Last Updated: November 19, 2026
Summary: An Nvidia H200 rents for $3.59 to $10.60 per GPU-hour on the official price lists we checked, which works out to roughly $2,600 to $7,700 per GPU per month if it runs around the clock. Nvidia does not publish a list price for buying one, but Reuters reported in December 2025 that Nvidia planned to price H200 chips for Chinese buyers at around $27,000 each.
The spread is the story. The same chip costs nearly three times as much per hour on Google Cloud on-demand as on RunPod's community cloud, and AWS has raised its H200 reservation prices more than once in 2026. Which column you land in decides whether an H200 project costs $21,000 or $62,000 a month for one 8-GPU server.
Below are the rates from each provider's own pricing page, checked on November 18, 2026, plus the monthly math, an H100 vs H200 price comparison, and a simple rent-vs-buy break-even.
Free download: The AI Stats Sheet 2026. 36 key AI numbers for business on four printable pages, with the source under every figure. Join 2,000+ business leaders who read AI Business Weekly and get it in your inbox right away.
In this guide
How Much Does an H200 Cost per Hour?
An H200 costs between $3.59 and $10.60 per GPU-hour on the public price lists below. Specialist GPU clouds sit at the low end, and Google Cloud's on-demand rate sits at the top. Every figure comes from the provider's own pricing page.
Provider and option | Price as listed | Per GPU-hour | One GPU, 730 hours |
|---|---|---|---|
RunPod Community Cloud | From $3.59/hr per GPU | $3.59 | $2,621 |
RunPod Secure Cloud | $5.29/hr per GPU | $5.29 | $3,862 |
Nebius on-demand (HGX H200) | $5.40 per GPU-hour | $5.40 | $3,942 |
CoreWeave HGX H200, 8 GPUs | $50.44/hr per instance | $6.31 | $4,603 |
AWS p5e.48xlarge Capacity Block, 8 GPUs | $54.924/hr per instance | $6.87 | $5,012 |
AWS p5en.48xlarge Capacity Block, 8 GPUs | $63.158/hr per instance | $7.90 | $5,763 |
Google Cloud A3 Ultra, 8 GPUs, on-demand | $84.81/hr per machine | $10.60 | $7,739 |
Per-GPU figures for 8-GPU instances are the instance price divided by eight. Monthly figures assume 730 hours, which is 24 hours a day for an average month. Prices exclude tax, storage and data transfer.
Two caveats matter. RunPod's $3.59 is a "from" price on a marketplace of hosts, so availability and the exact rate move. Nebius also offers preemptible H200 capacity "from $0.79" per GPU-hour, but its pricing page says spot prices are dynamic and can change as often as every 15 minutes, so that rate is not a fair comparison for steady production work.

H200 rental rates vary almost threefold between the providers we checked. Photo: Getty Images / Unsplash+
What Does an 8-GPU H200 Server Cost to Rent per Month?
An 8-GPU H200 server costs about $21,000 to $62,000 a month to rent full time, depending on the provider. Large training and inference jobs often run on a full eight-GPU node, so this is the number a finance team should budget against.
Option | Hourly, 8 GPUs | Monthly (730 hrs) | Yearly (8,760 hrs) |
|---|---|---|---|
RunPod Community Cloud (8 x from $3.59) | $28.72 | $20,966 | $251,587 |
Nebius on-demand (8 x $5.40) | $43.20 | $31,536 | $378,432 |
CoreWeave HGX H200 | $50.44 | $36,821 | $441,854 |
AWS p5e.48xlarge Capacity Block | $54.92 | $40,095 | $481,134 |
AWS p5en.48xlarge Capacity Block | $63.16 | $46,105 | $553,264 |
Google Cloud A3 Ultra, 3-year commitment | $37.21 | $27,162 | $325,946 |
Google Cloud A3 Ultra, 1-year commitment | $58.47 | $42,685 | $512,214 |
Google Cloud A3 Ultra, on-demand | $84.81 | $61,909 | $742,909 |
Google Cloud is the clearest example of how much commitment changes the bill. The same A3 Ultra machine drops from $84.81 an hour on-demand to $37.21 on a 3-year commitment, a 56% cut, which moves Google from the most expensive option on this list to one of the cheapest. Google's page lists a spot price of $50.87 an hour for the machine, and it notes that spot prices can change up to once a day.
