Why Are GPU Prices Rising in 2026?
GPU rental rates rose at AWS, Nebius and Lambda this year, and the reason is memory rather than silicon. Here is what moved, why the trackers disagree about how much, what the increase adds to a monthly cluster bill, and whether self-hosting still beats the API.
GPU prices are rising because memory is, not because silicon is. DRAM revenue rose 369% year over year in Q3 2026 on about 4% more bits shipped, and an AI accelerator carries 80GB to 180GB of that memory.
Published rate cards moved 15% to 20%, and market medians moved further. Lambda raised H100 and B200 about 15% effective November 16, Nebius raised 17% to 20% on October 1, and AWS raised Capacity Blocks about 15% on October 7.
Quoted H100 rates range from under $1 to over $11, and the spread is methodology. Contract sits below on-demand, per-GPU sits below per-node, and GetDeploying’s own two pages report a 13% and a 32% twelve-month change for the same GPU.
The same increase hits a contract buyer twice as hard in percentage terms. An eight-GPU H100 cluster running continuously gained $3,796 a month on both Nebius list and a one-year contract, which is 17% on one base and 38% on the other.
Self-hosting got weaker this year, not stronger. GPU rental rose while budget-tier API prices fell, and closed models like Claude and GPT cannot be self-hosted at any volume, so the comparison only exists for open-weight models.
What is actually pushing GPU rental rates up, what it does to a monthly bill, and whether self-hosting still beats the API.
GPU prices are rising because the memory inside the GPU got dramatically more expensive, not because the silicon did. DRAM revenue grew 369% year over year in the third quarter of 2026 while the industry shipped only about 4% more bits, according to Counterpoint Research. That is a price story rather than a volume story, and it lands on every accelerator built with high bandwidth memory.
GPU prices have moved across the market with it. GetDeploying’s on-demand price index put the median H100 at $3.96 per GPU-hour in the week of November 9, 2026, up 32% over twelve months, and SemiAnalysis tracked one-year H100 contracts rising from $1.70 an hour in October 2025 to $2.35 by March 2026. Nebius raised its published rates on October 1, and Lambda’s new rate card takes effect on November 16
1. How much have GPU rental prices risen in 2026?
Published rate cards moved roughly 15% to 20% this year, concentrated in the newest accelerators, while market-wide medians rose further. The moves below come from each provider’s own pricing page.
| Provider | GPU | Move | When |
| Lambda | H100 (8x instance) | $3.99 to $4.59 per GPU-hour, up 15% | Effective Nov 16, 2026 |
| Lambda | B200 (8x instance) | $6.69 to $7.69 per GPU-hour, up 15% | Effective Nov 16, 2026 |
| Nebius | H100 | $3.85 to $4.50 per GPU-hour, up 17% | Oct 1, 2026 |
| Nebius | H200 | $4.50 to $5.40 per GPU-hour, up 20% | Oct 1, 2026 |
| Nebius | B200 | $7.15 to $8.50 per GPU-hour, up 19% | Oct 1, 2026 |
| AWS | Capacity Blocks, all accelerators | About 15% | Oct 7, 2026 |
AWS is the clearest pattern. Capacity Blocks reservation rates moved three times in 2026, and AWS states on its own pricing page that reservation prices are updated regularly based on supply and demand trends, with the next update scheduled for January 2027. That is a provider telling you in writing that the rate card is not a fixed thing.
Why the trackers disagree. Quoted H100 rates range from under $1 to over $11, and the spread is methodology rather than error. Contract rates sit below on-demand, so SemiAnalysis’s $2.35 one-year figure and a $4.50 on-demand list rate describe different purchases. An index that takes one median per provider weights a small provider the same as AWS. And some sources quote a single GPU while others quote an eight-GPU node, an eightfold difference on its own. GetDeploying’s own pages illustrate it: its H100 model page reported a 13% twelve-month change on November 12 while its price index reported 32% for the week of November 9, because the two are computed over different provider sets. Always check which series you are reading and what date it carries. Our GPU pricing tracker stays closer to live than any blog post can.
