GPU Prices 2026
By Axis Intelligence Research
Co-author: Sarah Mitchell | Last updated: October 2, 2026 | License: CC BY 4.0
Prices verified on October 2, 2026. Every price on this page was read from the vendor’s live pricing page, public price API or retail listing on that date. AWS Capacity Block rates effective October 7, 2026 are taken from AWS’s own announcement on its pricing page.
Renting one NVIDIA H100 for an hour costs $3.49 on Runpod and $12.29 on Microsoft Azure — the same chip, listed on public pricing pages on October 2, 2026. Axis Intelligence Research’s GPU Hour Premium Ratio puts the hyperscaler list premium at 2.93x the neocloud median, and AWS is raising every GPU Capacity Block rate by roughly 15% on October 7.
Quick Answer
How much does a GPU cost in October 2026? In the cloud, an NVIDIA H100 lists between $3.49 and $12.29 per GPU-hour on-demand; the neocloud median is $3.99. A B200 lists from $6.69 to $8.60. On the consumer side, the cheapest new RTX 5070 on Newegg was $869.99 against a $549 launch price — 58% above MSRP.
Key Findings
- According to Axis Intelligence Research, hyperscalers list the NVIDIA H100 at 2.93 times the neocloud median price — $11.68 versus $3.99 per GPU-hour on-demand, as of October 2, 2026 (GPU Hour Premium Ratio, GHPR).
- AWS raises EC2 Capacity Block GPU rates about 15% effective October 7, 2026, taking an H100 accelerator-hour from $5.191 to $5.970 and a B200 from $12.355 to $14.208, per AWS’s pricing page.
- An 8-GPU H100 server rented on-demand for three years costs $2.58 million on Azure and $734,000 on Runpod, according to Axis Intelligence Research’s list-price TCO calculation.
- The cheapest new GeForce RTX 5070 on Newegg cost $869.99 on October 2, 2026, 58% above the $549 price NVIDIA announced at CES in January 2025.
- Azure’s spot price for an 8x H100 VM is 82% below its own on-demand rate — $18.17 versus $98.32 per hour in US East, per Microsoft’s Retail Prices API.
How Much Does It Cost to Rent an H100 GPU per Hour?
The H100 is the reference chip for AI compute pricing, and in October 2026 its price depends far more on who rents it to you than on the silicon.
H100 on-demand price per GPU-hour by provider (October 2, 2026)
| Provider | Instance / SKU | Listed price | GPUs | Price per GPU-hour | Source |
|---|---|---|---|---|---|
| Runpod | H100 SXM Pod (Secure Cloud) | $3.49/hr | 1 | $3.49 | Runpod pricing |
| Lambda | NVIDIA H100 SXM instance | $3.99/GPU/hr | 1–8 | $3.99 | Lambda pricing |
| CoreWeave | NVIDIA HGX H100 (North America) | $49.24/hr | 8 | $6.16 | CoreWeave pricing |
| Google Cloud | a3-highgpu-8g, us-central1 | $88.49/hr | 8 | $11.06 | Google Cloud accelerator-optimized pricing |
| Microsoft Azure | ND96isr H100 v5, East US | $98.32/hr | 8 | $12.29 | Azure Retail Prices API |
Per-GPU-hour figures for 8-GPU instances are Axis Intelligence Research calculations: instance price ÷ 8. Google Cloud and Azure prices include the host vCPUs, memory and local SSD bundled into the VM; neocloud prices also bundle host resources, but in smaller amounts per GPU.
The spread from cheapest to most expensive is 3.5x ($12.29 ÷ $3.49). That is the number most “GPU price” pages quote, and it overstates the useful comparison, because Runpod’s $3.49 Pod and Azure’s $98.32 VM are not the same product. Azure sells an eight-GPU VM with its host CPU, memory and storage bundled in; a single Pod is a single GPU.
