In August an H100 was reported available 34% of the time. In September it was 25%. This is the first month GPU Finder can compare against a complete earlier one, and the direction was the same almost everywhere: of the 76 provider × GPU × node-size records ranked in both months, 29 got scarcer, 8 improved and 39 stayed within one percentage point.
GPU Finder polls provider stock APIs roughly every hour and keeps the history. September 2026 is frozen in the second monthly edition of the GPU Cloud Availability Index: 120 provider × GPU × node-size records, of which 92 cleared the coverage floor and carry a score. Those 92 cover 199,881 provider-hours of known signal from 13,646 stored observations, 1 September to 1 October UTC. Numbers here are rounded; the index carries the exact values and is downloadable as CSV and JSON.
Terms. A record is one provider × GPU × node size, summed across every region we track. Its covered hours are the hours with a known signal (available, limited or unavailable); unknown hours are dropped. A record is ranked with at least 48 covered hours and 50% of the hours it could have had. Available % is hours reported available divided by covered hours. limited (a handful of units) does not count, so every figure below is a floor for "could I have launched one", not a ceiling.
Key facts (GPU Finder, GPU Cloud Availability Index, September 2026):
- In September 2026, H100 capacity was reported available 25% of covered hours across the 12 cloud providers ranked in GPU Finder's GPU Cloud Availability Index, down from 34% in August 2026 (hourly stock-API checks, 1 to 30 September 2026 UTC).
- In September 2026, H200 was reported available 40% of hours (52% in August) and A100 45% (52% in August).
- B200 was reported available 12% of hours in September 2026, up from 5% in August, and the rise comes from two newly ranked sources (Packet and Lium); among providers ranked in both months B200 was unchanged at about 5%.
- A single H100 at Lambda was reported available 22% of September 2026, against 51% of August; an 8x H100 node at Lambda was available under 1% of the month.
- 36 of the 92 ranked provider × GPU × node-size records showed no available stock at any point in September 2026.
H100, H200, A100 and B200 availability: September vs August 2026
H100 availability fell from 34% to 25% of hours
Across the 12 providers with enough H100 coverage to rank, H100 capacity was reported available for 25.5% of covered hours, coverage-weighted. Across all four GPUs and all ranked records it was 30.1%, from 37.5% in August.
The same caveat as last month applies. PrimeIntellect is a marketplace that resells other clouds, and its feed again scored 100% on every one of its ten records. Without those rows the numbers are H100 16%, H200 17%, A100 35% and B200 9%.
| GPU | Ranked providers | September | August | September excl. PrimeIntellect | August excl. PrimeIntellect |
|---|---|---|---|---|---|
| H100 | 12 | 25% | 34% | 16% | 25% |
| H200 | 7 | 40% | 52% | 17% | 28% |
| A100 | 5 | 45% | 52% | 35% | 44% |
| B200 | 9 | 12% | 5% | 9% | 1% |
The provider set changed between the two editions, so the fairer read is the cohort ranked in both months (76 records). On that cohort H100 went from 37% to 27%, H200 from 52% to 47%, A100 from 52% to 45%, and B200 from 4.9% to 5.5%.
B200 availability: the rise is new sources, not more stock
B200 moved from 5% to 12%, and none of that came from providers that were already ranked. Packet, ranked for the first time, reported its B200 available 98% of the month at 1x, 68% at 2x, 48% at 4x and 0% at 8x, with the larger sizes selling out in the second half of September. Lium, also newly ranked, was at 18% for a single B200. Hyperstack, Lambda, Runpod, Shadeform and Vultr reported 0% for every B200 size they list, and Nebius about 5%.
8x H100 availability vs 1x
Availability still fell with node size at the providers that sell every size, from a lower starting point. Lambda's H100 went from 22% at 1x to under 1% at 8x, Hyperstack from 30% to 0%, Runpod from 11% to 0%, Shadeform from 36% to 3%. Across all ranked records a single H100 was available 35% of the time (49% in August), a 2x node 23%, a 4x node 13% and an 8x node 23% (28% in August); the 8x figure is again lifted by providers at 100%.
Live per-provider stock at full-node size is on the 8x H100, 8x H200 and 8x B200 server price pages.
36 of 92 records were never available
Over a third of ranked provider × GPU × size combinations finished the month at 0%, up from 27 of 81. Twelve finished at 100%: PrimeIntellect's ten and Digital Ocean's two H100 sizes.
