Computable GPU Index Review (2026): Pricing, Features & Honest Verdict

Computable GPU Index (CGI) is a free, open-source reference rate for GPU compute, published as USD per GPU-hour and recomputable from the public record. We found the engineering excellent and the licence a problem: settling anything on it needs a separate written deal. Best for: teams buying GPU capacity. Price: Free (free plan: yes). Rating: 7.0/10.

What is Computable GPU Index?

CGI is a price index for renting GPUs. Every 15 minutes it collects the published on-demand rental rates of a fixed panel of providers, takes an interquantile mean over weighted votes, and prints one number in USD per GPU-hour for each accelerator it covers. H100, H200, B200 and B300 are live. On 1 September 2026 at 11:30 UTC it read $7.70 for the B300, $6.64 for the B200, $4.19 for the H200 SXM and $3.43 for the H100 SXM.

The point is not the number, it is the receipts. Every data point is sourced and timestamped, the calculation is in a public Python repository, and there is a ./reproduce h100 2026-09-01 command that re-derives a full day from raw public history and exits 0 only if every value and digest matches what was published. The index is served four ways, all free: a live page, an anonymous read-only REST API with no key, a flat-file corpus, and an MCP server at mcp.getcomputable.com/mcp that any agent can query. It launched on Product Hunt on 1 September 2026 and took second place with 185 upvotes.

The publisher is Computable, a Y Combinator company whose founders came from Jump Trading and Coinbase. That background shows. This is a market-data product built by people who have shipped market data before, and it arrives with a methodology document, a governance document, a change log and a trademark policy, which is more institutional scaffolding than most launches have after a year.

What are Computable GPU Index’s key features?

A number you can recompute yourself

Reproducibility here is a real feature rather than a claim. Clone the repo, install one runtime dependency, and run the reproduce script against any accelerator and date. It fetches the trailing history it needs, re-derives the attendance factors, liveness weights, votes and the observation values, and checks each artifact digest. Exit 0 means everything matched, 1 means a mismatch and 2 means nothing could be verified. There is a faster receipts-only mode for spot checks.

The governance rules around it are just as considered. The methodology is versioned, every published record embeds the parameter set that produced it, and the calculation refuses to extend a series under altered parameters. Published values are never revised; corrections publish forward. That is how a reference rate is supposed to behave, and it is the opposite of a dashboard that quietly restates last week.

An outlier-resistant method, tested live by a flapping provider

Each panel provider casts weighted votes at its price and the index averages the central band, so votes in the outer thirds have no direct effect on the value. You can watch this work on the live page. On 1 September the B300 panel had eight sources passing: hyperstack at $7.40, scaleway at $8.73, verda at $7.50, latitude at $16.00, nebius at $7.85, runpod at $7.89, massedcompute at $6.60 and vast at $5.63.

Vast is the interesting one. Its receipt log shows the price flipping between $5.63 and $12.50 hour to hour, and its liveness weight has been marked down to 8.6% while the other seven sit at roughly 13.2%. A source that cannot hold a price for an hour gets a smaller vote. That is the design working in public, on the day of launch, with the evidence on the page.

Distribution built for agents, not just for humans

The API needs no account and no key, the flat-file corpus publishes latest.json and dated archives, and the MCP server exposes three read-only tools so Claude or any other client can pull the index mid-conversation. For a product whose whole value is being cited, removing the signup step is the correct decision and a rarer one than it should be.

A market and a hedge sitting behind the index

The index is the free front door. The nav bar has two more doors. Buy is a marketplace selling GPU nodes by the calendar week, with live listings; on the day we looked, H100 nodes were $2.32 per GPU-hour for weeks 33 to 35 and $2.10 for weeks 42 to 45. Hedge sells cash-settled price floors on GPU compute: pay a premium, and if the index average falls below your floor you receive the difference on your covered volume.

How much does Computable GPU Index cost?

The index itself is free. The live page, the REST API, the flat files and the MCP server all cost nothing and need no account, and the collector code is Apache-2.0 on GitHub. If all you want is a defensible number to put in a board deck or a procurement model, you are done and you have paid nothing.