AWS Capacity Blocks work differently from normal on-demand instances. You reserve GPU capacity for a fixed window, and AWS charges the rate in force when you buy, even if the block starts after a price change. The AWS pricing page says its current prices are next scheduled to update in January 2027.
How Much Have AWS H200 Prices Risen in 2026?
AWS's H200 reservation prices are up sharply this year. The Register reported in January 2026 that AWS raised Capacity Block prices by about 15%, taking p5e.48xlarge from $34.61 to $39.80 an hour and p5en.48xlarge from $36.18 to $41.61. AI Weekly, citing The Information, later reported a further increase of about 20% from July 1, 2026 across the P5, P5e, P5en and Blackwell families.
AWS Capacity Block | Before Jan 2026 hike | After Jan 2026 hike | Price on Nov 18, 2026 | Change since start of 2026 |
|---|---|---|---|---|
p5e.48xlarge (8 x H200) | $34.61 | $39.80 | $54.924 | +59% |
p5en.48xlarge (8 x H200) | $36.18 | $41.61 | $63.158 | +75% |
The first two columns come from The Register's report, and the November figures come from the AWS Capacity Blocks pricing page. The percentage change is our calculation. AWS's own explanation is short: its page says reservation prices "are updated regularly based on trends in supply and demand."
For a company budgeting a year of H200 capacity on AWS, that gap is real money. One p5en.48xlarge node running all year cost about $317,000 at the pre-January rate and about $553,000 at today's rate. If you priced a project in late 2025, re-run the numbers before you sign.
How Much Does It Cost to Buy an H200?
Nvidia does not publish a retail price for the H200. The most cited figure comes from Reuters, which reported on December 31, 2025, citing sources, that Nvidia planned to price H200 chips for Chinese customers at around $27,000 each and that an eight-chip module was expected to cost around 1.5 million yuan. The same sources said pricing would vary with volume and customer arrangements, and Nvidia said only that it "continuously manages its supply chain."
That $27,000 is for the GPU alone. A working server also needs CPUs, system memory, networking, storage, power, cooling and data center space, and Nvidia sells the H200 through partners in HGX systems with 4 or 8 GPUs (the SXM version) or MGX systems with up to 8 PCIe cards (the NVL version). Reseller listings online vary widely by vendor and configuration, so get written quotes rather than trusting a single web price.
Rent vs buy: a simple break-even
Here is a rough way to compare. Take eight H200s at Reuters' reported $27,000 each, which is $216,000 for the GPUs alone. Divide by what you would pay to rent an 8-GPU node:
Rental option | 8-GPU hourly rate | Hours to spend $216,000 | Days at 24/7 |
|---|---|---|---|
Google Cloud A3 Ultra, on-demand | $84.81 | 2,547 | 106 |
AWS p5en.48xlarge Capacity Block | $63.16 | 3,420 | 142 |
CoreWeave HGX H200 | $50.44 | 4,282 | 178 |
Nebius on-demand | $43.20 | 5,000 | 208 |
Google Cloud A3 Ultra, 3-year commitment | $37.21 | 5,805 | 242 |
This flatters buying, because it ignores the rest of the server, power, staff and the fact that GPUs lose value as newer chips arrive. It still shows the basic rule: if you will keep eight GPUs busy almost every hour for more than a year, owning or a long commitment deserves a serious look. If your usage is bursty, a pilot, or uncertain, renting is cheaper and far less risky.
H100 vs H200 Price: Is the Upgrade Worth It?
