In 1983, the New York Mercantile Exchange listed the first crude oil futures contract. At the time, the idea was met with skepticism bordering on derision. Oil was priced through long-term bilateral contracts between producers and refiners, managed by a cartel, and traded informally through a network of brokers who communicated by telephone. The notion that you could standardize this opaque, relationship-driven market into exchange-traded contracts through transparent pricing, central clearing, and daily mark-to-market, was dismissed by industry incumbents as naive at best and destabilizing at worst. Within five years, NYMEX crude oil futures had become the global benchmark for energy pricing, fundamentally reshaping how oil was financed, produced, traded, and consumed. Within a decade, the derivatives market for oil dwarfed the physical market by a factor of ten.
We are standing at the same inflection point for AI compute. But with an important complication that the current market-building efforts have not yet fully confronted.
In the first two installments of this series, we established that token pricing has collapsed 1,000x in three years while total enterprise AI spending is continuing to surge; that AI inference is rapidly displacing (at least in part) human labor as a primary operating expense, shifting corporate cost structures from predictable to volatile; and that no financial instrument exists to hedge this exposure. The conclusion was stark: AI compute is the largest unhedgeable variable cost in modern corporate finance.
Historical precedents exists, but perhaps nothing matches the scale of what we’re seeing, and more importantly expect to see in AI in the years ahead.
This final installment answers the question that logically follows: what needs to be built?
Not in the abstract, but concretely. What are the layers of financial infrastructure required to transform AI token costs from an unmanageable expense into a hedgeable one? Where are the analogies to existing commodity markets instructive, and where do they break down? And, critically, why is the infrastructure currently under construction necessary but insufficient to solve the corporate CFO's actual problem?
The answer requires distinguishing between two separate but related asset classes that most commentary conflates: compute and intelligence. GPU hours and AI tokens are not the same commodity. They are linked the way crude oil is linked to jet fuel, but the relationship between them is variable, non-linear, and driven by forces that move independently.
This distinction is not academic. It determines whether the hedging infrastructure being built will actually reduce the risk that corporate buyers face, or merely create a new form of basis risk that replaces one problem with another.
The two-layer problem: compute is not intelligence
The emerging financial infrastructure for AI, the indices being published, the futures contracts being designed, the exchange partnerships being announced, is focused almost entirely on the compute layer: the hourly cost of renting GPU hardware. Ornn Data LLC's Compute Price Index (OCPI) tracks live spot prices for GPU compute across hardware types including H100, H200, B200, and RTX 5090.¹ Silicon Data's SDH100RT index, backed by DRW and Jump Trading, tracks the hourly rental cost of H100s across global providers.² Architect Financial Technologies has announced perpetual futures contracts linked to GPU and DRAM pricing.³ These are valuable, necessary developments. They bring transparency to a market that has operated in near-total opacity. They are necessary for the hyperscalers and others investing directly in AI infrastructure.
But they solve the wrong problem for the corporate CFO.
Here is why. The Fortune 500 CFO whose AI budget this series has been tracking does not buy GPU hours. She buys tokens — units of intelligence produced by a model, metered per million, billed by her AI provider.
The relationship between the GPU spot price and the token price she pays is mediated by at least five independent variables, each of which moves on its own timeline and for its own reasons.
Model architecture efficiency. DeepSeek V3.2 has 671 billion total parameters but activates only 37 billion per token through its mixture-of-experts design.⁴ A dense model of equivalent capability activates all parameters on every inference pass. Same GPU, same hourly rental rate, radically different token throughput, and therefore radically different token cost.
When a lab ships a more efficient architecture, token prices fall without any change in the underlying GPU price. A compute index does not capture this, and using compute-based derivatives would introduce too much basis risk to create the budget predictability CFO’s are after.
Inference optimization. Quantization, speculative decoding, KV-cache optimization, PagedAttention, and continuous batching improvements all increase the number of tokens produced per GPU-hour. These techniques improve continuously and independently of hardware upgrade cycles. Roblox achieves a 90% cost reduction through 2-bit quantization with speculative decoding on the same hardware, at the same GPU rental rate.⁵
The compute price did not move. The effective token price collapsed. A compute index registers nothing.
