Computable GPU Index wants to be the CPI of AI compute

GPU compute has a pricing problem so severe that most buyers navigate it blind. Cloud vendors quote per-hour rates that vary by region, by instance family, by commitment term, and by whether you are buying reserved capacity or spot. Third-party resellers broker H100 clusters at prices that move weekly. Nobody publishes a consolidated, comparable number. The Computable GPU Index (CGI) launched this week as the first open-source price index for GPU compute, and it is trying to become the CPI of the AI hardware market.

The pitch is straightforward. CGI aggregates listed prices for GPU instances across major cloud providers and resellers, normalizes them to a per-GPU-hour basis, and publishes the result as an open dataset. Anyone can inspect the methodology, fork the code, or contribute new sources. The project launched on Product Hunt with a Discussion post attached, which suggests the team behind it wants community input on what gets indexed and how.

What is genuinely new here is not the idea of tracking GPU prices. Analysts at firms like TrendForce and SemiAnalysis have published GPU cost estimates for years. Cloud providers themselves publish pricing pages. What is new is the open-source framing. An index that anyone can audit, extend, and challenge is a different beast from a consultancy report that arrives as a PDF with a paywall.

The AI economy has a measurement gap. Model training runs cost millions of dollars, inference workloads scale with usage, and every serious AI company carries a compute line item that rivals its payroll. Yet there is no agreed-upon reference price for the underlying commodity. The Computable GPU Index is a bet that the market is ready for one.

The timing makes sense. GPU supply has loosened considerably since the peak shortage of 2023 and 2024. Nvidia’s production ramp, plus entrants like AMD’s MI300 series and the custom silicon from Google and Amazon, has created a multi-vendor market where price competition actually exists. When supply is tight and allocation is rationed, an index is less useful because buyers take whatever they can get. When supply is adequate, price discovery becomes the whole game. The CGI arrives at the moment when buyers finally have choices to compare.

The index also arrives at a moment when the cost structure of AI is under scrutiny from multiple directions. Publicly traded cloud providers report AI revenue growth but rarely break out GPU unit economics. Startups raise enormous rounds premised on compute costs that they often estimate from a single vendor quote. Researchers publish efficiency claims that are hard to compare across hardware. A transparent price index would give all of these parties a shared reference point.

There are real questions about whether the CGI can deliver on its promise. The first is coverage. GPU pricing is fragmented across dozens of providers, each with its own instance naming conventions, region multipliers, and discount structures. An index that misses the long tail of resellers will produce a distorted picture. The second is timeliness. Spot prices fluctuate by the hour; committed-use discounts change quarterly. An index that updates weekly will be stale for the most volatile segments. The third is methodology. Normalizing a p4d.24xlarge on AWS against an A100-80GB node on a reseller platform requires assumptions about utilization, power costs, and amortization that reasonable people will dispute.

None of these problems are fatal. The open-source model is precisely the right answer to the methodology question, because disputes can be resolved in public rather than in a vendor’s spreadsheet. The coverage problem can be solved incrementally, as contributors add sources. The timeliness problem is a design choice, not a flaw.

What the CGI cannot solve is the deeper opacity in GPU pricing: the negotiated deals. The largest AI buyers do not pay list price. They sign multi-year commitments with cloud providers, or they buy directly from Nvidia in volume, or they strike confidential agreements with resellers that never appear on a public price sheet. An index built from published prices will capture the retail market, not the wholesale market. That is a meaningful limitation, and the CGI team should be upfront about it.

The comparison to the CPI is instructive. The consumer price index does not capture every transaction in the economy, and it is frequently criticized for methodology choices. But it remains the reference point for inflation because it is consistent, transparent, and long-running. The CGI does not need to capture every GPU deal to be useful. It needs to be consistent, transparent, and long-running. If it achieves those three things, it will become the default citation for “what does GPU compute cost,” even if sophisticated buyers know the real number is lower.

For AI builders, the practical value is immediate. A founder pricing out a fine-tuning run, a researcher budgeting an evaluation sweep, a procurement officer sanity-checking a reseller quote, all of them currently rely on word of mouth and vendor marketing. The CGI gives them a baseline. The number will not be the final price they pay, but it will be the number they start from.

There is also a strategic value that is harder to quantify. A public price index shifts negotiating leverage. When a reseller quotes $3.20 per H100 hour, a buyer who can see that the indexed average is $2.80 has a concrete anchor. The index does not need to be perfect to change the conversation. It just needs to be credible.

The broader implication is for the AI economy itself. Every commodity market eventually develops a reference price. Oil has Brent, gold has the London fix, memory chips have DRAMeXchange. GPU compute is the most important new commodity of the decade, and it has been trading without a reference price for years. The Computable GPU Index is an attempt to close that gap, and the fact that it is open-source rather than a proprietary product is a meaningful choice. It signals that the creators believe the index’s value comes from adoption and trust, not from selling access to the data.

The index will succeed or fail based on adoption. A price index with no users is a spreadsheet. The CGI team has the Product Hunt launch, which gives it initial visibility in the developer community, but sustaining that requires ongoing maintenance, contributor outreach, and a clear governance model for how the index is updated. Open-source projects die from neglect more often than from competition.

The most interesting question is whether the major cloud providers will engage or ignore it. If AWS, Google Cloud, and Azure treat the CGI as a reference point, it becomes a de facto standard. If they ignore it, the index will track the reseller and secondary markets more heavily, which would still be useful but would tell a partial story.

The GPU market has matured to the point where it needs this kind of infrastructure. The Computable GPU Index is a small project with a large ambition: to make the price of AI compute legible. Whether it becomes the Brent crude of GPUs or a footnote depends on whether the community treats it as a public good worth maintaining. The first open-source price index for GPU compute is a start, and for a market this opaque, a start is worth having.