Sarah Mitchell, AI & Machine Learning: The honest comparison is node to node. CoreWeave’s 8-GPU HGX H100 node at $49.24 an hour against Azure’s 8-GPU VM at $98.32 is still a 2x gap for a like-for-like training box. That gap is not a typo and not a promo. It is what the hyperscalers charge for being where your data, your IAM policies and your enterprise agreement already live — and for most inference teams, that bundle is worth less than they’re paying for it.
What does AWS charge for an H100?
AWS’s per-accelerator figure that we verified is the EC2 Capacity Blocks for ML rate — a reservation you book for a fixed window, not open-ended on-demand. On October 2, 2026 the p5.48xlarge (8x H100) Capacity Block rate in US East (N. Virginia) was $41.528 per instance-hour, or $5.191 per H100. AWS’s pricing page states that from October 7, 2026, the P5 rate becomes $5.970 per accelerator across all available regions. Axis Intelligence Research did not verify a current AWS on-demand p5 rate for this snapshot, so AWS is excluded from the on-demand table rather than estimated.
Why Is AWS Raising GPU Prices in October 2026?
AWS publishes the reason in one line: Capacity Block reservation prices are dynamic and are updated periodically based on supply and demand. The October 7, 2026 update moves every GPU family up by the same proportion.
AWS EC2 Capacity Block rate change, effective October 7, 2026
| Instance family | GPU | Rate before (per accelerator-hour) | Rate from Oct 7, 2026 | Change |
|---|---|---|---|---|
| P6-B300 | B300 | $14.040 | $16.146 | +15.0% |
| P6-B200 | B200 | $12.355 | $14.208 | +15.0% |
| P5en | H200 | $6.865 | $7.895 | +15.0% |
| P5e | H200 | $5.970 | $6.866 | +15.0% |
| P5 | H100 | $5.191 | $5.970 | +15.0% |
| P4de | A100 80GB | $2.214 | $2.546 | +15.0% |
| P4d | A100 40GB | $1.475 | $1.696 | +15.0% |
Source: AWS EC2 Capacity Blocks for ML pricing, retrieved October 2, 2026. “Before” rates are the US East (N. Virginia) or US-region rates listed on the same page; percentage change is an Axis Intelligence Research calculation.
A uniform 15% move across seven GPU generations — including a 2020-era A100 40GB — tells you this is not a scarcity signal about one chip. Capacity Blocks are AWS’s mechanism for guaranteed short-term cluster access, and a flat repricing across the whole catalog reads as AWS re-pricing guaranteed availability itself. Buyers who only need interruptible capacity should read this as a reason to look at spot and neocloud rates, not as proof that GPUs in general got 15% more expensive.
How Much Does a B200 or H200 Cost per Hour?
Blackwell and H200 per-GPU-hour prices (October 2, 2026)
| GPU | Provider | Pricing model | Price per GPU-hour | Source |
|---|---|---|---|---|
| B200 | Lambda | On-demand instance | $6.69 | Lambda pricing |
| B200 | Runpod | On-demand Pod | $6.79 | Runpod pricing |
| B200 | CoreWeave | On-demand, 8x HGX ($68.80/hr) | $8.60 | CoreWeave pricing |
| B200 | Lambda | 1-Click Cluster, 256+ GPUs, 2 weeks–1 year | $8.87 | Lambda pricing |
| B200 | Google Cloud | DWS Flex-start, a4-highgpu-8g ($64.44/hr) | $8.06 | Google Cloud pricing |
| B200 | AWS | Capacity Block, from Oct 7, 2026 | $14.21 | AWS pricing |
| B300 | Runpod | On-demand Pod | $7.89 | Runpod pricing |
| H200 | Runpod | On-demand Pod | $4.59 | Runpod pricing |
| H200 | CoreWeave | On-demand, 8x HGX ($50.44/hr) | $6.31 | CoreWeave pricing |
| H200 | Google Cloud | On-demand, a3-ultragpu-8g ($84.81/hr) | $10.60 | Google Cloud pricing |
One detail that most rate cards hide: Lambda’s B200 is cheaper on-demand ($6.69) than in its own 1-Click Clusters ($8.87–$9.86). The cluster price buys guaranteed multi-node capacity with interconnect, not a per-GPU discount. Google Cloud does not list an on-demand price for its A4 (B200) VM at all; the cheapest public path is DWS Flex-start at $64.44 per 8-GPU hour.