On most clouds an H200 costs 15% to 33% more per hour than an H100, but it carries 76% more memory (141 GB vs 80 GB). On CoreWeave the premium is only 2%, and on Google Cloud on-demand the H200 machine is actually cheaper than the H100 A3 High machine.
Provider | H100 per GPU-hour | H200 per GPU-hour | H200 premium | H200 cost per GB of GPU memory per hour |
|---|---|---|---|---|
RunPod Community (from) | $2.69 | $3.59 | +33% | 2.5 cents (H100: 3.4 cents) |
RunPod Secure | $4.39 | $5.29 | +21% | 3.8 cents (H100: 5.5 cents) |
Nebius | $4.50 | $5.40 | +20% | 3.8 cents (H100: 5.6 cents) |
AWS Capacity Blocks (p5 vs p5e) | $5.970 | $6.866 | +15% | 4.9 cents (H100: 7.5 cents) |
CoreWeave (8-GPU instance / 8) | $6.16 | $6.31 | +2% | 4.5 cents (H100: 7.7 cents) |
Google Cloud on-demand (A3 High vs A3 Ultra) | $11.06 | $10.60 | -4% | 7.5 cents (H100: 13.8 cents) |
Memory is why the H200 exists. Nvidia's H200 page lists 141 GB of HBM3e and 4.8 TB/s of memory bandwidth, which it describes as nearly double the H100's capacity with 1.4 times the bandwidth. Nvidia also claims the H200 runs Llama 2 70B inference 1.9 times faster and GPT-3 175B inference 1.6 times faster than the H100. Those are vendor benchmarks, so treat them as a ceiling rather than a promise for your workload.

The H200's main upgrade over the H100 is memory: 141 GB versus 80 GB. Photo: Getty Images / Unsplash+
Worked example: one H200 or two H100s?
A model's weights take roughly 2 bytes per parameter at 16-bit precision and 1 byte at 8-bit. A 70-billion-parameter model served at 8-bit therefore needs about 70 GB just for weights, before the working memory each conversation uses.
On an 80 GB H100, that leaves about 10 GB of headroom, which limits how many users or how long a context you can serve at once. On a 141 GB H200 it leaves about 71 GB. If the tight fit pushes you to split the model across two H100s, the H200 is the cheaper choice: at Nebius, one H200 costs $5.40 an hour against $9.00 for two H100s, a saving of 40%, or about $2,600 per month per serving replica at 730 hours.
If your model fits comfortably on one H100 with room to spare, the H100 is usually the better buy. Pay for the H200 when memory is your bottleneck, not by default. For a broader look at how serving costs add up, see our guide to AI inference.
H200 vs B200: How Much More Does Blackwell Cost?
A B200 rents for about 36% to 107% more per hour than an H200 on the same provider. On Nebius the B200 is $8.50 per GPU-hour against $5.40 for the H200. CoreWeave lists an 8-GPU B200 instance at $68.80 an hour versus $50.44 for H200, and RunPod lists the B200 at $7.99 an hour.
Provider | H200 per GPU-hour | B200 per GPU-hour | B200 premium |
|---|---|---|---|
CoreWeave | $6.31 | $8.60 | +36% |
RunPod | $5.29 (Secure) | $7.99 | +51% |
Nebius | $5.40 | $8.50 | +57% |
AWS Capacity Blocks | $6.866 (p5e) | $14.208 (p6-b200) | +107% |
CoreWeave and RunPod list the B200 with 180 GB of memory, more than the H200's 141 GB. Whether the newer chip is cheaper per unit of work depends on how much faster it runs your specific model, so benchmark both on a short rental before committing to a long contract. Our breakdown of Nvidia's competitors and the TPU vs GPU comparison cover the non-Nvidia options, and our Nvidia AI chips guide covers the rest of the lineup.
Why Is the H200 So Expensive?
