Silicon Data Raises $30.5M to Build AI Compute Benchmarks

AI companies have spent years obsessing over how many GPUs they can get. Wall Street is beginning to ask a different question: what should that computing power actually be worth?

That question has created an unusual opportunity for Silicon Data, a startup founded by former Bloomberg data executive Carmen Li. Instead of building AI models or data centers, Silicon Data is trying to build something much more familiar to financial markets, a reliable price benchmark.

The timing may be ideal. Goldman Sachs estimates that roughly $7.6 trillion could be invested in AI compute, data centers and power infrastructure between 2026 and 2031. At that scale, GPU capacity is no longer just an IT expense. It starts behaving like a major economic input whose price companies may eventually need to hedge in the same way airlines hedge fuel or manufacturers hedge metals.

The numbers show why a benchmark matters

GPU prices can differ dramatically depending on where capacity is rented.

Silicon Data's August 19 benchmarks put Nvidia H100 rental capacity at about $2.68 per GPU-hour on neocloud platforms, while hyperscaler capacity was priced around $7.28 per hour. Its newer B200 benchmark stood at roughly $5.65 per GPU-hour, while A100 capacity ranged from about $1.65 on neoclouds to $3.73 at hyperscalers.

Those spreads matter when a company is renting thousands of GPUs for months at a time.

They also move. Silicon Data's B200 index began 2026 around $4.40 per hour and reached $6.11 in late March. The company recorded a maximum single-day move of nearly 19% during that period. Even the older H100 and A100 markets have experienced renewed pricing pressure as inference demand grows.

In other words, the industry's most important input does not yet have anything resembling the universally understood price benchmarks available in oil, electricity or metals.

Silicon Data wants to change that.

From $4.7 million seed round to Wall Street infrastructure

Silicon Data emerged from stealth with a $4.7 million seed round backed by trading firms DRW and Jump Trading. At the time, it had already aggregated about 3.5 million market data points across 50 GPU chipsets and platforms.

Now the company has announced a $30.5 million initial Series A closing, led by the Valor Atreides AI Fund.

CME Group, DRW, F-Prime, Samsung, VanEck, Further, Jump, Tectonic and Wintermute also invested, alongside several other venture firms. Combined with its earlier seed round, Silicon Data has disclosed at least $35.2 million in funding.

The funding will go beyond collecting GPU prices. Silicon Data plans to expand performance benchmarking, institutional market data and infrastructure for derivatives, insurance and credit markets.

That last part is important because Silicon Data is increasingly looking less like a conventional AI startup and more like a financial-data company built around AI infrastructure.

CME plans to make GPU prices tradable

The biggest validation of that strategy may arrive on October 5, 2026.

CME Group plans to launch two compute futures contracts tied to Silicon Data's benchmarks, pending regulatory review: one tracking Nvidia H100 rental prices and another tracking Blackwell B200 prices.

Each futures contract will represent roughly a month's worth of GPU rental capacity and will be listed under NYMEX rules.

The idea is straightforward.

An AI company worried that GPU rental prices could rise could lock in future costs. A cloud provider concerned that rental rates could fall could hedge its future revenue. Financial firms could also trade the contracts based on where they think AI infrastructure demand is heading.

This is essentially the financial machinery that already exists around oil, natural gas, electricity, metals and agricultural commodities — applied to computing power.

Regulators are now paying attention

The market is becoming significant enough to attract Washington's attention.

On August 19, the U.S. Commodity Futures Trading Commission moved toward seeking public input on compute derivatives as demand for AI infrastructure continues to accelerate. The regulatory scrutiny comes as CME, Intercontinental Exchange and other firms develop markets tied to computing capacity.

CME's planned contracts are therefore not guaranteed to launch exactly as proposed. Regulatory approval and, perhaps more importantly, market liquidity still have to develop.

AI compute is also harder to standardize than a barrel of oil.

An Nvidia H100 running in one data center may not deliver exactly the same effective performance as another because networking, memory configuration, cluster architecture, geography and software optimization all matter. Silicon Data itself has highlighted measurable performance differences even between nominally identical GPUs.

That complexity is one reason pricing data could become valuable rather than unnecessary.

The AI boom is becoming a capital-markets story

The larger shift goes far beyond Silicon Data.

Big technology companies are expected to spend hundreds of billions of dollars this year building AI infrastructure. Reuters reported that spending by major hyperscale AI providers including Google, Microsoft and Meta is expected to reach roughly $725 billion in 2026.

At the same time, Goldman Sachs' longer-term estimate suggests cumulative AI infrastructure investment could reach $7.6 trillion by 2031.

That amount of capital needs financing, valuations, collateral models, insurance and ways to manage risk.

Silicon Data is betting that before Wall Street can finance AI compute efficiently, it first needs to agree on what compute is worth.

If that happens, one of the most important financial tickers of the AI era may not represent Nvidia, OpenAI or another AI company at all.

It could simply represent the price of an hour of computing power.