How Much Cheaper Are Spot and Reserved GPUs?
The list price is the ceiling. Most production GPU spend in 2026 is bought below it — through spot capacity, reservations or committed-use contracts.
H100 price per GPU-hour by commitment model (October 2, 2026)
| Provider | Model | Price per GPU-hour | vs. same provider on-demand |
|---|---|---|---|
| Azure (East US) | On-demand | $12.29 | — |
| Azure (East US) | 1-year reservation ($551,221 per VM) | $7.87 | −36% |
| Azure (East US) | 3-year reservation ($1,134,310 per VM) | $5.40 | −56% |
| Azure (East US) | Spot ($18.17/hr per VM) | $2.27 | −82% |
| Google Cloud (us-central1) | On-demand | $11.06 | — |
| Google Cloud | 1-year resource CUD | $7.67 | −31% |
| Google Cloud | 3-year resource CUD | $4.86 | −56% |
| Google Cloud | DWS Flex-start | $4.79 | −57% |
| Google Cloud | Spot | $6.62 | −40% |
| CoreWeave | Spot ($19.71/hr per 8 GPUs) | $2.46 | −60% |
Per-GPU-hour = VM price ÷ 8. Azure reservation per-hour = total term price ÷ (term years × 8,760 hours) ÷ 8. All Axis Intelligence Research calculations from vendor list prices.
Two figures in this table should change how buyers shop. Azure’s 8x H100 spot rate of $2.27 per GPU-hour is below every neocloud on-demand H100 price we verified — if your workload can survive eviction. And on Google Cloud, spot ($6.62) is more expensive than DWS Flex-start ($4.79) for the same a3-highgpu-8g machine — a 38% premium for spot over a queued-start option that most teams never price.
What Does an 8-GPU H100 Server Cost Per Year in the Cloud?
Axis Intelligence Research cloud TCO: one 8x H100 node, list price, 100% utilization
| Provider and model | 1 year | 3 years |
|---|---|---|
| Azure, on-demand | $861,283 | $2,583,850 |
| Google Cloud, on-demand | $775,172 | $2,325,517 |
| Azure, 1-yr / 3-yr reservation (term price) | $551,221 | $1,134,310 |
| CoreWeave, on-demand | $431,342 | $1,294,027 |
| AWS Capacity Block, rate from Oct 7, 2026 | $418,378 | $1,255,133 |
| Google Cloud, 3-year resource CUD | $340,452 | $1,021,356 |
| Lambda, on-demand ($3.99 × 8) | $279,619 | $838,858 |
| Runpod, on-demand ($3.49 × 8) | $244,579 | $733,738 |
Formula: hourly rate for 8 GPUs × 8,760 hours × years. Reservation rows use the vendor’s published term price. The AWS row assumes Capacity Blocks could be chained continuously at the October 7 rate, which AWS does not guarantee — read it as a price benchmark, not an available contract. Storage, egress, support plans and idle time are excluded.
Sarah Mitchell: The 3.5x spread between the top and bottom of this table is $1.85 million over three years for one node. At that size, the cloud bill stops being an IT line and starts being a capital decision. And notice what a 3-year Azure reservation actually buys: $1.13 million, which is still 55% more than three years of Runpod on-demand with no commitment at all. Commitments discount the hyperscaler’s own list price. They do not close the gap to the neoclouds.
How Much Do Gaming GPUs Cost Right Now?
Consumer graphics card prices are being pulled by the same force as data-center rates: memory supply diverted to AI accelerators. Axis Intelligence Research verified two current street prices directly on Newegg on October 2, 2026, against NVIDIA’s official launch prices.