Provider pricing strategy and subsidization. OpenAI is projecting $14 billion in losses in 2026.⁶ Google prices Gemini Flash below marginal cost to drive ecosystem adoption. DeepSeek prices at a fraction of Western alternatives for a combination of geopolitical, strategic, and cost-structure reasons.⁷ Token prices reflect corporate strategy, competitive dynamics, and investor pressure, not just cost-plus markup on compute inputs.
When OpenAI eventually raises prices 30% to improve its path to profitability, the GPU spot market will not register that event at all. But every enterprise buyer's token cost spikes immediately.
Capability tiers. A million output tokens from Claude Haiku 4.5 costs $5. The same million from Claude Opus 4.6 costs $25.⁸ The compute consumed is different, but so is the intelligence delivered, and the intelligence quality is what determines the token's economic value to the buyer. A GPU index treats compute as homogeneous raw material.
Token-level pricing inherently encodes the quality dimension that matters to the end consumer.
Competitive dynamics across the stack. Anthropic could migrate from NVIDIA GPUs to Google TPUs. Meta could shift Llama inference to custom silicon. An AI lab could vertically integrate into hardware, fundamentally restructuring its cost basis.⁹ In each scenario, the GPU spot price is unchanged or moves for unrelated reasons, while the token price shifts substantially based on the provider's new cost structure, competitive positioning, or capacity decisions.
The abstraction layer between hardware and tokens means that competitive moves at the infrastructure level propagate to token pricing in non-linear, difficult-to-predict ways.
The cumulative effect of these five variables is that the correlation between GPU compute prices and AI token prices, while positive, is loose, unstable, and subject to structural breaks.
A corporate CFO who hedges his token cost exposure using GPU compute futures is carrying enormous basis risk: the risk that the hedge and the underlying exposure diverge, potentially leaving him worse off than if he had not hedged at all.
The crude oil analogy — and why it matters
The cleanest analogy is one that every corporate treasurer will immediately recognize.
Airlines are exposed to jet fuel prices. Jet fuel is refined from crude oil. Crude oil futures are the most liquid energy derivatives market in the world. Yet airlines that hedge their fuel exposure exclusively with crude oil futures carry significant crack spread risk — the variable refining margin between crude and the specific refined product they actually consume. When refining capacity tightens, or when specific refined product demand spikes, jet fuel prices can diverge sharply from crude. An airline that sold crude oil puts to hedge against rising fuel costs can find itself with a hedge that moved in the right direction, but by the wrong magnitude, or, in extreme cases, in the wrong direction entirely.
This is why the jet fuel swap market exists. Airlines that can hedge directly with jet fuel derivatives, like swaps and options that reference the actual commodity they consume, not its upstream input, have materially lower residual risk.
The jet fuel swap market was not the first energy derivative. It emerged after crude futures were well established, building on the price discovery infrastructure that crude provided, but adding a layer of specificity that crude alone could not deliver.
It created the tool that the Airlines actually needed, allowed low-cost carriers like Ryanair, EasyJet and (originally) Southwest Airlines to exist. US carriers have opted to forgo most hedging, instead passing on rising costs to consumers, assuming that demand is sufficiently elastic to absorb higher prices.
If you’ve booked an airline ticket recently, you know how that feels as the end consumer.
Pricing dynamics in AI is no different.
AI tokens are the jet fuel. GPU compute is the crude oil. The compute derivatives market currently being built, the GPU spot indices, the hardware futures, the DRAM contracts, is the crude oil layer. It is necessary. It provides the foundational price discovery that the entire derivatives stack requires. Data center operators, GPU cloud providers, lenders against GPU collateral, and hardware manufacturers will use these instruments directly, because their economic exposure genuinely lives at the hardware level.
But the corporate CFO's exposure lives at the token level. Their costs are denominated in tokens per million, billed by model tier, subject to provider pricing decisions and architectural efficiency gains that have nothing to do with GPU spot rates. They need the equivalent of a jet fuel swap, a derivative instrument that references AI token pricing directly, at the capability tier she actually consumes, with settlement against an index that captures the full cost stack from silicon to intelligence.
The compute layer is being built. The intelligence layer is not.
Until it is, the corporate hedging problem described in Part Two of this series remains fundamentally unsolved.
