EP55 Token Prices Dropping on Model Shifts with Carmen Li, CEO of Silicon Data & Compute Exchange

Summary

• Li attributes falling token prices to customers shifting their model mix toward cheaper alternatives. Silicon Data tracks consumption across 300 open-source/open-weight models plus closed models, observing “people shifting from premium models to cheaper models after the token-mixing era.” Doug suggests workflow routing could explain the shift, but Li cannot see customer use cases and does not confirm that mechanism.
• B200 on-demand rates were around $5 per GPU-hour at the global neocloud rate, while its implied forward curve still sloped upward. Last year’s downward-sloping curves had shifted into contango; more recently, A100 and H100 curves flattened. These curves come from reserved contracts, ahead of futures trading, and Li’s warning is explicit: “There are cycles.”
• Older GPUs can retain earning power beyond a simple depreciation schedule. Li sees three-year A100 contracts despite the architecture already being roughly four to six years old, and cites secondary prices of $4,000–$9,000 per GPU. Doug challenges whether that reflects durable economics rather than physical longevity; Li points to cooling constraints, replacement delays, and workflow suitability, while cautioning, “I’m not saying every A100 lasts 9 years.”
• Compute hedging addresses the mismatch between long financing commitments and shorter customer contracts. GPU financing can run five to seven years against reservations lasting three to five, leaving later rental income exposed. CME futures would cover the first 36 months, with longer hedges requiring OTC transactions; the objective is to let operators lock their “cost of production” and make financing easier to underwrite.
• Silicon Data’s commercial opportunity depends on trusted benchmarks becoming embedded in compute finance. Li says CME would launch H100 and B200 futures “in the next few weeks,” with four ETFs already filed, including a 1x inverse product. Her potential $30 trillion mature derivatives market is annual notional activity, not a forecast for data-licensing revenue.
• More chip designers do not automatically eliminate compute supply constraints. Doug stresses that fabrication capacity is expanding; Li’s response is that fabrication, packaging, infrastructure, and approvals must expand together. Otherwise, output remains capped by the slowest-growing part, while specialized inference chips may supplement Nvidia rather than replace its entire role.
• Li leaves the ultimate AI value-capture question open. Doug points to internet platforms that owned customers and distribution, while Nvidia currently captures substantial economic rent. Li calls AI “a gigantic baby”: the infrastructure, models, harnesses, and applications are still developing, and she does not claim to know which future layer will produce the biggest winners.

Deep dive

1. Fixed-price AI revenues create a need for compute hedges

• Li’s move from DRW energy-options trading and Bloomberg data integrations began with a 2024 customer problem: AI SaaS companies collected fixed monthly revenue while on-demand GPU charges swung from $3 to as much as $9 or $16 per GPU-hour. More clients did not necessarily repair negative margins. Her airline analogy carries the argument: “It’s not their job to speculate on GPU cycles.” Operators should secure production costs and concentrate on products.

• Her two companies address different functions. Silicon Data supplies indices, valuations, and risk-management information—the “referee layer for compute”—without providing GPUs or token access. Compute Exchange supplies liquidity through long-term and forward contracts on compute, GPUs, and tokens, as well as secondary server transactions. Li compares the businesses to Bloomberg and Nasdaq, respectively.

• The first natural hedgers, Li expects, are participants with concentrated GPU exposure, including neoclouds, hyperscalers, or financiers holding servers on their balance sheets. Her example is a GB300 purchase financed for seven years alongside a three-, four-, or five-year reservation: later rental income remains exposed even after securing an initial customer.

• Doug’s financing argument is that locking future income could make a creditor comfortable funding an otherwise uncontracted purchase. Li agrees, while distinguishing useful hedging tools from unlimited liquidity: CME futures cover only 36 months. Years four through seven would generally require block or OTC trades, informed by the futures curve rather than automatically supplied by it.

2. Benchmark credibility underpins the licensing opportunity

• At recording, Li says CME would launch H100 and B200 compute futures “in the next few weeks”; she expects options eventually, without announcing them. Four ETFs had already been filed based on the futures, including a 1x inverse product. She finds that unusually early: in her experience, issuers normally wait to observe futures liquidity before building ETFs around it.

• A B200 label does not describe a standardized server. Li says roughly 50 configurations can differ in CPU, RAM, disk, and geography. Silicon Data normalizes them to a base configuration and monitors changing relationships daily; a location premium may reflect latency-sensitive demand. Its data includes about 250,000 rows daily, and licensing arrangements can include exclusive rights or a paying-client relationship.

• Doug presses on conflicts, including DRW’s investment and opportunities for index arbitrage. Li says DRW is “not involved,” by design. She says she is outside the index calculation committee because the head of research and head of indices are better suited to the job, and that she will work with external academic advisers. She also says index providers should have a third-party-audited compliance program and separate index calculation; only a few participants do those things. Her governing principle is neutrality: “You can’t participate and be the referee at the same time.” An index that stops reflecting supply and demand loses its usefulness regardless of competition.

• Doug asks how much revenue the data itself could generate. Li instead offers a conditional market-sizing framework: NVIDIA’s spot market notional transactions could be $1–$3 trillion annually today, with mature compute derivatives activity potentially ten times spot, reaching $30 trillion at the upper end. Her licensing argument is that heavier trading creates demand for granular benchmarks, regional exposures, and risk tools; she supplies no corresponding revenue estimate.

