An Nvidia H100 rents for $2.69 to $6.16 per GPU-hour as of October 16, 2026, across six providers' published price lists, and the cheapest B200 lists at $6.69. Those prices went up this year, not down, which is the clearest sign the AI chip shortage is not over.
This page is a price sheet first. Value Add VC pulled the on-demand rate card from each provider's own pricing page on October 16, 2026, converted every node price to a per-GPU hourly figure, and set the results against the 2026 reporting on rising rents and scarce memory. Rates change often, so every row links the page it came from.
GPU pricing in 2026: is the AI chip shortage over?
No. GPU pricing in October 2026 says the AI chip shortage is still on: listed H100 rates run $2.69 to $6.16 per GPU-hour, on-demand B200 rates $6.69 to $8.60, and both list prices and market rents rose during 2026, even for H100s that first shipped in volume four years ago. Cheaper compute would show up as falling rents, and rents rose.
The best single data point is Nebius. Its pricing page showed the H100 at $2.95 an hour in early May; after an October 1 increase the same chip costs $4.50, and its B300 went from $6.10 to $9.50, according to The Motley Fool, which put the two increases at about 53% and 56% in five months. The same report says Nebius's second-quarter shareholder letter cited more than 30% higher pricing for older-generation GPUs versus the first quarter. The $4.50 H100 rate is live on Nebius's price list today.
How much does an H100 cost per hour?
The median of the six providers' cheapest on-demand H100 SXM rates is $4.25 per GPU-hour (Value Add VC's arithmetic on the table below, taking one rate per provider). Specialist GPU clouds sit at the low end and the two node-priced sellers, CoreWeave and AWS, at the top.
| Provider | GPU | $/GPU-hr | Configuration |
|---|---|---|---|
| RunPod (Community Cloud) | H100 SXM 80GB | $2.69 | Single GPU, per-second billing |
| Lambda | H100 SXM 80GB | $3.99 | 8x instance ($4.29 on a 1x instance) |
| Together AI | HGX H100 | $3.99 | GPU Clusters, pay as you go |
| RunPod (Secure Cloud) | H100 SXM 80GB | $3.99 | Single GPU, Tier 3+ data centers |
| Nebius | HGX H100 | $4.50 | Per GPU-hour, after Oct 1 increase |
| AWS EC2 Capacity Blocks | H100 (p5.48xlarge) | $5.97 | $47.757 per 8-GPU instance, reserved block |
| CoreWeave | HGX H100 | $6.16 | $49.24 per 8-GPU node |
How we counted: on-demand (pay-as-you-go) list prices for H100 SXM-class GPUs; published pricing pages read October 16, 2026; 7 rate cards from 6 providers; node prices divided by GPUs per node (Value Add VC's arithmetic); taxes, storage and egress excluded; as of October 16, 2026.
Source: each provider's pricing page, linked in the Provider column. AWS figures are EC2 Capacity Blocks reservation rates in US regions.
Two cheaper H100 tiers exist outside the table. RunPod lists an H100 PCIe card at $1.99 on Community Cloud and $2.89 on Secure Cloud; PCIe parts have lower memory bandwidth than SXM, so they are not like-for-like. Lambda also sells H100 1-Click Clusters of 16 to 256 GPUs at $5.54 to $6.16 per GPU-hour on two-week to one-year terms, more than its $3.99 self-serve instance because the clusters bundle high-speed interconnect.
H100 SXM On-Demand Price per GPU-Hour, October 2026
A 2.3x spread separates RunPod Community Cloud from CoreWeave for the same chip
Provider pricing pages (RunPod, Lambda, Together AI, Nebius, AWS, CoreWeave), read Oct 16, 2026
Data
| Item | Value |
|---|---|
| CoreWeave | $6.16 |
| AWS Capacity Blocks | $5.97 |
| Nebius | $4.50 |
| Together AI | $3.99 |
| Lambda (8x) | $3.99 |
| RunPod Community | $2.69 |
How much does a B200 cost per hour?
A B200 costs $6.69 to $8.60 per GPU-hour on demand, or $14.21 on reserved AWS Capacity Blocks, a median of $8.35 across six providers (Value Add VC's arithmetic). That is roughly twice the H100 median. Rack-scale GB200 NVL72 capacity, where 72 GPUs share one NVLink domain, prices near $10.50 per GPU.
| Provider | GPU | $/GPU-hr | Configuration |
|---|---|---|---|
| Lambda | B200 SXM6 180GB | $6.69 | 8x instance ($6.99 on 1x) |
| RunPod | B200 180GB | $7.99 | Single GPU pod |
| Together AI | HGX B200 | $8.19 | GPU Clusters, pay as you go |
| Nebius | HGX B200 | $8.50 | Per GPU-hour |
| CoreWeave | HGX B200 | $8.60 | $68.80 per 8-GPU node |
| AWS EC2 Capacity Blocks | B200 (p6-b200.48xlarge) | $14.21 | $113.666 per 8-GPU instance |
| CoreWeave | GB200 NVL72 | $10.50 | $42.00 per 4-GPU instance |
| AWS EC2 Capacity Blocks | GB200 (u-p6e UltraServer) | $10.58 | $761.904 per 72-GPU UltraServer |
How we counted: on-demand list prices for Nvidia Blackwell GPUs (6 B200 rate cards, 2 GB200 rate cards); pricing pages read October 16, 2026; node and UltraServer prices divided by GPU count (Value Add VC's arithmetic); as of October 16, 2026.