2. Which GPUs are rising, and where did the H100 price hold?
GPU prices rose on new, memory-heavy accelerators bought on demand. Older GPUs, spot capacity and some providers did not move at all, and the gap between hyperscalers and dedicated GPU clouds widened.
| GPU | GPU clouds | Hyperscalers | Difference |
| H100 | $4.50 | $7.89 | 75% |
| H200 | $5.40 | $10.85 | 101% |
| B200 | $8.55 | $14.24 | 67% |
| A100 | $2.00 | $3.67 | 84% |
| RTX PRO 6000 | $2.15 | $4.50 | 109% |
Median on-demand rates per GPU-hour from GetDeploying, where hyperscalers means AWS, Azure and Google Cloud.
Three places prices held. Older accelerators stayed flat. In the same repricing that lifted Lambda’s H100 and B200 about 15%, the A100 40GB stayed at $1.99 per GPU-hour and the V100 stayed at $0.79. Spot and preemptible capacity sits far below on demand. CoreWeave lists HGX H100 spot at $19.71 an hour against $49.24 on demand, about $2.46 per GPU rather than $6.16, and Nebius lists preemptible GPU-hours from $0.79 against $4.50 on demand. Some providers held. CoreWeave’s published HGX H100 node has stayed at $49.24.
3. Why are GPU prices rising?

GPU prices are rising for four reasons, and memory is the largest, with committed demand, the end of neocloud discounting and power constraints adding to it. An AI accelerator is substantially a memory device, so when memory doubles in price the bill of materials moves with it.
Memory. Counterpoint Research put DRAM revenue at just under $200 billion in the third quarter of 2026, up 369% year over year, on bit shipments that grew only about 4%. Blended DRAM sits near $1.90 per gigabit against the 20 to 30 cents it once sold for, and Counterpoint expects HBM4 to head toward $4. The benchmark 16Gb DDR5 contract price rose from $8.70 in October 2025 to $48 by September 2026, as reported by Fortune. This is not ending quickly. SK Group’s chairman said in March 2026 that the shortage will last until 2030, with wafer supply trailing demand by about 20%.
An H100 carries 80GB of HBM, an H200 carries 141GB and a B200 carries 180GB. That is why the increases cluster on newer parts while a V100 holds its rate.
Committed demand. Capacity bought in advance holds GPU prices up by removing spare supply. NVIDIA’s quarterly filing shows customer advances at $2.8 billion as of July 26, 2026, against $160 million six months earlier, and raw materials inventory at $11.3 billion against $3.8 billion. Buyers are paying earlier to hold a place in line, which removes the slack that used to push spot rates down.
The discounting phase ended. Providers that won customers on price are now raising it. Gartner forecast in October 2026 that cost pressure and a lack of a differentiating business model will force more than 80% of neocloud providers to cease independent operations by 2030, so the cheapest quote on the market is no longer a reliable guide to what you will pay next year.
Power as a pressure. The International Energy Agency reports data center electricity demand rose 17% in 2025 against 3% growth in global electricity demand, and describes AI deployment as running into physical bottlenecks including grid connection delays. This constrains how fast capacity can be added rather than setting rental rates directly, but it is why new supply has not arrived to compete the price down.
4. What does the increase do to your monthly bill?
An eight-GPU H100 cluster running continuously uses 5,840 GPU-hours a month. At Nebius list rates that bill went from $22,484 to $26,280 on October
The more useful comparison is what the same increase does to a contract buyer.
| Pricing basis | Old rate | New rate | Old monthly | New monthly | Change |
| Nebius on-demand list, Oct 2026 | $3.85 | $4.50 | $22,484 | $26,280 | +$3,796 (17%) |
| One-year contract, Oct 2025 to Mar 2026 | $1.70 | $2.35 | $9,928 | $13,724 | +$3,796 (38%) |

Both rates rose by $0.65 an hour, so both bills rise by $3,796 a month, about $45,500 a year. Our GPU pricing calculator runs the same math on your own GPU count, hours and utilization. The contract buyer pays the same extra dollars on a much smaller base, so the percentage more than doubles. Note that the two rows cover different periods: the Nebius move is a single October 2026 change from its rate card, while the contract figures are SemiAnalysis index values for October 2025 and March 2026.
