The Rise of Compute as a Financial Asset
- The AI industry is experiencing significant capital investment, with billions of dollars annually directed toward data centers and GPUs, making compute the primary expense for AI product development 10s.
- Silicon Data aims to establish a reference price for GPU rentals, positioning itself as the index against which Wall Street futures contracts would settle 10s.
- Steve How joined Silicon Data as the head of research in May, bringing experience from his previous role as a senior researcher in systematic equity indices at Bloomberg 42s.
- Bloomberg’s index business operates as a distinct vertical pillar, separate from the company's news and terminal platform businesses 42s.
Mechanics and Purpose of Futures Markets
- Futures contracts are ancient financial instruments designed to allow parties to lock in prices for future delivery of goods, such as crops, to mitigate the risk of price volatility 1m35s.
- A producer might use a futures contract to guarantee a sale price above their production costs, protecting against the possibility that market prices could drop by the time of harvest 1m35s.
- Futures markets function because participants hold differing expectations regarding future market conditions, with some individuals seeking to hedge against risk while others speculate on price movements to generate profit 1m35s.
- Futures markets allow entities to lock in prices for assets, a concept historically applied to commodities like oil to manage uncertainty regarding future events 0s.
Regulatory and Structural Challenges of Compute Futures
- The CME (Chicago Mercantile Exchange) plans to launch silicon data-linked compute futures on October 5, subject to regulatory approval 15s.
- While some question the classification of compute as a commodity due to issues like technological obsolescence and the lack of physical storability compared to oil, proponents argue that physical existence or storability is not a requirement for a commodity 25s.
- Many major futures contracts, such as those for the S&P 500, are cash-settled, meaning participants trade based on the profit or loss relative to the underlying index rather than taking physical delivery 42s.
- Even in oil markets, the majority of futures contracts do not result in physical delivery; the negative oil prices observed in March 2020 occurred because traders sought to avoid physical delivery when storage facilities were full 1m15s.
Market Demand and Hedging Strategies
- The scale of investment in compute justifies the creation of a futures market, with hyperscalers like Google and AWS spending approximately $750 billion on compute this year and projected to spend roughly $1 trillion next year 1m45s.
- The massive financial exposure to compute creates a demand for hedging, as entities may want to lock in earnings, manage uncertainty regarding market demand, or generate cash from their compute holdings 2m15s.
- Potential participants in the compute futures market include compute providers like hyperscalers and neoclouds looking to hedge downside risk, as well as AI labs and inference platforms seeking to hedge upside risk 2m45s.
Participants in the Compute Futures Ecosystem
- Financial markets typically consist of three distinct sides: natural longs, natural shorts, and market makers who facilitate transactions between the two 0s.
- Cloud providers and data center operators act as natural shorts because they possess available compute capacity and seek to lock in rental income through futures contracts 15s.
- AI labs, such as Anthropic or OpenAI, and inference platforms that run open-source models act as natural longs because they purchase compute and have an incentive to hedge against rising prices 35s.
- Enterprises are less likely to buy compute directly, instead opting to purchase AI tokens or model tokens from AI labs and inference platforms 45s.
- Market makers, such as DRW, Jane Street, or HRT, serve as intermediaries to create a "double coincidence of wants" when the timing of buyers and sellers does not align 1m5s.
- The composition of market participants can vary, as large financial institutions with exposure to AI data center income may seek to hedge their existing positions regardless of their classification as a natural long or short 1m25s.
Portfolio Management and Risk Mitigation
- Trading compute futures allows companies with "skin in the game" to hedge their bets or adjust their exposure without needing to physically build out infrastructure or sell off tangible assets 1m45s.
- Financial instruments like futures contracts are not exclusively for betting on outcomes; they serve as tools for participants to dynamically adjust their portfolio exposure 2m5s.
- An analogy for this hedging process is a real estate owner in New York City who, fearing a market downturn, could use a real estate index futures contract to reduce their exposure without having to sell their physical apartments 2m20s.
