Who actually publishes GPU prices? We audited 77 providers to find out
Sixty-six GPU cloud providers publish their hourly rates openly, with no account and no sales call. That is the direct answer, and it is also the entire reason this panel can exist: those 66 providers are the ones whose prices anyone - a buyer, a researcher, an LLM answering a question - can actually check. This post is the census of everyone else: who hides prices behind a signup, who gates them behind a sales conversation, and which providers publish a teaser while the real catalog waits behind a login. Every classification below comes from reading the provider's own pages this week, and the full list with per-provider evidence is in our public repo.
| Category | Count | What it means for a buyer |
|---|---|---|
| Open rate cards | 66 | Full hourly rates readable with no account |
| Signup-gated | 6 | Pricing inside a provisioning console - free account required to see anything |
| Quote-only | 5 | No self-serve rates; sales conversation required |
| Thin slice on the panel | 14 | Public page shows 1-16 listings; more behind access gates |
Who publishes real rates?
The 66 providers on the panel tonight, covering 1,409 listed configs and 4,453 GPUs. The spread of what "publishing" means is wide. RunPod, Vast.ai and SF Compute expose live APIs where prices move intra-day. Most providers publish a rate card page - Lambda Labs, Crusoe, Nebius, CoreWeave - that changes every few weeks. And a growing tail of regional providers publishes honest hourly floors in local currency: iRender and EZYCLOUDX in Vietnam, NovaGPU and TeamCloud in Malaysia, ACC in Singapore, Biznet Gio in Indonesia, E2E and JarvisLabs in India, Sakura Internet in Japan. If you want the cheapest RTX 5090 hour sold by a datacenter provider tonight, it is EZYCLOUDX in Hanoi at $0.69. You would never find that number on any Western comparison site, because none of them read these pages.
Who makes you sign up first?
Six providers show nothing until you create an account: TensorDock, Civo's fuller catalog, JarvisLabs' console rates, VNG Cloud and GreenNode in Vietnam, FPT Cloud's AI Factory, and vnso. The pattern is not malice - it is that their pricing lives inside a provisioning console, and the console requires an identity. But the effect on buyers is real: a market where comparing six providers costs six account registrations is a market where most people compare two and overpay.
Who wants a sales call?
Five providers we logged this week publish no self-serve rates at all: Sustainable Metal Cloud in Singapore, gpuyard, readyspace, saidi, and the bare-metal arm of IP ServerOne. Quote-only pricing is a legitimate choice for cluster-scale deals - the price of 512 H200s genuinely depends on term, network fabric and power commitments. But it means the advertised market and the traded market are different things, and anyone quoting "GPU prices" from quote-only sources is quoting an opening bid.
Who shows you a slice?
The subtlest category, and the one we did not expect to be this large. Fourteen providers already on the panel publish a thin public slice - one to sixteen listings - while their own pages admit more inventory sits behind access gates. Akamai publishes three hourly GPU plans and labels the RTX PRO 6000 Blackwell "access-gated" on the same page. Scaleway publishes the L4 family openly and zone-gates its L40S and H100 rows. SF Compute's homepage order book shows one average price; the live book is behind the dashboard. Voltage Park publishes a single on-demand H100 rate while listing H200 and B200-class hardware. We keep collecting what is public, flag the rest, and the registry of all fourteen - with the exact gate for each - is in the repo.
Frequently asked questions
How many GPU cloud providers publish their prices?
As of September 18, 2026: 66 providers publish hourly GPU rates readable without an account, covering 1,409 listed configurations. Six providers require signup before showing any rates, five are quote-only, and fourteen publish a partial catalog with more behind a login.
Why do some GPU providers hide their pricing?
Three reasons, in order of frequency: pricing lives in a provisioning console that requires an account (most common), capacity is allocated through sales conversations at cluster scale (legitimate for large blocks), and access tiers gate premium SKUs to qualified accounts (Akamai's Blackwell, Scaleway's zones).
Are published GPU prices the prices you actually pay?
Published on-demand hourly rates are the ceiling, not the clearing price. Committed-use and reserved terms run 30-70% below the sticker on the providers that publish both. Floors marked "from" are marketing minimums. The panel collects the on-demand published rate with its receipt, labels floors, and excludes commitment pricing - so the numbers you see are the rates anyone can book at, tonight.
What happens next
The census is a living document. Phase two is already scheduled: account-backed collection for the six signup-gated providers and the twelve crackable thin-slice ones, each logged with its gate in the registry. And every future sweep - the next countries are already queued - adds to the same public list. The market gets more legible every time we count it.