Carmen Li

Carmen Li is the founder and CEO of Silicon Data and CEO of Compute Exchange. She builds pricing data and trading infrastructure for AI compute markets, helping buyers understand their costs, hardware owners assess potential earnings, and both sides manage changing prices. Her work brings experience in commodity trading and financial data to the economics of GPUs.

From energy trading to compute markets

Li began her career at DRW in Chicago as an energy-options market maker, developing experience in commodity pricing, volatility, and risk management. She earned an MBA from Harvard Business School, then worked in strategic planning at American Express and global working-capital products at Citi. At Bloomberg, she led strategic alliances for enterprise data and managed global data integrations, connecting financial information with customers’ systems, cloud services, and AI applications. Her career spans trading, financial products, and data infrastructure.

Conversations with AI startups at Bloomberg brought a familiar financial problem into view. Companies could win customers paying fixed monthly subscription fees while still losing money because their GPU costs fluctuated. Adding customers did not necessarily fix their margins. Li saw a need for the benchmarks and risk-management tools already used in other commodity markets.

She left Bloomberg and founded Silicon Data in 2024. DRW founder Don Wilson, whom she knew from her trading career, provided early backing. The company developed GPU pricing benchmarks and indices to make compute costs easier to measure and compare. Li subsequently took on leadership of Compute Exchange, adding a marketplace business alongside the data company. Her founding story connects two strands of her earlier work: understanding market risk and getting useful data into customers’ systems.

Pricing information and a place to transact

Silicon Data supplies indices, valuations, and market intelligence. Compute Exchange handles transactions involving compute capacity, forward contracts, and physical servers. Together, the businesses address two needs: reference prices that help participants evaluate a deal, and a marketplace where they can arrange it. Compute Exchange describes these complementary roles as the intelligence and marketplace functions of compute finance.

Running both has required Li to adapt her management approach. She describes a data business and a marketplace as having different cultures, margins, and routes to customers. Their related products do not make them interchangeable organizations; lessons from one still need reconsideration before she applies them to the other. Her account of running the two companies emphasizes how much she has had to learn after moving from established financial firms into entrepreneurship.

How she approaches compute economics

• Hedge the cost of production. Li argues that AI companies should be able to secure compute costs and concentrate on their products. Hardware owners face the reverse exposure: a server purchase financed over seven years might have a customer reservation covering only three to five, leaving later rental income uncertain. Benchmarks and hedging contracts can help participants understand and manage that exposure, although the availability of a benchmark does not guarantee a willing counterparty for every term.

• Make unlike machines comparable. A GPU model name does not describe a complete server. CPU, memory, storage, networking, and location can affect performance and price. Silicon Data normalizes different configurations to a common base before calculating benchmarks, then monitors how their pricing relationships change. A regional premium, for example, may reflect demand for lower latency. Comparable pricing requires continuing measurement rather than a simple average of advertised rates.

• Assess hardware through the work it can still earn. Li challenges the assumption that a new chip generation immediately eliminates older hardware’s economic usefulness. She points to older A100 GPUs still attracting three-year contracts, while explicitly cautioning that this does not establish a nine-year earning life for every unit. Cooling requirements, delivery times, capital constraints, and workload suitability can keep older equipment productive. Software improvements such as quantization may also allow models to run on hardware they previously outgrew.

These ideas shape Li’s approach to compute economics. For Compute Exchange, she wants procurement to take hours rather than months. For Silicon Data, she wants businesses to understand what their hardware is worth, how their costs and revenues can change, and which risks they can hedge.