GeForce RTX 50 Series: launch price vs. verified street price
| Card | Launch price (NVIDIA, Jan 6, 2025) | Cheapest new, in-stock on Newegg (Oct 2, 2026) | Premium over launch price |
|---|---|---|---|
| GeForce RTX 5070 | $549 | $869.99 (MSI Gaming Trio OC) | +58% |
| GeForce RTX 5070 Ti | $749 | $1,169.99 (Gigabyte WindForce OC) | +56% |
| GeForce RTX 5080 | $999 | not verified in this snapshot | — |
| GeForce RTX 5090 | $1,999 | not verified in this snapshot | — |
Launch prices: NVIDIA Newsroom, January 6, 2025. Street prices: Newegg search results sorted by lowest price, excluding used, refurbished and out-of-stock listings, retrieved October 2, 2026.
Refurbished RTX 5070 cards sat at $799.99 on the same results page — still 46% over launch price. Axis Intelligence Research does not publish RTX 5090 or RTX 5080 street prices in this edition because we could not verify them from a retailer listing on the snapshot date; figures circulating elsewhere range widely by channel and should be checked against a live listing before use.
What Moves GPU Prices?
Who you rent from. The single largest variable in the data above. Same chip, 3.5x price range on list.
How much availability you’re buying. Guaranteed cluster capacity (Capacity Blocks, 1-Click Clusters, reservations) is priced above interruptible capacity — sometimes above plain on-demand, as Lambda’s B200 shows.
Bundled host resources. Hyperscaler VMs include large vCPU, RAM and local-SSD allocations in the GPU price. A Google Cloud a3-highgpu-8g bundles 208 vCPUs and 1,871 GB of memory; you pay for them whether your job uses them or not.
Commitment length. Three-year commitments cut Azure and Google Cloud H100 list prices by 56%.
Memory supply. On the consumer side, graphics card prices above launch MSRP track the cost and availability of GDDR7, which competes for production capacity with the HBM used in data-center GPUs.
GPU Hour Premium Ratio (GHPR): Methodology
The GPU Hour Premium Ratio (GHPR) is an Axis Intelligence Research metric that measures how much more the major hyperscalers charge, at list, for one GPU-hour than the specialist GPU clouds do for the same GPU model.
Formula: GHPR = (mean hyperscaler on-demand price per GPU-hour) ÷ (median neocloud on-demand price per GPU-hour)
Panel (this reading): hyperscalers — Google Cloud (us-central1) and Microsoft Azure (East US); neoclouds — CoreWeave (North America), Lambda, Runpod. AWS is included once a verified on-demand rate is logged. Prices are normalized to one GPU (8-GPU instance price ÷ 8).
Readings as of October 2, 2026 (baseline):
| GPU | Hyperscaler mean | Neocloud median | GHPR |
|---|---|---|---|
| H100 | $11.68 (GCP $11.06, Azure $12.29) | $3.99 (CoreWeave $6.16, Lambda $3.99, Runpod $3.49) | 2.93 |
| H200 | $10.60 (GCP) | $5.45 (CoreWeave $6.31, Runpod $4.59) | 1.95 |
| A100 80GB | $5.07 (GCP a2-ultragpu-1g) | $2.70 (CoreWeave $2.70, Lambda $2.79, Runpod $1.59) | 1.88 |
A GHPR of 1.0 would mean hyperscalers and neoclouds list the GPU at the same price. This is the baseline reading; GHPR will be recomputed with every price refresh of this page. The H200 and A100 readings rest on a single hyperscaler each and should be read with that in mind. GHPR compares list prices, so it does not capture negotiated enterprise discounts, which are not public.
Questions Buyers Ask About GPU Prices
Is it cheaper to rent an H100 or buy one?
At list prices verified October 2, 2026, renting one H100 on Lambda for a full year costs $34,952 ($3.99 × 8,760 hours). Axis Intelligence Research did not verify a current new H100 purchase price from a primary seller for this edition, so we do not publish a buy-versus-rent breakeven; compare a written vendor quote against the hourly figure above.
Which cloud has the cheapest H100 on-demand?