The money is in the two products the index feeds. On the marketplace, the platform fee is 2% of an order’s total value and it is charged to both the maker and the taker of every trade, so a round trip costs 4% of notional. Auction bids need a 10% deposit that is forfeited in full on default, and capacity not redeemed by the end of its window expires with no residual value, refund or credit. On the hedging side there is a free tier giving indicative quotes and 30 days of index history, a Pro tier at $500 a month adding unlimited repricing, custom floors and indices, full index history and CSV export, and a bespoke tier for eight-figure structures papered bilaterally.

Now the part that matters most, and it is in the data licence rather than on the pricing page. The code is Apache-2.0 but the index values are CC BY-NC 4.0, which is a non-commercial licence, and section 2 goes further. Using any index value as the basis of a financial product, including the settlement value, strike, trigger, margin or payout of any future, option, swap, forward or structured product, requires a separate written licence. So does non-display use in a commercial pricing, risk or execution system, and so does redistribution in a paid data feed.

Read that against the audience the homepage names, which is people who trade, rent or finance compute. Almost every one of them is a commercial user, and the single most valuable thing you could do with a GPU reference rate is settle a contract on it. The one party that can do that freely today is the company that publishes the index and also sells the derivative. That is a normal commercial arrangement for an index provider and it is disclosed precisely, in writing, in the repository. It is also not what most people hear in the phrase “open-source price index.”

Who is Computable GPU Index best for?

Use CGI if you buy GPU capacity and need a defensible number. It is free, it is citable, the method is published, and you can hand a finance team a command that reproduces the figure you quoted. For budgeting, procurement negotiation, board reporting or writing a compute line into a model, it is the best free option we have seen and the MCP server makes it trivially available inside an agent.

Use it too if you are researching the shape of the GPU market rather than a single price. The live panel view showing $16.00 at one provider and $5.63 at another for the same accelerator on the same day is a more useful finding than the index level itself.

Skip CGI if you need a settlement rate. The licence says no without a separate deal, and building a product on a number you are not licensed to use that way is a problem you discover late. Skip it if you need long history: the index has been running since 10 August 2026, so there is under a month of record and no meaningful cycle in it. And treat the marketplace and the index as two separate decisions rather than one, because the company that prints the reference rate also sells the asset it references, and on the day we looked its own weekly H100 capacity was listed 32% to 39% below its own on-demand H100 print. Those are different products, weekly reserved against on-demand, and a reader skimming both pages will not see that.

Best Computable GPU Index alternatives

Silicon Data is the established competitor and the direct comparison is instructive. It publishes an H100 rental price index on a Bloomberg ticker, SDH100RT, standardised for rental term, cluster scale and interconnect, with separate neo-cloud and hyperscaler readings. Its neo-cloud H100 reading is $2.53 per GPU-hour against Computable’s $3.43 on-demand print for the same accelerator, which tells you how much the panel and the standardisation rules decide the answer. Basic is free with 30 days of history, Pro is $998 a month with a 7-day trial, Enterprise is quoted per organisation. It is the pick if you need history, breadth and a ticker; CGI is the pick if you need the calculation itself.

Provider price pages and the Vast.ai console API are the free do-it-yourself route, and CGI’s own receipts show exactly what you get: real prices, no standardisation, and a $5.63 to $16.00 spread you now have to reason about yourself. Fine for a spot check on one provider, unusable as a reference rate.

SF Compute attacks the same market from the trading side with an order book for GPU clusters rather than an index. It is the alternative if what you actually wanted was to buy compute at a market-cleared price rather than to measure it, though it publishes no reference rate you can cite.

Final verdict: is Computable GPU Index worth it?

CGI is the most carefully built free thing we have reviewed in a while. The reproduce command works, the governance rules read like they were written by people who have been burned by a revised series before, published values are never restated, and the whole distribution stack is open to anonymous clients including agents. Compute has been rented, resold and financed at commodity scale for three years without the reference rate every other commodity has, and someone with the right background is finally building one in public. Take it, cite it, and check it yourself, because you actually can.

Two things keep it off an 8. The licence contradicts the pitch: an index sold as open carries a non-commercial data licence and reserves every settlement use for a separate written deal, which means the headline use case is the one you cannot have. And the index publisher runs the market in the underlying and sells the derivative that settles on it, with panel membership decided by Computable alone, no independent oversight committee, and a documented route for a provider to have its collection paused within five business days. Full disclosure and open code are a real mitigation and they are more than any closed index offers. They are not the same thing as independence, and a reference rate is a claim about independence. Use it as the best free market observation available. Do not treat it as a settled benchmark yet.