NextFin News - The artificial-intelligence arms race is entering a new phase: the contest is no longer just about who can manufacture the most advanced chips, but about who can price, hedge, and finance the compute those chips deliver. Carmen Li, founder and chief executive of Silicon Data, made the point on the sidelines of the 33rd CITIC Securities International Investors' Forum in Hong Kong, arguing that the center of gravity in AI is moving from semiconductor fabrication to the economics of GPUs and data centers.
The shift matters because it changes what investors should be watching. For the past two years the market has fixated on chip shipments, wafer starts, and fab capacity. Li's argument is that the next battleground is the financial infrastructure layered on top of that hardware: transparent benchmarks, forward curves, futures contracts, and the unit economics of the AI tokens those machines produce. The company that once needed a semiconductor analyst now needs a commodities trader.
The Price Signal: GPUs Are Defying the Normal Cycle
The clearest evidence that something unusual is happening sits in Silicon Data's own pricing indices. The firm's Neo Cloud H100 index rose from 2.20 to 2.64, a 20% gain over three months. The Neo Cloud B200 index, tracking Nvidia's next-generation Blackwell chip, climbed from 4.40 to 5.35, up 22%. Even the Hyperscaler H100 index ticked up, from 7.26 to 7.46, a 3% rise.
These moves run against the textbook pattern for hardware. In a normal cycle, prices spike when a new chip launches, then gradually ease as supply catches up and the product moves down the adoption curve. Instead, B200 pricing has stayed elevated and kept climbing months after launch, a sign that capacity constraints are still binding across the AI stack - chips, memory, power, and data-center space. The constraint is not a single bottleneck but a chain: even when GPUs arrive, networking, cooling, and grid connections can keep them from being deployed at full utilization.
"The price is going pretty nuts," Li said of H100 rental prices.
The dispersion between venues is as telling as the direction. An H100 costs almost three times as much to rent from a hyperscaler such as Amazon, Microsoft, Google, or Oracle than from a neocloud provider such as CoreWeave. Customers pay the premium for convenience, existing integrations, and guaranteed access to capacity inside environments where their data and workflows already live. That spread is itself a market inefficiency - two buyers of identical GPU hours paying wildly different prices with no public way to know who got the better deal - and it is precisely the kind of opacity that benchmarking businesses exist to eliminate.
Perhaps the most striking data point is depreciation. In the second year of ownership, a refurbished H100 can still be sold for 85 cents on the dollar; in year three, 84 cents. "My car depreciates a lot faster than that," Li said. For the hundreds of billions of dollars of GPUs now sitting on tech-company balance sheets, slow depreciation is the difference between clean earnings and a future impairment charge. So far, the feared write-down cycle has not arrived - because demand for compute continues to outstrip supply, keeping the resale market for used accelerators unusually firm.
Financializing Compute: Benchmarks, Futures, and the CME Deal
Li is not merely observing this market; she is helping to build its plumbing. On August 11, 2026, Silicon Data announced an initial closing of $30.5 million in Series A funding, led by the Valor Atreides AI Fund, with participation from CME Ventures, DRW, F-Prime, Samsung Next, VanEck, Further, Jump, Tectonic, and Wintermute. The same day, CME Group confirmed plans to launch two compute futures contracts on October 5, 2026, pending regulatory review, listed on NYMEX.
The contracts - the Silicon Data H100 Rental Index Futures (ticker GPU1, market code SDH100RT) and the Silicon Data B200 Rental Index Futures (GPU2, SDRB200RT) - each represent one month of hourly GPU rental capacity, or 730 GPU-hours, and are financially settled. Each tracks a daily neocloud index published by Silicon Data. The value per tick is $7.30, and monthly contracts will be listed out 36 months, giving builders a view of expected compute costs three years ahead.
"Compute futures give the market something it's never had: a public, tradable reference price for the resource every AI system runs on," Li said. "Silicon Data's benchmarks make that price real; CME makes it tradable. Together, that turns compute from something enterprises negotiate blindly into a market they can actually plan around."
The analogy CME executives draw is deliberate and historically grounded. Oil fueled the 20th-century economy and evolved from opaque, bilateral spot trading into a global derivatives market with transparent benchmarks such as Brent and WTI. Compute is being set up to follow the same arc: first fragmentation and information asymmetry, then independent price discovery, then standardized, exchange-cleared risk transfer. Pete Keavey, global head of energy and environmental products at CME Group, called compute "the currency of the AI age" and said the contracts would turn it into "a standardized, tradable commodity." Gavin Baker, managing partner and CIO at Atreides Management, put the financing angle more bluntly: "Futures markets are what let farmers finance the seed and equipment for next season instead of guessing. You can't build against a price you can't see or lock in."
The infrastructure behind the futures is deeper than price alone. Part of the Series A proceeds funds SiliconMark, which independently benchmarks how physical GPU infrastructure actually performs - because two clusters built on identical chips can deliver meaningfully different real-world output depending on networking, topology, and configuration. Normalizing performance, not just price, is what would eventually allow physical delivery of compute resources rather than cash settlement. That distinction matters: a cash-settled contract tells you what the market expects to pay; a deliverable contract tells you what you can actually obtain.
The second-order effect of a listed futures curve is that it changes corporate behavior before a single contract trades. CFOs at AI labs and cloud providers will begin modeling GPU costs against a public forward curve rather than against private broker quotes. Procurement teams negotiating annual reserved-capacity deals will have a reference point for what "fair" looks like. And lenders underwriting data-center construction or GPU-backed loans will have an observable collateral value. Price transparency does not just inform trades; it lowers the cost of capital for the entire buildout.