The five layers of infrastructure that must be built
Transforming AI token costs from an unhedgeable expense into a manageable risk requires five distinct layers of infrastructure, each dependent on the one below it. The first two layers apply to both compute and intelligence. The remaining three are where the intelligence layer introduces genuinely novel challenges.
Layer One: the compute benchmark
Every derivatives market begins with a price that both sides of a trade can observe, verify, and trust. For the compute layer, this foundation is now being laid. The OCPI tracks GPU spot prices built on executed transaction data — not surveys, not rate cards, not provider list prices.¹⁰ Its Bloomberg Terminal availability signals institutional readiness.¹¹ The design choice to build on cleared transaction prices rather than self-reported data mirrors the methodology of the most successful commodity benchmarks: Platts for oil, ERCOT for Texas electricity, EEX for European power. And it avoids manipulation.
Ornn's adoption of Asian-style settlement, averaging the volume-weighted price over the contract period rather than settling on a single expiration date, reflects a sophisticated understanding of how compute is consumed.¹² GPU compute, like electricity, is a flow commodity. It cannot be stored. What market participants care about is not the price on a single day but the average price paid over the period of use. This structure aligns the financial payoff with the economic reality of compute consumption.
The compute benchmark is real, it is live, and it is sufficient to support the first generation of GPU-focused derivatives. This layer is largely solved.
Layer Two: the intelligence benchmark ← the unsolved problem
The token layer requires its own benchmark: an index that tracks the cost of AI intelligence at standardized capability tiers, across providers, normalized for quality.
This is, in my assessment, the single most important unsolved problem in AI financial infrastructure. And it is substantially harder than building a GPU spot index.
The challenge is multidimensional. A token is not a fungible unit. One million output tokens from a frontier reasoning model delivering PhD-level analysis is a fundamentally different product than one million tokens from a lightweight model handling customer service triage. A useful intelligence benchmark must capture this quality dimension while remaining standardized enough to serve as a derivatives settlement reference. It must aggregate across providers like OpenAI, Anthropic, Google and open-source alternatives, without being dominated by any single provider's pricing decisions. It must account for subsidization, so that the index reflects sustainable market pricing rather than temporary loss-leader strategies. And it must be constructed from verifiable data, not self-reported list prices that providers can change unilaterally without warning.
The design constraints point toward a tiered index architecture, multiple indices tracking different capability classes, linked by basis relationships that the market discovers through trading. The analogy to crude oil grades (WTI, Brent, Dubai, Urals) or electricity delivery points (PJM Western Hub, ERCOT North, SP15) is instructive. Each grade or node has its own price, reflecting local supply-demand conditions, and the relationships between them are actively traded as basis spreads.
For AI intelligence, the "grades" would be capability tiers defined by performance benchmarks rather than brand names. The methodology for defining these tiers, selecting the reference benchmarks, normalizing across providers, and adjusting for subsidization represents the core intellectual property of this emerging market. I will note only that the design space is well understood by those working on the problem, that the solutions draw more heavily from electricity market design than from traditional commodity benchmarks, and that the companies that solve this problem will own the most strategically valuable piece of the entire AI financial infrastructure stack.
Any such index should be constructed in alignment with the IOSCO Principles for Financial Benchmarks, which provide the governance standard (independent oversight, transparent methodology, third-party audits) required for institutional adoption.¹³
Until the intelligence benchmark exists, all the downstream infrastructure (forward curves, swaps, options, structured products) for the token layer cannot be built on a sound foundation. This is where the market-building bottleneck sits today.
Layer Three: the forward curve — pricing the future
In mature commodity markets, the forward curve is arguably more important than the spot price. It aggregates expectations about future supply and demand, incorporates information about storage costs, seasonality, and risk premia, and provides the price signals that guide investment and hedging decisions. An oil producer deciding whether to drill a new well looks at the five-year forward curve, not today's spot price. A utility deciding whether to build a gas plant examines the ten-year power forward curve. The price expected in the future is far more relevant than the price today.
AI compute, at both the hardware and intelligence layers, has no forward curve.
None.