- Portfolio adjustments allow individuals and entities to manage risk, such as protecting against potential downturns or unforeseen events like a pandemic, by reallocating capital into other asset classes like bonds, equities, or commodities 2m45s.
- Financial derivatives and futures contracts serve to provide liquidity, which increases confidence when engaging in large-scale expenditures 42s.
- Hedging against the price of compute allows entities to mitigate risk, providing psychological comfort even if the hedged event is not expected to occur 0s.
Financialization of AI Infrastructure
- The existence of futures markets for compute could enable new behaviors, such as hedge funds expressing views on the AI industry by shorting compute or NeoClouds selling futures against GPU fleets to finance data centers 25s.
- Financial derivatives function similarly to oil futures, where market participants express various points of view through the contract, thereby increasing the liquidity of the underlying physical exposure 42s.
- Dario Amodei previously noted that the risk of misaligning the timing of compute investment creates a binary outcome between significant profit and bankruptcy, leading to risk-averse scaling of physical infrastructure 1m5s.
- Liquid financial instruments could allow companies to build more infrastructure and dynamically adjust their exposure after the fact, following the historical pattern of Wall Street facilitating the capitalization and financialization of industrial investments 1m25s.
- Companies like CoreWeave and Nebius operate as landlords for GPUs, requiring significant capital expenditure and debt to build out data centers 1m45s.
- Futures contracts could potentially allow NeoClouds to lock in rental streams, thereby making infrastructure debt easier to underwrite in the event of declining rental prices 2m6s.
- Industry participants are working to make the financing of expensive capital assets, such as AI compute, more accessible and affordable through the development of financing platforms and futures markets 0s.
Market Transparency and Counterparty Risk
- The establishment of a futures market is expected to increase pricing transparency, allowing market participants to observe prices with greater certainty 0s.
- Some data centers currently mitigate risk by pre-selling compute capacity, ensuring that construction only proceeds if there are firm commitments from buyers at specific price points 0s.
- A significant concern in the market involves counterparty risk, specifically regarding whether companies like OpenAI or Anthropic will remain credible and capable of fulfilling their long-term compute purchase commitments by the time the capacity is delivered 0s.
- The AI compute industry is characterized as a "giant baby," representing a sector that is both enormous in scale and very young in development 0s.
Data Sourcing and Index Methodology
- Silicon Data does not focus on forecasting, as the company maintains that publishing an index of current market activity does not inherently improve one's ability to predict future market conditions 35s.
- Data for GPU rental contracts is sourced from two primary channels: publicly available published prices and data partners, including the sister exchange, Comput Exchange, which provides information on actual rental transactions 55s.
- To create an "apples to apples" comparison, the company uses machine learning and statistics to normalize rental contracts that vary by factors such as geolocation, memory specifications, CPU, and contract length 1m15s.
- The company consolidates these observations to produce benchmark indices, which are published daily on their website and made available on data platforms like Bloomberg Terminal 1m15s.
Trends in AI Hardware Demand and Utilization
- Conventional narratives regarding AI infrastructure often fluctuate based on stock market performance rather than underlying data trends 0s.
- Analyzing the AI compute market requires looking at a cross-section of various chip models, such as the H100, B200 (Blackwell), and A100, rather than focusing on a single chip's price 25s.
- The A100 chip, released in May 2020, experienced a decline in rental rates throughout the previous year, but rates have been steadily increasing and remaining strong since late last year 42s.
- The sustained demand for the A100, despite it being an older and less powerful chip, indicates a significant surge in inference demand 1m15s.
- The rise of agentic AI—where AI systems call other AI and utilize tools—has caused a substantial increase in overall AI demand 1m0s.
- As machine intelligence has become more powerful and cheaper per token, usage has increased, marking the current year as a period where AI has become widely utilized 1m5s.