The H200 is expensive because demand for large-memory AI chips still outruns supply. Reuters reported in December 2025 that Chinese technology companies had ordered more than 2 million H200 chips for 2026 while Nvidia held about 700,000 units in stock, after the US decided to allow H200 sales to China with a 25% fee. Beijing had not yet approved shipments at the time.
The chip is also a big step up in memory and power. Each H200 carries 141 GB of HBM3e high-bandwidth memory, and Nvidia lists a configurable maximum power draw of up to 700 W for the SXM version, so the servers around it need serious power and cooling.
Cloud pricing reflects the same pressure. AWS ties its Capacity Block rates to supply and demand, and its H200 reservations cost far more now than at the start of the year. For the wider picture on where AI budgets are going, see our AI spending statistics and Nvidia statistics, or the latest on Nvidia's limits on neocloud financing, which affects many of the specialist clouds in the tables above.
How Should a Business Choose Where to Rent H200s?
Pick by how steady your workload is, then by where your data already lives:
Testing or short projects: start on a specialist cloud with per-GPU pricing, such as RunPod or Nebius, so you can rent one GPU instead of a full 8-GPU node.
Steady production on one cloud: if your data and security reviews already sit with Google Cloud or AWS, compare their committed or reserved prices, not on-demand. Google's 3-year A3 Ultra rate is less than half its on-demand rate.
Large reserved clusters: Nebius says customers can pay less than on-demand rates by reserving large clusters for multiple months, but its page lists no figures, so ask providers for written quotes.
Memory-bound models: choose H200 over H100 when your model or context length does not fit comfortably in 80 GB. Otherwise the H100 is usually cheaper.
Whichever you choose, price the full bill, including storage, networking and idle time. A GPU you reserve but leave idle costs the same as one doing useful work. Our AI pricing hub tracks what the major AI tools and APIs cost if renting GPUs turns out to be more than you need.
Frequently Asked Questions
How much does an H200 cost per hour?
An Nvidia H200 costs $3.59 to $10.60 per GPU-hour on the official price lists we checked in November 2026. RunPod's community cloud starts at $3.59, Nebius charges $5.40, CoreWeave works out to $6.31, AWS Capacity Blocks run $6.87 to $7.90, and Google Cloud on-demand is $10.60.
Why is the H200 so expensive?
Demand for large-memory AI chips still exceeds supply. Reuters reported that Chinese companies ordered more than 2 million H200s for 2026 against about 700,000 units in Nvidia's inventory, and each chip carries 141 GB of HBM3e high-bandwidth memory.
Is the H200 faster than the H100?
Nvidia says yes for large language models: it claims 1.9 times faster Llama 2 70B inference and 1.6 times faster GPT-3 175B inference than the H100. The gain comes mostly from memory, with 141 GB versus 80 GB and 1.4 times the bandwidth.
Is the H200 Nvidia's best chip?
No. Nvidia's newer Blackwell chips, such as the B200 and B300, come after it, and Reuters reported that Nvidia has been focused on ramping Blackwell and its upcoming Rubin line. The B200 costs roughly 36% to 107% more per hour to rent than the H200 on the providers we checked.
The Bottom Line
The H200's price depends far more on where you rent it than on the chip itself. The same GPU runs from $3.59 to $10.60 an hour, an 8-GPU node costs about $21,000 to $62,000 a month, and AWS's H200 reservations have climbed by more than half this year.
Start small on a per-GPU cloud, measure whether memory is really your bottleneck, and only then lock in a commitment. Check each provider's page on the day you buy, because these prices move.
Related Articles
Nvidia Statistics: revenue, market share and data center numbers.
Nvidia Competitors: AMD, Google, Amazon and the other chip challengers.
TPU vs GPU: how Google's chips compare with Nvidia's.
AI Inference Explained: what it costs to run AI models in production.
AI Spending Statistics: where corporate AI budgets are going.
⭐ Get more AI Business Weekly in Google: add us as a preferred source on Google.
Written by Sameer Khan, with support from the AI Business Weekly Editorial Team. AI-assisted and fact-checked against the sources linked above.