A company evaluating a $200 million AI infrastructure investment has no market-derived signal about what compute or tokens will cost in twelve months. Every decision is made on the basis of internal forecasts, vendor conversations, and educated guesswork. This is exactly the way oil companies operated before 1983.
The absence of a forward curve has concrete, measurable consequences. CoreWeave holds approximately $18.8 billion in GPU-collateralized debt. Because there is no forward market for GPU compute, lenders cannot hedge their collateral exposure.¹⁴ The result is that GPU financing is more expensive, more opaque, and more concentrated than it needs to be. One analyst described CoreWeave as looking less like a modern cloud company and more like a 1990s independent power producer with leveraged infrastructure financed through bilateral relationships, with risk concentrated rather than distributed.¹⁵
Building forward curves requires reliable spot benchmarks and sufficient trading activity to establish prices at multiple tenors. For the compute layer, this process can begin relatively soon, anchored to the OCPI and similar indices. For the intelligence layer, the token-level benchmark described above must come first.
The natural starting points are one-month, three-month, and twelve-month forwards, the tenors most relevant to enterprise budgeting cycles. Initially established through bilateral OTC trades, these will eventually migrate to exchange-traded futures that create continuous curves visible to all participants.
The critical insight for CFOs: the moment a twelve-month forward price for AI inference becomes observable, corporate budgeting for AI changes fundamentally. Instead of guessing what tokens will cost next year, a CFO can observe a market-derived price, compare it to internal forecasts, and make hedging decisions accordingly.
The forward curve firmly migrates AI cost planning from a technology procurement exercise into a treasury risk management function.
Layer Four: the contract architecture
The design of hedging instruments must reflect the two-layer structure of the market. Several categories of instruments are needed, each serving a different participant and purpose.
Compute-layer instruments like GPU futures, hardware swaps, DRAM forwards, serve data center operators, cloud providers, lenders, and hardware manufacturers. These are the instruments currently being designed by Ornn and Architect.¹⁶ Cash-settled against GPU spot indices, with Asian-style averaging, they will allow participants whose economic exposure genuinely lives at the hardware level to manage that exposure with precision. This market is already being developed, because the benchmark infrastructure exists and the participant base is identifiable.
Intelligence-layer instruments like token forwards, inference swaps, capability-tiered options, serve enterprise AI consumers, AI-native SaaS companies, and the CFOs whose budgets are increasingly denominated in, and dependent on AI tokens. These instruments must settle against the intelligence benchmark (Layer Two) and must be designed with sufficient specificity to meaningfully reduce the hedger's actual cost variance, while remaining standardized enough to support liquidity. The design challenge is that AI inference exposure is multidimensional: a buyer's cost depends on the model tier, the prompt complexity, the output length, the batch versus real-time mix, and the geographic routing. A well-designed instrument abstracts away enough of this complexity to be tradeable while retaining enough granularity to be useful.
Basis instruments like compute-to-intelligence spreads serve participants with exposure to the relationship between the two layers. This is the "crack spread" of the AI market. AI labs are the natural participants: they are long compute (they buy GPU hours) and short intelligence (they sell tokens). Their margin is literally the basis between the two layers. A basis swap allows a lab to lock in its processing margin regardless of whether GPU prices rise or token prices fall. It also allows institutional market participants to take views on the efficiency trajectory of inference, essentially betting on whether the AI industry will produce more or fewer tokens per GPU-hour in the future.
Several design principles apply across all instrument categories:
Cash settlement is essential. Physical delivery of "compute" or "intelligence" is operationally impractical. All contracts must settle financially against benchmark indices. This is how electricity futures, weather derivatives, and freight derivatives all operate.
Asian-style averaging is superior to terminal settlement. Corporate buyers' exposure is to the average cost over a budget period, not the spot price on a single date. Averaging reduces basis risk and aligns the financial instrument with economic reality.¹⁷
Standardization must balance liquidity and hedge effectiveness. Contracts must be standardized enough to support secondary market trading and central clearing, yet specific enough, through capability tiering and tenor structure, to meaningfully reduce hedgers' actual cost variance.
Layer Five: the institutional framework
No institutional derivatives market functions without regulatory clarity, clearing infrastructure, and accounting standards that allow corporate hedgers to use the instruments cleanly.