- The high utilization of older hardware like the A100 challenges the conventional wisdom that data center assets are rapidly depreciating and losing value within two to three years 2m6s.
- Similar to how older smartphone models remain in high demand when there is a shortage of total devices, the continued use of older AI chips suggests that current production levels are insufficient to meet the total demand for compute 2m35s.
- Each new generation of AI chips, such as the transition from H100 to Blackwell and future iterations like Rubin, typically offers increased memory and power, with demand for these resources remaining additive 2m15s.
- Market participants may be underestimating the lifespan of AI hardware, as these assets might depreciate more slowly than previously budgeted 0s.
Infrastructure Constraints and Market Growth
- Equity valuations are influenced by the second derivative of growth, meaning that even if AI development continues to increase, a shift from exponential to super-linear growth can trigger negative market reactions 0s.
- Reports of data center construction pauses in regions like Texas and New York have raised questions about whether the AI infrastructure buildout is slowing down 35s.
- Concerns regarding AI infrastructure are multifaceted, involving debates over the return on investment, potential lack of demand, and local political pushback regarding electricity prices, inflation, water pollution, and noise 1m5s.
- Physical constraints, specifically a lack of available land and electrical power, serve as significant barriers to new data center projects 1m5s.
- Supply chain shortages for critical infrastructure components, such as gas turbines and high-voltage electrical transformers, are creating long-term delays for new developments 1m5s.
- The pause in Texas data center construction may be a reaction to the high volume of existing buildouts in the state rather than a broader rejection of AI infrastructure 2m6s.
Compute Pricing Dynamics and Market Indicators
- Data from July indicated that rental prices for B200, H100, and A100 chips were projected to be lower over a 36-month period, reflecting market expectations that compute costs would decrease 2m35s.
- Over the three to six months following the July data, the rental price curve for compute shifted upward across all contract lengths, which is attributed to rising demand for agentic AI 2m35s.
- Rising rental prices for GPU compute serve as a market indicator of increased demand, similar to how rising apartment rental costs reflect higher demand in the housing market 0s.
- Data tracked from March 2026 shows that GPU rental prices have experienced a steady, monotonic increase over the subsequent four months 18s.
Competitive Landscape of Cloud Providers
- NeoClouds are defined as companies that specialize exclusively in providing AI compute as a service 42s.
- Hyperscalers price their compute at a significant premium, typically two to three times the cost of NeoClouds, because they offer bundled services including enterprise-level trust, compliance, data security, and software analytics 52s.
- A specific comparison of H100 index capacity showed NeoCloud pricing at $2.70 per hour, while hyperscalers charged $7.28 per hour 1m15s.
Silicon Data Funding and Future Roadmap
- Silicon Data plans to launch compute futures for H100 and B200 rental indices in October, provided there are no regulatory issues 1m35s.
- Silicon Data intends to trade these compute futures on the CME 1m52s.
- Silicon Data has successfully raised a Series A funding round 1m58s.
- The startup secured $30 million in an initial Series A funding round led by Valor and the Trades AI fund, with Antonio Garcia and Gavin Baker serving as the primary leaders 0s.
- Additional participants in the funding round included F-prime, Fidelity, the CME venture fund, Samsung Next, and Tectonic 0s.
- The company aims to build an ecosystem that facilitates financial instruments for physical AI compute financing 0s.
- The long-term goal is to move beyond futures contracts that lack a physical underlying asset toward a system that allows for the physical delivery of compute 0s.
- The company is developing indices and a tool called Silicon Mark, which evaluates individual GPUs at specific data centers to provide a grade for those assets 0s.
- This grading system is intended to allow users to take instances of on-demand or reserved AI compute at the expiration of futures contracts 0s.
- Increased enterprise adoption of AI, particularly the growth of inference, is expected to make compute more interchangeable and fungible across different platforms and providers 0s.
- The company plans to enable users to identify the cheapest compute available at the time of a futures contract's expiration 0s.
