GPU-hour list prices reflate — and the discount moves to interruptible capacity
Published GPU-hour rates for Hopper and Blackwell hardware are going up. Between 2026-07-14 and 2026-09-24, eight GPU clouds in the corpus raised an on-demand or reserved H100/H200/B200/B300 rate, and a ninth, Anyscale, pulled its H100 and H200 prices for a Contact Us row. The cheap hour did not disappear: it moved to spot, preemptible and community capacity the vendor can take back.
What's happening — and why
What's happening: GPU clouds spent the first half of 2026 cutting. Together AI cut on-demand cluster H100 twice in June ($5.49 to $4.79 to $3.99), and RunPod cut H100 SXM from $3.29 to $2.99 on 2026-07-14. From mid-July the direction flipped, on exactly the hardware AI training and inference runs on.
Nine dated raises across eight vendors in ten weeks. DeepInfra's dedicated H100 went from $1.79 to $2.20. Hyperbolic's H100 SXM went from $1.50 to $2.89 and then to $3.19, two raises 14 days apart. Fal's H100 list went from $3.99 to $4.50. Fireworks moved on-demand B200 from $10 to $13 effective 2026-09-01. RunPod's Secure Cloud H100 went from $2.99 to $3.49, Lightning AI's H100 from $3.29 to $4.50, and Nebius's HGX H100 from $3.85 to $4.50 effective 2026-10-01. Anyscale stopped printing a Hopper price at all.
The list is converging upward. Four corpus vendors now print an H100 at exactly $4.50 an hour: Fal, Lightning AI, Nebius and Hugging Face's AWS H100.
What did not move: consumer and older cards held or fell. Novita cut its RTX 5090 23%, DeepInfra kept A100 flat at $0.89, and Nebius left RTX PRO 6000 and L40S unchanged. And every vendor that raised on-demand kept a cheap price somewhere else. Together added a preemptible H100 at $1.99 (50% off on-demand), Nebius moved preemptible to dynamic spot from $0.79, Fal's negotiated $1.89 floor held under the rising list, and RunPod left every Community Cloud rate unchanged.
Why: the corpus records no vendor-stated reason for the raises, so no driver is claimed here. What is observable is that the guaranteed, non-interruptible hour got more expensive while the interruptible hour did not.
How it works
Evidence over time
11 supporting · 5 counter — hover or tap a point for detail, click to jump to the row.
Evidence
| Company | Date | What happened |
|---|---|---|
| DeepInfra | Jul 2026 | Raised dedicated GPU-hour rates 16-32%: H100 $1.79 to $2.20, H200 $2.19 to $2.69, B200 $2.79 to $3.69, B300 $4.20 to $4.89 per GPU-hour, with A100 flat at $0.89. Logged at the time as a reversal for a vendor known for cutting. |
| Hyperbolic | Jul 2026 | Nearly doubled H100 SXM from $1.50 to $2.89/GPU/hr, raised H200 $2.40 to $3.49 and B200 $3.50 to $5.99, and pulled its consumer-GPU hourly rates from the page. |
| Together AI | Jul 2026 | Raised reserved GPU-cluster H100 rates on every tenor: 7-30 days $3.59 to $3.69, 31-90 days $3.29 to $3.45, 91-180 days $3.09 to $3.19 — its first reserved increase after two cuts in June. |
| Hyperbolic | Aug 2026 | Second raise in 14 days: H100 SXM $2.89 to $3.19 (+10%), H200 $3.49 to $3.99 (+14%). The marketplace "starting at" claim moved from $0.20/GPU/hr to $3.19, the H100 rate itself. |
| Fal | Aug 2026 | H100 80GB List Price raised $3.99 to $4.50/h (+12.8%), now identical to H200's list. The negotiated "as low as" floor held at $1.89/h, so the gap between list and floor widened to 2.4x. |
