Carmen Li and the price nobody could agree on

A software subscription makes a tidy promise: pay the same amount each month, receive the service. Add artificial intelligence, and the promise acquires a less tidy companion. Somewhere, a computer is doing work on the customer’s behalf. Its owner wants paying, too. The two bills can behave very differently.

Carmen Li noticed that mismatch among AI businesses while working around financial data. Their customers paid predictable subscription fees. Their GPU costs moved. Winning more customers could increase exposure to a price they could neither confidently forecast nor readily hedge. Growth, usually the answer to a startup’s problems, had become an additional question.

For someone trained in trading, this was a familiar sort of discomfort in an unfamiliar place. A business was exposed to an essential input, yet lacked the financial tools used elsewhere to manage that exposure. Li began building Silicon Data in April 2024. Her subject was compute. Her starting point was the cost of keeping a commercial promise.

A career spent making numbers usable

Li grew up in Hangzhou, China. At 15, she won a scholarship to study in Singapore and had one hour to accept it. She did. At an all-girls boarding school, keeping the funding required straight As. She later chose the United States over a scholarship at the National University of Singapore, studying finance and computer science at Illinois. Decisions about uncertain futures had arrived early.

Before there were GPU indices, there were option models and trading platforms. At DRW in Chicago, Li built both, then moved into trading. She later attended Harvard Business School and worked at American Express and Citi. At Bloomberg, she managed data partnerships and integrations with outside businesses. The career joined two demanding activities: deciding what a price means and making its underlying data useful to somebody else.

Her Bloomberg work exposed her to the routes financial information takes through other people’s systems: cloud providers, databases and the businesses using them. That experience gives her compute project a practical emphasis. Producing data is one task. Getting it into a form that another business can rely on is another. The distinction helps explain why she talks about infrastructure alongside the underlying prices.

Her relationship with DRW founder Don Wilson endured. In March 2024, a conversation in Boca Raton turned to compute-price risk. Wilson agreed to provide seed capital. Silicon Data followed the next month. A professional connection built in trading had found a new application.

Li can be entertainingly frank about the confidence required to start. She has described feeling unusually well placed to connect GPUs with derivatives and market data, while cheerfully acknowledging her own bias. Her account allows room for both expertise and the founder’s necessary leap. Experience supplied a reason to begin; it could hardly supply a guarantee.

“Think of me as the Bloomberg for GPU pricing.”Carmen Li / July 2025

The comparison explains the job she wanted Silicon Data to perform. Buyers need information they can use before making a transaction. Lenders need numbers they can examine when financing equipment. A dashboard might look like a screen full of prices. The actual product is an agreed basis for a decision.

The chip on the invoice, the machine in the room

Silicon Data’s first daily H100 rental index, announced in May 2025, put a GPU reference price onto Bloomberg terminals. Its construction required adjusting for differences in configurations, locations and rental conditions. The neat little number at the end had a considerably less neat job behind it: deciding which offers could fairly be compared.

Think about the buyer’s predicament. Two offers can carry the same GPU name while differing in the surrounding system. A price per hour is useful only after asking what that hour buys. Otherwise, the spreadsheet performs a splendid impression of precision while leaving the expensive questions unanswered.

Li also wanted a way to examine delivery. SiliconMark addresses GPU and cluster performance. SiliconNavigator supplies market and pricing information; SiliconPriceIQ adds pricing analytics; SiliconCarbon estimates the emissions associated with compute use. The initial suite reflected several questions that arrive together when someone actually has to procure capacity.

The distinction between price and performance runs through her thinking. A system can be cheaper for a reason. It can also be expensive without delivering the expected output. Measuring the machine gives the buyer something firmer than the provider’s description, and gives a capable provider a way to demonstrate its work.

Two companies, two kinds of work

In a November 2024 conversation about becoming a first-time founder, Li discussed agile teams and encouraging more women to pursue startups. Entrepreneurship was still a new role for someone who had worked inside established financial businesses. The move put her accumulated experience to a different test: choosing which problems to solve first, building products around them and assembling the people to do the work. Her public account gives the career change room to feel like a transition.