3. A100 contracts challenge a uniform depreciation story

• Doug raises the bear argument associated with Michael Burry: extending server useful lives could overstate earnings. Li’s counterexample is transactional: A100s have existed for roughly four to six years, yet she still sees three-year contracts being signed. For maintained units, she says, six- and three-year contracts can mean at least nine years of earning cash, but explicitly limits the inference: “I’m not saying every A100 lasts 9 years.”

• Doug’s pushback—worth keeping—is that physical survival differs from economic usefulness: a newer chip might justify replacement through productivity and energy savings. Li agrees newer architectures can be more efficient and support larger models. But cooling systems, floor strength, available capital, and delivery schedules constrain replacement; buyers might wait six months or longer for new equipment while secondhand equipment is available “tomorrow.”

• Li quotes A100 secondary prices of $4,000–$9,000 per GPU. Separately, she says H800 residual values have been declining but remain above the five-year straight-line depreciation path. Her point is an uneven, longer earnings tail—not that older hardware avoids depreciation or that every server deserves the same accounting life.

• Software can also change hardware economics. Larger models may outgrow an older GPU’s memory, but better optimization or quantization might make those models deployable on older chips again. That could change demand and pricing: the GPU layer’s value depends partly on what happens in the software and model layers above it.

4. The forward curve has already changed its view of the cycle

• Most GPU consumption still happens inside hyperscalers, Li says, but neoclouds have expanded from a handful two years earlier to more than 300 globally, with a new provider appearing every day. More suppliers mean more customer options and harder-to-manipulate benchmarks; transaction data can also come from customers’ bills, allowing verification beyond the provider’s own account.

• Over the preceding 30 days, most chip prices had stabilized, with B200 global neocloud on-demand rates around $5 per GPU-hour. Silicon Data’s implied forward curves aggregate reserved contracts, including 36-month and five-year terms. Last year’s curves sloped down; this year they moved into contango. Recently, A100 and H100 flattened, while B200 remained in contango.

• Doug’s framing: “The only thing I know about the future curve is that it’s always wrong.” Li agrees that a curve represents today’s consensus rather than knowledge of future spot prices. Asked whether B200 finishes the year above $5, she says, “I actually don’t know.” She also declines forward guidance or trading exposure as an index provider.

5. Token prices reveal changing consumption rather than model quality

• Silicon Data’s token indices estimate how much society would pay if it were a person. Li tracks consumption across 300 open-source and open-weight models plus closed models, accounting for different input and output prices and volumes. The product measures observed spending choices; it does not grade model quality or merely repeat a premium provider’s published tariff.

• The mechanism Li emphasizes is migration: people are moving at least some workflows toward cheaper, open-source and open-weight models “after the token-mixing era.” Doug proposes that harnesses route simpler tasks away from frontier models. Li leaves that explanation conditional: “We don’t really have information about customer use cases.” Summarization is her example of work that might not require the strongest model.

• Li compares the token index with an altcoin revenue chart, describing it as “almost a leading indicator.” When the indices ramp up, altcoin revenue ramps up; the index later crashed to about a dollar, around when revenue began to stagnate. She frames the index as willingness to pay; she does not establish causation or identify why each customer changed models.

• Token services remain largely on-demand, unlike GPUs with reserved pricing. Li warns providers can change price or precision without clear disclosure. Doug wants controls against agents generating surprise bills; Li adds quality, deadline, latency, security, and location preferences. Owning or renting a GPU offers a different consumption structure—run models against hardware capacity—although, as Doug reminds her, “Depreciation’s real.”

6. Chip competition still runs into physical bottlenecks

• Doug asks whether hyperscaler-designed chips command premiums or discounts. Li recalls AWS announcing price cuts on Trainium and other models but supplies no numerical benchmark spread. Her distinction is functional: some specialized inference chips aim to supplement Nvidia use cases; others seek to replace Nvidia for particular inference workloads. They are not interchangeable across every task.

• Li’s fabrication example: a factory previously making 100 chips for Nvidia might instead make 80 for Nvidia and ten each for two other designers. More designers alone have not increased total output. Doug stresses that factories are expanding; Li agrees, but packaging or another slower step can still cap delivered capacity. Design companies can multiply faster than physical production.

• Regulatory pauses add another possible constraint. When Li raises moratorium risks, Doug says he has not heard of the broad scenario she describes; she cites New York State saying there would be no data centers for a few years, while acknowledging the policy could change. She also rejects futures launches as inherently a bubble signal: futures allow bearish participants to short and contribute to price discovery in a market otherwise dominated by long exposure.

7. AI’s eventual winners remain open while market infrastructure develops

• Doug asks where investors should sit across fabrication, hardware, tokens, data, and applications. Li looks back to internet infrastructure and the platforms that ultimately captured value. Doug points to Google and Meta owning customers, and adds Amazon’s distribution. That historical outcome does not settle which AI businesses will repeat it.

• Li calls AI “a gigantic baby”: enormous already, but early in development. Nvidia currently captures substantial economic rent, Doug observes; Li argues the foundational infrastructure is still being built. Models, harnesses, and applications sit above it. Another layer of winners might emerge, but her answer remains, “I don’t know.”

• Running both companies has also challenged easy transfer of experience. Li describes a data business and a marketplace as having different cultures, margins, and routes to customers—like daughters with different personalities. Her entrepreneurial advice follows from the pressure: choose something sufficiently compelling to sustain constant learning and unexpected obstacles.

• Her concrete ambition for Compute Exchange is to reduce procurement from “2 months” to “3 hours,” serving smaller buyers as well as requests for 20,000 nodes. Silicon Data’s parallel ambition is trusted valuation and benchmarking across compute layers, so participants can mark exposures daily, understand financing costs, and decide how to hedge.