| Fireworks AI | Aug 2026 | Published a forward-dated increase on every on-demand GPU tier effective 2026-09-01: H100/H200 $7 to $8/hr (+14%), B200 $10 to $13 (+30%), B300 $12 to $15 (+25%), GB300 $18 to $20 (+11%) — the first repricing of an already-published on-demand SKU since the product launched in January 2024. |
| Anyscale | Sep 2026 | Pulled its published H100 (AC 9.2880/hr) and H200 (AC 10.6812/hr) rates for a single unpriced "NVIDIA H/B/GB GPU families — Contact Us" row, while CPU through A100 stayed public. The withdrawal hit exactly the GPU classes the rest of the cohort was raising. |
| RunPod | Sep 2026 | Raised 11 of 19 Secure Cloud Pod rates: H100 SXM $2.99 to $3.49, B200 $5.89 to $6.79, H200 $4.39 to $4.59, B300 $7.39 to $7.89, A100 PCIe $1.39 to $1.59. Every Community Cloud rate was unchanged. The H100 raise reverses RunPod's own $3.29 to $2.99 cut of 2026-07-14. |
| Together AI | Sep 2026 | Added a preemptible pay-as-you-go column to GPU Clusters at $1.99/hr H100, $2.99 H200, $4.09 B200 — 50% off on-demand. The cheap H100 price returned, but only on capacity that can be taken back. |
| Lightning AI | Sep 2026 | Repriced its GPU card upward at the entry and middle: T4 $0.19 to $0.55, L4 $0.48 to $0.79, A100 40GB $1.29 to $2.19, H100 $3.29 to $4.50. In the same change H200 FELL from $6.53 to $4.50, so H100 and H200 now cost the same. |
| Nebius | Sep 2026 | Announced on-demand increases effective 2026-10-01: HGX H100 $3.85 to $4.50, H200 $4.50 to $5.40, B200 $7.15 to $8.50, B300 $7.85 to $9.50 (RTX PRO 6000 and L40S unchanged). Preemptible moves from fixed rates to dynamic spot from $0.79 (H100/H200) and $0.99 (B200/B300). |
Counterexamples
- Together AI · Sep 2026 — Ran a promotion cutting Dedicated Inference on-demand H100 from $5.49 to $3.99/hr (27% off) through 2026-09-30 — a Hopper price moving DOWN inside the window, albeit dated. Together also cut on-demand cluster H100 twice in June ($5.49 to $4.79 to $3.99).
- Novita AI · Sep 2026 — Cut self-serve RTX 5090 23% to $0.56/hr. Its H100 SXM self-serve rate held at $3.39/hr ($1.70 spot) through two removals and re-additions (2026-08-26, 2026-09-07, 2026-09-23) — a data-center GPU that did not reprice.
- Lightning AI · Sep 2026 — The same change that raised H100 to $4.50 cut H200 from $6.53 to $4.50, a 31% decrease on a Hopper-class GPU.
- RunPod · Jul 2026 — Cut H100 SXM from $3.29 to $2.99 and RTX Pro 6000 from $2.09 to $1.99 — a cut that lasted 68 days before the 2026-09-20 raise.
- Hugging Face · Jul 2026 — Added an AWS H100 on Inference Endpoints at $4.50/hr against its own GCP H100 at $10/hr — a new listing priced 55% under the incumbent one.
Trivia
-
Four corpus vendors now list an H100 at exactly $4.50 an hour — Fal (list price, from 2026-08-11), Lightning AI (2026-09-21), Nebius (effective 2026-10-01) and Hugging Face's AWS H100 (2026-07-28) — and three of the four got there by RAISING, from $3.99, $3.29 and $3.85.
-
Hyperbolic's H100 SXM went from $1.50 to $3.19 an hour in 14 days (2026-07-21 and 2026-08-04), and its marketplace "starting at" floor went from $0.20 to $3.19 — the floor price became the H100 price.
-
RunPod's 2026-09-20 raise touched only Secure Cloud. Before it, Community Cloud was priced ABOVE Secure on two GPUs (B200 and L4); after it, Community is cheaper on every GPU on the card. The tier that did not move became the discount tier.