Source: provider pricing pages linked in the Provider column.
The newer B300 lists at $8.99 on RunPod, $9.50 on Nebius and $9.99 on Together AI's GPU Clusters, while an AWS p6-b300.48xlarge capacity block works out to $16.146 per GPU. The AWS premium holds across every generation: its B200 block at $14.208 per GPU is about 70% above the six-provider median. For how much work each generation does per dollar, see the H200 vs B200 vs GB200 comparison.
Median On-Demand Price per GPU-Hour by Nvidia Generation, October 2026
Each Nvidia generation roughly steps the hourly rate up by $1 to $4
Value Add VC arithmetic on provider pricing pages read Oct 16, 2026 (medians of each provider's listed rate)
Data
| Item | Value |
|---|---|
| H100 | $4.25 |
| H200 | $5.70 |
| B200 | $8.35 |
| B300 | $9.75 |
H200 median from four rate cards: RunPod $5.29, Nebius $5.40, Together AI $5.99, CoreWeave $6.31 (its $50.44 eight-GPU node divided by eight). B300 median from four: RunPod $8.99, Nebius $9.50, Together AI $9.99, AWS $16.15.
Why are GPU rental prices rising in 2026?
Rents climbed across chip generations in 2026. Silicon Data figures cited by Wccftech on August 3, 2026 put the H100's market rental rate just shy of $2 an hour in January and approaching $3 by August, while B200 rates moved from just below $5 to $5.50 to $5.80. Those index figures sit below the list prices in the tables above because they track transacted and marketplace rates rather than rate cards.
The newest chips tightened fastest. Market strategist Warren Pies posted that B200 rental rates were up 30% year to date and called demand "insatiable," with availability at 0% in his own index, TIKR reported on September 29, 2026.
Memory is the other squeeze. High-bandwidth memory (HBM) is stacked on every data-center GPU, and its scarcity is now moving chip prices. KED Global reported on August 24, 2026 that Nvidia is raising prices for its AI accelerators amid a surge in memory chip prices. Chinese chipmakers including Huawei and Cambricon are also raising accelerator prices because HBM has become scarce, with Huawei's Ascend 950DT reportedly quoted 20% to 50% above earlier pricing, per Reuters reporting summarized by Yahoo Finance.
Nebius List Price per GPU-Hour: May vs October 2026
The Motley Fool, Sep 27, 2026 (May prices); nebius.com/prices, read Oct 16, 2026
Data
The older H100 rose about as fast as the newest B300, which points to a market-wide shortage rather than a launch premium on one chip.
What does the cheapest GPU capacity cost?
Interruptible capacity is far cheaper than on-demand. Together AI lists preemptible HGX H100 at $1.99 per GPU-hour against $3.99 on demand, and its reserved H100 rate falls from $3.69 for 3 to 30 days to $3.19 for 181 days or more, per its pricing page. CoreWeave lists an eight-GPU H100 spot node at $19.71 an hour, about $2.46 per GPU, and Nebius advertises spot H100 capacity from $0.79.
Commitments cut the B200 bill too: Together AI's B200 rate drops from $8.19 on demand to $6.79 on terms of 181 days or more, and Lambda's self-serve B200 at $6.69 already undercuts most providers' cluster pricing. For what those hourly rates mean per user or per token, see the true cost of running an AI product in 2026 and how inference costs per token have fallen even as GPU-hours got pricier.
What higher GPU prices mean for AI clouds and buyers
For GPU clouds, rising rates lift the value of unsold capacity. Nebius had about $37.5 billion of remaining performance obligations at the end of June, more than 60 times its second-quarter revenue of $582 million, and those contracts were priced when signed, so the October increase does not reprice them, The Motley Fool noted in the same report. The same article says Nebius's annual contract value per megawatt rose from a 2026 base of around $12 million to over $20 million on deals closed in the second quarter.
That pricing power is what investors are paying for in the neocloud group: see CoreWeave's backlog and Lambda's path to an IPO. For buyers, the read is simpler. A startup budgeting for 2027 should not assume GPU-hours get cheaper on their own; the capex behind the supply is tracked on the AI buildout tracker and the AI spending tracker.
What the headline misses
List prices are not what large buyers pay. Most GPU capacity moves on multi-month or multi-year contracts negotiated off the rate card, and marketplace indices such as Silicon Data's sit well below the published on-demand rates in the tables here. A rate card shows the direction of the market more reliably than its level.
Shortage pricing can also reverse. The Motley Fool flagged that if supply catches up while Nebius and its rivals keep building, hourly rates could drop as fast as they rose, and Nebius alone guided to $20 billion to $25 billion of 2026 capital spending. TIKR raised the same question for the B200: whether its squeeze is a lasting trend or a short window for the newest chip that closes once supply arrives.
One read on this: per-token AI prices can keep falling while per-hour GPU prices rise, because each new chip and model does more work per hour. A higher hourly rate is a cost problem only for buyers whose throughput per GPU stands still.
Published by Value Add VC · Editor-in-chief: Trace Cohen. Prices are read from each provider's public pricing page and change often; check the linked page before you buy. Get the weekly newsletter at tracecohen.substack.com.
Latest from the Pulse
Get VC data most people never see
— free to subscribe
Trace's notes on venture, AI and startups, a few times a week. Join 5,000+ subscribers. No spam.
No spam, unsubscribe anytime. Sent via Substack · Privacy · or subscribe on Substack ↗