H100 availability by provider, September 2026
Coverage-weighted across every node size a provider sells.
| Provider | September | August | Covered hours | Node sizes |
|---|---|---|---|---|
| PrimeIntellect | 100% | 100% | 11,520 | 1x, 2x, 4x, 8x |
| Digital Ocean | 100% | 100% | 1,440 | 1x, 8x |
| Scaleway | 63% | 90% | 1,440 | 1x, 2x |
| Nebius | 62% | 81% | 1,440 | 1x, 8x |
| Shadeform | 22% | 39% | 27,600 | 1x, 2x, 4x, 8x |
| Lambda | 12% | 30% | 13,680 | 1x, 2x, 4x, 8x |
| Lium | 10% | not ranked | 10,206 | 1x, 2x, 4x, 8x |
| Gcore | 10% | 7% | 7,200 | 8x |
| Runpod | 6% | 10% | 12,960 | 1x, 2x, 4x, 8x |
| Hyperstack | 6% | 8% | 8,640 | 1x, 2x, 4x, 8x |
| AceCloud | 0% | 0% | 2,880 | 1x, 2x, 4x, 8x |
| Vultr | 0% | 0% | 720 | 8x |
Gcore is the only H100 provider that improved, and its 8x A100 went from 0% in August to 42%. AceCloud reported no H100 in either month but its H200 rose to 90% from 75%.
As in August, a high score on a narrow catalogue is not the same claim as a middling score on a wide one. Records sum configuration-hours across every tracked region, so a provider tracked in many regions with stock in one of them scores a fraction, while a single-region provider scores on one place to launch.
Caveats: what changed in the provider set, and one that cuts against Shadeform
Cudo is gone, Lium and Packet are in. Cudo stopped renting GPUs on demand and was delisted from GPU Finder on 19 September, so it is excluded from this edition; it scored 0% for H100 in August. Lium (added 20 August) and Packet (feed live 25 August) had their first full month. Seven of Lium's records have partial coverage, 69% to 78% of possible hours, because it is a marketplace whose per-country listings appear and disappear; they are ranked, and flagged partial in the data.
Configurations a provider stopped listing still count as unavailable. The index scores every configuration that has a stock signal, including ones that have since dropped out of the price catalogue. In September that matters most for Shadeform, whose API kept reporting several older listings as unavailable all month: its H200 at 1x, 2x and 4x and its B200 rows score 0% on listings last priced in July or August, and three of its four 8x H200 listings are in the same state. The one 8x H200 listing still in the catalogue was available about 67% of the month, against the pooled 17% in the index. Restricted to configurations still in the price catalogue during September, the month reads H100 29%, H200 49%, A100 51% and B200 16%. August was scored the same way, so the month-over-month direction holds on either basis, but Shadeform's H200 fall from 25% to 6% is mostly listings ending, not stock running out.
PrimeIntellect's 100% is the same signal as before. Its endpoint reports a stock level per upstream listing and GPU Finder records the best level across all of them for each GPU and node size. 100% means that at some upstream cloud, in some data centre, that configuration was at high stock in every covered hour.
Vast.ai and Theta EdgeCloud are not ranked. Vast's availability history is kept for seven days, so about a third of September survived to the freeze, under the 50% floor. Theta's cloud catalogue carries no capacity signal and is recorded as unknown. Both have live listings on GPU Finder; Vast.ai is often the cheapest H100 on the site, and its absence here is a retention limit, not a verdict.
stale on the index page means "no new record in the six hours before the freeze". GPU Finder stores a history record only when a provider's signal changes, plus one heartbeat per UTC day. The edition was frozen at 12:44 UTC on 1 October; 76 of 120 rows read stale because their signal had not changed since that day's heartbeat. It does not feed the monthly percentage.
How to find an H100 available right now
- Shop by node size first. If you need 8x H100, start from the 8x H100 server price page, which shows live stock per provider at that size, rather than the headline 1x price.
- Read reliability next to price. Every listing on the H100 cloud pricing page carries its 30-day reliability score (how it is computed: GPU cloud reliability); the cheapest row is only cheap if it was launchable. The same column sits next to every price on the cheapest cloud GPU table.
- Set an alert instead of refreshing. With a single H100 available about a third of the time, a price and availability alert from any GPU page is the cheaper way to catch stock.
- Citing these numbers. The GPU Cloud Availability Index, September 2026 edition is frozen, dated and downloadable. Suggested citation: Alexander Salikov, "GPU Cloud Availability Index — September 2026," GPU Finder, generated 1 October 2026.
Methodology: how GPU cloud availability is measured
Each provider's stock API is polled roughly hourly; a history record is stored whenever the signal changes, and at least once per UTC day. Each stored signal holds until the next record, capped at 30 hours, and a longer gap counts as missing, not downtime. The score is the share of covered time reported available, so a provider available for six days and out for one reads 86%, not 50%. The method is unchanged from the August 2026 edition, which is the baseline for every comparison above. Full definitions and the data dictionary are on the index page.