For buyers
Rebuild any GPU budget that was set from 2026-H1 on-demand list rates, because it is now low. The same on-demand H100 hour costs 17% more at Nebius and 113% more at Hyperbolic than each vendor's own mid-2026 rate. Then sort your workloads by whether they can be interrupted. Anything that checkpoints cleanly (batch inference, most training runs, evaluation sweeps) can still buy the old price on spot, preemptible or community capacity: Together's preemptible H100 at $1.99, Nebius's dynamic spot from $0.79, RunPod's unchanged Community Cloud. Anything that cannot be interrupted should be budgeted from the new list, or moved onto a commitment, which is the other route back to a lower rate. Watch forward-dated notices too: Fireworks published its increase three weeks ahead and Nebius a week ahead, which is the window to lock a reservation.
For vendors
If you are raising, the cohort shows the pattern that holds the customer: raise the guaranteed hour and keep a visible cheap tier for workloads that can tolerate preemption. Together's 50%-off preemptible column and RunPod's untouched Community Cloud both kept a low headline price on the card while the on-demand rate rose. Forward-date the change as Fireworks and Nebius did, because a surprise raise on an already-published SKU invites churn. And be deliberate about withdrawing prices: Anyscale's Contact Us row on exactly the GPU classes the rest of the cohort was raising reads as a price increase to buyers even though no number was printed.
Outlook — what to watch
The arc is ten weeks old and directional across eight vendors, which is why it is logged as `new` rather than `emerging`. Two things would flip it. If, within 90 days, two or more of these eight vendors cut Hopper or Blackwell on-demand list below its pre-raise level, the reflation was a blip. If spot and preemptible rates start rising in step with on-demand, the split between the guaranteed and interruptible hour collapses and this becomes a plain price increase rather than a two-track market. Keep the counter-moves in view: Together's dated promotion cut Dedicated Inference H100 from $5.49 to $3.99 through 2026-09-30, Lightning AI cut H200 31% in the same change that raised H100, and Hugging Face listed an AWS H100 at $4.50 against its own $10 GCP H100.
Bottom line
The published price of a guaranteed Hopper or Blackwell hour went up at eight GPU clouds in ten weeks, and four now list an H100 at exactly $4.50. The cheap hour survives only on capacity the vendor can take back, so the practical question for a buyer is which of your workloads can accept preemption.
FAQ
Are GPU cloud prices going up in 2026?
For data-center Hopper and Blackwell GPUs, yes. Between 2026-07-14 and 2026-09-24 eight corpus GPU clouds raised a published on-demand or reserved rate: DeepInfra, Hyperbolic (twice), Together AI, Fal, Fireworks AI, RunPod, Lightning AI and Nebius. Anyscale withdrew its H100 and H200 prices to Contact Us. Consumer and older cards such as the RTX 5090 and A100 held or fell.
How much did H100 hourly prices rise?
It varies by vendor. DeepInfra's dedicated H100 went from $1.79 to $2.20, RunPod's Secure Cloud H100 from $2.99 to $3.49, Lightning AI's from $3.29 to $4.50, Nebius's from $3.85 to $4.50 and Hyperbolic's H100 SXM from $1.50 to $3.19 in two steps. Four vendors (Fal, Lightning AI, Nebius and Hugging Face's AWS H100) now list an H100 at exactly $4.50 an hour.
Where can I still get a cheap H100 hour?
On capacity that can be interrupted. Together AI added a preemptible H100 at $1.99/hr, 50% off on-demand. Nebius is moving preemptible to dynamic spot from $0.79. Fal's negotiated floor held at $1.89 under a $4.50 list, and RunPod's Community Cloud rates did not change, so Community is now cheaper than Secure on every GPU on its card.
Why are GPU-hour prices rising?
The corpus records no vendor-stated reason, so none is claimed here. What is observable is the shape: the raises land on guaranteed data-center GPU hours, while interruptible capacity and consumer-class cards held or fell.