5. Does self-hosting still beat the API?
For closed models there is no break-even at any volume, because you cannot self-host them. Claude and GPT models are available only through their provider’s API or resold inside a cloud, so renting H100s never replaces them. The question only applies to open-weight models such as Llama, Qwen and DeepSeek.
For those, the test is cost per million tokens at the throughput you actually achieve, not a token volume. A node has a ceiling, so adding volume means adding nodes rather than crossing a threshold.
| Node rate (8x H100) | At 15,000 tokens/sec | At 5,600 tokens/sec |
| $31.92 per hour, old Lambda rate | $0.59 per 1M | $1.58 per 1M |
| $36.72 per hour, new Lambda rate | $0.68 per 1M | $1.82 per 1M |
Figures assume 100% utilization and count GPU rent only, with no staff, storage or egress. At 70% utilization the new-rate numbers become $0.97 and $2.60. Published throughput for a 70B-class open model on an eight-GPU H100 node spans roughly 5,600 to 15,000 tokens per second depending on prompt length, batch size, quantization and engine, so measure your own.
Because the GPU rate sits on top of that fraction, a 15% rate rise moves your self-hosted cost per token by 15%. The direction of travel matters more than any single number: GPU rental rose through 2026 while budget-tier API prices fell, so the case for self-hosting a standard open model got weaker this year, not stronger. Self-hosting still pays for control, privacy, custom models or guaranteed capacity. If your inference load is steady, a dedicated GPU server on a fixed monthly price, such as OpenMetal’s H200 servers, takes the hourly rate card out of the calculation. Our AI token economics guide covers the API side of the same comparison.
6. What should you do before 2027?
Budget for GPU prices to rise again rather than return to 2025 rates, and date every number you carry forward. Two increases are already scheduled: AWS says its next Capacity Blocks update is due in January 2027, and Microsoft has announced virtual machine and storage increases effective February 1, 2027, reaching 17% in North Europe.
For instance-level detail across the three major clouds, see our cloud GPU pricing comparison, and our cloud cost optimization guide for the wider practice that keeps GPU prices from surprising you.
Frequently Asked Questions
Mainly because high bandwidth memory got much more expensive. DRAM revenue rose 369% year over year in the third quarter of 2026 on only about 4% more bits shipped, according to Counterpoint Research, and an AI accelerator carries 80GB to 180GB of that memory. Committed demand, the end of neocloud discounting and power constraints add to it.
AWS raised EC2 Capacity Blocks reservation rates about 15% on October 7, 2026, one of three increases during the year. AWS states that these rates are updated regularly based on supply and demand, and that the next scheduled update is January 2027.
Nothing in the supply picture points that way for the newest accelerators. SK Group’s chairman said in March 2026 that the memory shortage will last until 2030 with wafer supply trailing demand by about 20%, and AWS and Microsoft both have increases scheduled for early 2027. Older GPUs and interruptible capacity are where rates have stayed flat.
About 75% more for an H100, at median on-demand rates of $7.89 against $4.50 per GPU-hour, according to GetDeploying. The gap is wider on newer parts, reaching 101% on the H200. Hyperscaler rates buy proximity to your existing data, identity and compliance setup, which is worth paying for in some cases and not in others.
Only for open-weight models, since closed models like Claude and GPT cannot be self-hosted at any price. For an open model, a busy, well-tuned eight-GPU H100 node runs about $0.59 to $1.82 per million tokens on GPU rent alone, before staff, storage and idle time. Self-hosting usually wins on control, privacy or guaranteed capacity rather than on rate.
Connect your cloud accounts to a platform that reports GPU instance spend and AI API spend against the same dimensions. Economize covers AWS, Azure and GCP alongside OpenAI, Anthropic, Bedrock, Vertex AI and Azure AI Foundry, and you can start free up to $100,000 a month in tracked spend.
Simran Sardar
FinOps enthusiastProduct Manager at Economize with over 3 years of experience, focused on FinOps strategies and cloud cost optimization. Dedicated to helping organizations streamline cloud expenses and drive financial efficiency.
Maximize Cloud Efficiency and Optimize Costs
Get started free in our sandbox or book a personalized call with our experts






