Among the five providers verified on October 2, 2026, Runpod listed the lowest on-demand H100 SXM price at $3.49 per GPU-hour, followed by Lambda at $3.99. Microsoft Azure listed the highest, at $12.29 per GPU-hour in East US.
Will AWS GPU prices go up in October 2026?
Yes, for Capacity Blocks. AWS’s pricing page states that from October 7, 2026, per-accelerator Capacity Block rates rise to $5.970 for P5 (H100), $14.208 for P6-B200 and $16.146 for P6-B300 — about 15% above prior rates. AWS states On-Demand and Savings Plans prices are unchanged.
Is spot GPU pricing worth the eviction risk?
For checkpointed training and batch inference, often yes: Azure’s 8x H100 spot rate works out to $2.27 per GPU-hour, 82% below its on-demand rate. For latency-sensitive serving, no list price compensates for capacity that can disappear mid-request.
Why is the RTX 5070 so far above MSRP?
On October 2, 2026, the cheapest new in-stock RTX 5070 on Newegg was $869.99, 58% above its $549 launch price. Consumer cards use GDDR7 memory, whose supply competes with the HBM going into AI data-center GPUs.
What is the GPU Hour Premium Ratio?
The GPU Hour Premium Ratio (GHPR) is an Axis Intelligence Research metric dividing the mean hyperscaler on-demand price per GPU-hour by the median neocloud price for the same GPU. The H100 baseline reading on October 2, 2026 is 2.93.
Related Axis Intelligence Research Data
- Nvidia GPU Market Share 2026 — who sells the chips priced on this page
- GPU Shipment Statistics 2026 — unit volumes behind consumer and data-center pricing
- LLM API Pricing 2026 — what the same compute costs once it is sold per token
- AI Data Center Statistics 2026 — the capacity constraints behind GPU availability
- Nvidia Blackwell Statistics — the B200 and B300 generation
Methodology
Collection. On October 2, 2026, Axis Intelligence Research read published GPU prices from: Lambda, Runpod and CoreWeave pricing pages; Google Cloud’s accelerator-optimized VM pricing page (us-central1); Microsoft’s Azure Retail Prices API (Standard_ND96isr_H100_v5, East US); AWS’s EC2 Capacity Blocks for ML pricing page; NVIDIA’s January 6, 2025 press release (launch prices); and Newegg search results sorted by lowest price (street prices).
Normalization. All cloud prices are converted to US dollars per GPU-hour. Multi-GPU instances are divided by GPU count. Azure reservation term prices are divided by hours in the term (8,760 per year). Linux pricing only.
Scope. On-demand, spot, reservation and committed-use rates as published. Negotiated enterprise discounts, egress, storage, support and taxes are excluded. Regions are stated per row.
Street prices. The lowest-priced new, in-stock listing matching the exact GPU model on Newegg; used, refurbished and out-of-stock listings excluded. One retailer, one day — a snapshot, not a market average.
About this dataset. gpu-prices.csv contains every price and every calculated figure in this article, one row per observation, with source URL, retrieval date and a flag for Axis-calculated values. License CC BY 4.0. Cite as: Axis Intelligence Research, GPU Prices 2026, 2026.
Disclosure. Axis Intelligence Research has no commercial relationship with any provider listed. Provider order in tables follows price, not partnership.
Cite This Page
APA: Axis Intelligence Research. (2026, October 2). GPU prices 2026: What an H100 hour, an RTX 5070 and an 8-GPU year actually cost. Axis Intelligence. https://axis-intelligence.com/gpu-prices/
MLA: Axis Intelligence Research. “GPU Prices 2026: What an H100 Hour, an RTX 5070 and an 8-GPU Year Actually Cost.” Axis Intelligence, 2 Oct. 2026, axis-intelligence.com/gpu-prices/.
Chicago: Axis Intelligence Research. “GPU Prices 2026: What an H100 Hour, an RTX 5070 and an 8-GPU Year Actually Cost.” Axis Intelligence, October 2, 2026. https://axis-intelligence.com/gpu-prices/.