Tokenomics: When Tokens Become an Export Commodity
The "tokenomics" in the discussion points to the demand side of the equation, and it is where the AI race becomes a race against unit economics. AI products generate revenue by producing tokens - the units of text, image, or code that models emit - and the largest single cost of producing each token is rented GPU time. When GPU rental prices stay high, the marginal cost of a token stays high, and the gross margins of every AI application built on rented compute remain under pressure. Efficiency gains from model architecture and inference optimization are, in effect, racing against stubborn rental rates.
Li has extended the logic beyond individual companies to entire countries. In a separate panel appearance, she argued that nations may soon talk about "exporting tokens as the next export product." A country cannot export electricity or energy directly across borders, but it can build data centers powered by domestic energy and sell the resulting tokens - effectively converting stranded or low-cost power into a tradable digital export. The Middle East, with its low-cost gas and sovereign capital, and regions with surplus hydroelectric or geothermal capacity, fit the model naturally.
This reframes the AI buildout as an energy-arbitrage trade dressed in semiconductor clothing. The countries that win the next phase may not be the ones with the best chip designers, but the ones with the cheapest, most abundant power and the regulatory capacity to permit and build data centers quickly. Token output becomes the measurable unit of that conversion, and token cost per unit becomes the metric by which AI products live or die. An AI assistant that costs $0.002 per token to produce competes on entirely different terms from one that costs $0.02 - and the gap between them is largely a function of where the GPUs sit and how much their owners paid to rent them.
There is also a currency dimension. Countries that export tokens earn foreign exchange from a service that consumes domestic electricity and domestic capital equipment. For economies seeking to diversify away from raw commodity exports, token exports offer a value-added outlet for the same underlying resource base. The idea remains early-stage, but it illustrates how thoroughly the financialization of compute could reorganize global trade patterns.
The Counter-Thesis: What If the Growth Rate Rolls Over?
The bear case against this narrative is straightforward and deserves weight. Hardware cycles normalize; that is the rule, not the exception. The relevant question is not whether absolute spending falls - it is not projected to. A consensus of sell-side estimates puts combined hyperscaler capital expenditure at roughly $602 billion in 2026, up about 36% from 2025, after a 73% jump the year before. Growth remains strong, but the pace of acceleration is roughly halving.
That deceleration is the crack in the bullish thesis. Every percentage point of slowing capex growth reduces the order flow that keeps GPU rental rates firm. Meanwhile, a wave of new supply - Nvidia's ramped Blackwell output, custom silicon at the large cloud providers, and neocloud capacity financed during the boom - is scheduled to come online through 2026 and 2027. If AI inference demand fails to compound fast enough to absorb that supply, the price indices that now look so firm could reverse quickly. Rental rates would fall, the refurbished market would soften, and the depreciation story would flip: GPUs bought at peak prices would become impairment candidates on balance sheets. The futures contracts, designed as hedges, would then reveal just how much volatility the market had been underpricing.
Li's own data contains the seeds of this risk. The threefold premium that hyperscalers charge over neoclouds is a sign of scarcity rents, not structural value. Scarcity rents attract capacity. Capacity, once built, does not un-build. And the refurbished H100 holding 84 cents on the dollar in year three is a function of current demand, not a law of physics.
The counter-thesis is credible enough that it defines the falsifying signal for the bullish view: if the Neo Cloud B200 index trades below 4.40 - the level it held earlier in 2026 - for two consecutive months, or if hyperscaler AI capital expenditure growth slows to single digits year over year, the structural-tightness narrative is wrong and the market is in a cyclical peak, not a regime shift.
What Comes Next: Three Horizons
Short term (sentiment and liquidity): The launch of CME compute futures on October 5, 2026, is the immediate catalyst. Initial volumes will be thin - new commodity contracts typically take quarters to build liquidity - but the mere existence of a regulated reference price will change how procurement teams negotiate and how CFOs model AI costs. Watch the first week's open interest and the spread between the futures curve and spot indices. A futures curve trading at a steep premium to spot would signal that buyers expect scarcity to persist; a curve at a discount would suggest the market expects relief.
Medium term (fundamentals): The question is whether token demand grows fast enough to absorb the capacity coming online. If AI application usage compounds at current rates, rental rates hold and the neocloud-hyperscaler spread narrows only modestly, because the premium for integration and reliability retains value. If usage growth disappoints - if the killer application proves to be a suite of modest-margin tools rather than a usage explosion - the spread collapses toward the marginal cost of power and the depreciation story reverses. The medium-term watch item is the ratio of AI revenue to AI capex at the large cloud providers; as long as that ratio climbs, the buildout is self-financing. If it stalls, the pressure moves upstream to GPU lessors and equipment lenders.
Long term (structural): The financialization of compute - benchmarks, futures, performance verification, forward curves - is likely durable even if prices fall. Markets that develop transparent price discovery rarely return to opaque bilateral negotiation. The regime shift, in other words, is in the plumbing, not necessarily in the price level. Even in a price downturn, the participants who survive will be those who used the new tools to hedge, and the benchmarks will outlast the cycle that created them.
Base case: compute prices remain elevated through 2026 as capacity lags demand, futures volumes build slowly, and token economics improve only incrementally as model efficiency gains offset stubborn rental costs. Upside case: a killer application drives token demand that outstrips even the planned capacity additions, pushing rental indices higher and pulling neocloud valuations with them. Downside case: hyperscaler capex deceleration accelerates into an absolute cut, leaving a glut of GPU hours and forcing depreciation charges that ripple through mega-cap earnings.
The AI race has not ended; it has simply changed terrain. The winners in the chip phase were the companies that could fabricate the most advanced silicon. The winners in the next phase will be the ones that can measure it, hedge it, and convert electricity into tokens at the lowest cost. Compute is becoming a commodity - and commodities, eventually, reward the low-cost producer, not the fanciest factory.
Explore more exclusive insights at nextfin.ai.
