GPUs with 3 or more providers. Hyperscalers are AWS, Microsoft Azure, Google Cloud and Oracle Cloud. All other providers are GPU clouds. The split by provider type uses smaller groups (2 to 12 providers), and the table shows the count for each group.
- Hidden prices. When a page shows prices only through JavaScript, we use the official price API (Azure Retail Prices API, Oracle price list API) or a public mirror of the AWS price list.
- Cost per token. MLPerf Inference v6.1 Closed division, Offline scenario, as published by NVIDIA (16 September 2026). We divide system throughput by the number of accelerators.
- Update cycle. Monthly, in the first week. Next update: 7 November 2026.
- Limits. List prices change often, sometimes daily. Many providers give lower prices to large buyers.
How to cite this index
Uvik Software (2026). GPU Price Index 2026: NVIDIA H100, H200, B200 and AMD MI300X prices to rent and buy. Prices captured on 7 October 2026. https://uvik.net/blog/gpu-price-index/
You can reuse the charts and the data under CC BY 4.0. Please credit “Uvik Software” with a link to this page.
Sources
- Uvik Software. GPU Price Index, October 2026: 87 on-demand prices from 19 providers, with the source URL for each price. https://uvik.net/wp-content/uploads/2026/10/uvik-software-gpu-price-index-2026-10.csv
- Provider price pages and APIs, read on 7 October 2026: AWS (EC2 on-demand list prices), Azure Retail Prices API, Google Cloud accelerator-optimized pricing, Oracle Cloud price list, CoreWeave, Lambda, Nebius, Crusoe, Together AI, RunPod, Jarvislabs, Hyperstack, DigitalOcean, Paperspace, Verda, Thunder Compute, Vultr, OVHcloud and Hot Aisle (URLs in the CSV).
- AWS News Blog. Announcing up to 45% price reduction for Amazon EC2 NVIDIA GPU-accelerated instances, 5 June 2025. https://aws.amazon.com/blogs/aws/announcing-up-to-45-price-reduction-for-amazon-ec2-nvidia-gpu-accelerated-instances/
- DatacenterDynamics. AWS cuts costs for H100, H200 and A100 instances by up to 45%, 9 June 2025. https://www.datacenterdynamics.com/en/news/aws-cuts-costs-for-h100-h200-and-a100-instances-by-up-to-45/
- Network World. AWS cuts prices of some EC2 Nvidia GPU-accelerated instances, 9 June 2025. https://www.networkworld.com/article/4003551/aws-cuts-prices-of-some-ec2-nvidia-gpu-accelerated-instances.html
- SemiAnalysis. GPU Pricing Index. https://semianalysis.com/gpu-pricing-index/
- Nebius. Prices, read on 7 October 2026. https://nebius.com/prices
- Silicon Data. H100 rental index (SDH100RT), read on 7 October 2026. https://www.silicondata.com/products/silicon-index/h100
- Ornn. Compute price index markets, read on 7 October 2026. https://data.ornn.com/markets
- NVIDIA. MLPerf Inference v6.1 results (data center). https://developer.nvidia.com/deep-learning-performance-training-inference/ai-inference
- MLCommons. MLPerf Inference: Datacenter results. https://mlcommons.org/benchmarks/inference-datacenter/
- NVIDIA. H100, H200, HGX and DGX B200 product pages. https://www.nvidia.com/en-us/data-center/h100/ · https://www.nvidia.com/en-us/data-center/h200/ · https://www.nvidia.com/en-us/data-center/hgx/ · https://www.nvidia.com/en-us/data-center/dgx-b200/
- AMD. Instinct MI300X accelerator. https://www.amd.com/en/products/accelerators/instinct/mi300/mi300x.html
- CNBC (via NBC San Diego). Nvidia’s latest AI chip will cost more than $30,000, CEO says, 19 March 2024. https://nbcsandiego.com/news/business/money-report/nvidias-latest-ai-chip-will-cost-more-than-30000-ceo-says/3465447
- TechRadar. Nvidia’s fastest AI chip ever could cost a rather reasonable $40,000, March 2024. https://www.techradar.com/pro/nvidias-fastest-ai-chip-ever-could-cost-a-rather-reasonable-dollar40000-but-chances-that-you-will-actually-be-able-to-buy-one-on-its-own-are-very-very-low-and-for-a-good-reason
- Tom’s Hardware. Nvidia’s Jensen Huang says Blackwell GPU to cost $30,000 to $40,000, later clarifies that pricing will vary, March 2024. https://www.tomshardware.com/pc-components/gpus/nvidias-jensen-huang-says-blackwell-gpu-to-cost-dollar30000-dollar40000-later-clarifies-that-pricing-will-vary-as-they-wont-sell-just-the-chip
- Club386. Nvidia Blackwell B200 GPU is a snip at between $30K and $40K, March 2024. https://www.club386.com/nvidia-blackwell-b200-gpu-is-a-snip-at-between-30k-and-40k
Get senior AI and ML engineers. Uvik Software provides senior Python and AI engineers: more than 50 engineers, no juniors, and at least 7 years of experience each. You get matched profiles 48 hours after the SOW and a 30-day no-cost replacement. Rates: $50 to $99 per hour. Uvik Software is a member of the Claude Partner Network and a Databricks Bronze partner.
Request matched profiles | Hire AI and ML developers