By October 2025, Li was publicly discussing a second CEO role at Compute Exchange. Silicon Data supplies measurements and intelligence. Compute Exchange provides a place to transact in compute capacity. Running both brought the information business into close proximity with the practical business of arranging commitments between buyers and sellers.

She admitted being unsure beforehand about what the schedule would feel like. Two weeks into the arrangement, she said the work had made her more focused and creative. Her analogy was domestic: two children, each with a different personality and different needs. Anyone expecting a diagram of corporate synergies received something rather easier to recognize.

There is a useful tension in those responsibilities. A marketplace has to help a deal happen. A benchmark has to retain its meaning regardless of who likes the resulting number. Li’s account of the companies emphasizes their complementary roles, while her explanations of index governance make independence central to the measurement business.

By August 2026, Silicon Data announced a $30.5 million initial Series A closing, led by the Valor Atreides AI Fund. The company said the funding would support pricing benchmarks, performance measurement, institutional datasets and risk infrastructure. Its ambitions had widened from helping someone inspect a quote to supplying references for institutions underwriting the compute economy.

Putting time into the price

A price today answers one question. An infrastructure commitment asks many more. Someone financing equipment needs to consider what it might earn later. A business reserving capacity needs to weigh the commitment against demand that may change. Time turns a shopping decision into a balance-sheet decision.

Li’s writing on GPU forward curves tackles that extension. A term rate describes a rental commitment available now. An implied forward rate is derived from the structure of those commitments. It offers information about how the market currently prices future periods. That distinction deserves care: today’s pricing of tomorrow can change before tomorrow arrives.

It also changes the conversation with a lender. Instead of relying entirely on an equipment owner’s revenue projection, the parties can examine a market reference, then discuss assumptions and downside scenarios. A shared starting point cannot remove uncertainty. It can make the disagreement specific enough to work on.

CME Group’s proposed H100 and B200 rental-index futures take the reference-price idea into financial contracts. The published design specifies cash settlement, 730 GPU-hours per contract and monthly maturities extending across 36 consecutive months. Participants would gain exposure to changes in an index; the contract itself would settle financially rather than allocate a particular machine.

The market remained a proposal as this profile was prepared on October 25, 2026. The CFTC had extended its review through November 9, and CME’s revised notice listed the trade date as TBD. That pause belongs in the story. A published benchmark, an announced contract and an operating futures market are separate milestones.

The distinction also keeps Li’s ambition in proportion. Financial infrastructure requires cooperation beyond the founder’s company. Exchanges, customers, counterparties and regulators each have work to do. The appearance of a contract specification is evidence of a project taking shape, with the harder test of actual use still ahead.

A referee has to earn the whistle

In her more recent discussions, Li has emphasized how much interpretation sits inside an apparently simple index. CPU, memory, storage and geography can distinguish machines sharing a GPU label. Normalize carelessly and the average conceals the very differences a customer needs to understand. Publish responsibly and the methodology becomes part of the product.

She has also described keeping herself outside the index calculation committee, leaving the technical responsibility to research and index specialists. Her argument is straightforward: a benchmark provider needs a structure that protects the measurement. Being convinced that compute will become a major market does not entitle anyone to choose the day’s answer.

“Every mature market eventually develops independent referees”Carmen Li / August 2026

That metaphor gives her career a pleasing destination. A trader makes decisions about prices. A data executive helps information reach the people who need it. Li is now trying to build a business whose measurements other people can use to make decisions of their own. The ambition requires becoming useful to parties who may disagree with one another quite vigorously.

It is an unusually concrete way to approach the future of AI. Start with a bill. Ask what was purchased, whether it performed, and how the economics might change. Then keep doing the measuring. For all the money and machinery surrounding compute, Li’s work comes back to a conversation recognizable in almost any business: before we sign, can we agree what this is worth?