On the regulatory side, the pathway is clearer than one might expect. Cash-settled futures and swaps on price indices fall squarely within existing CFTC authority under the Commodity Exchange Act. No new legislation is required.¹⁸ A designated contract market (DCM) would self-certify new contracts under CFTC Regulation 40.2(a). A derivatives clearing organization (DCO) would provide central clearing and margin management. The current CFTC leadership under Chairman Michael Selig has launched an Innovation Task Force dedicated to advancing clear rules for novel products and technologies. The regulatory environment has not been this favorable for new commodity market development in a generation.¹⁹
The accounting treatment is equally critical. Under ASC 815 (U.S. GAAP) and IFRS 9 (international standards), derivatives used for hedging can receive special treatment that reduces P&L volatility, but only if the hedging relationship meets specific criteria for designation, documentation, and effectiveness testing.²⁰ For a token derivative to qualify for hedge accounting, the hedger must demonstrate that the hedged item (AI token cost exposure) and the hedging instrument (token futures or swaps) are highly correlated. This is where the intelligence benchmark design matters directly: a well-constructed token-level index that tracks the actual cost experience of enterprise AI buyers will enable hedge accounting qualification. A poorly constructed index, or, worse, a GPU-level index used to hedge a token-level exposure, will fail effectiveness testing, forcing hedgers to mark positions through the income statement and potentially creating more P&L volatility than the underlying exposure.
The practical implication is that index construction, contract design, and accounting treatment are not separate workstreams. They must be designed together, with the accounting outcome informing the index methodology from the outset.
The carbon credit market learned this lesson painfully. Early instruments designed without sufficient attention to accounting treatment left many corporate buyers with hedges that created more earnings volatility than the exposures they were hedging.
The market participant map — who sits on each side
A derivatives market requires counterparties with opposing economic interests. One of the most encouraging features of the AI market is that natural two-sided demand exists at both layers, and a rich ecosystem of intermediaries connects them.
Compute-layer participants.
- Buy side: AI labs purchasing GPU capacity, cloud providers building out data centers, enterprises running on-premises inference.
- Sell side: GPU cloud operators preselling capacity, hardware manufacturers hedging production risk, lenders managing GPU collateral depreciation.
- Between them: energy trading firms and quantitative market makers (DRW's CEO has publicly stated that compute is becoming the world's largest commodity²¹), commodity trading desks at major banks, and the specialized exchanges being built by firms like Architect.
Intelligence-layer participants.
- Buy side: Fortune 500 enterprise AI consumers seeking cost predictability, AI-native SaaS companies managing variable COGS, financial institutions modeling AI cost exposure across portfolio companies.
- Sell side: AI labs locking in inference revenue, cloud providers offering reserved capacity at guaranteed rates, model hosting platforms managing utilization.
- Between them: the same intermediary ecosystem, augmented by enterprise technology advisory firms and the corporate treasury consultants who currently advise on FX and commodity hedging programs. VC, PE and similar financing firms that have a vested interest in increasing AI adoption across their portfolio of owned companies.
Basis participants. AI labs themselves are the natural basis traders. They are long compute, short intelligence, their margin defined by the spread. Quantitative hedge funds seeking uncorrelated exposure will trade the basis as a macro view on AI efficiency. And corporate hedgers running multi-layer programs will use basis instruments to fine-tune their risk profiles.
The Architect-Ornn partnership, with Architect bringing FINRA-registered broker-dealer status and a perpetual futures exchange, and Ornn bringing the compute benchmark, represents the first concrete market structure for the compute layer.²² Ornn CEO Kush Bavaria told The Block that the company has already executed its first compute swap, with more in the pipeline.²³
The intelligence layer awaits its own market builders. The opportunity is open.
The timeline: when does this become real?
Historical precedents suggest this market will develop faster than most expect.
Weather derivatives went from the first bilateral OTC trade in September 1997 to CME-listed exchange-traded products in September 1999 — twenty-four months.²⁴ The EU carbon market went from legislative mandate in 2003 to the first EUA futures trading in 2005 — two years.²⁵ In each case, the catalyst was an institutional recognition that the underlying risk had become too large to manage bilaterally.
AI compute is at that moment now.
For the compute layer: the OCPI is live on Bloomberg. Ornn has executed its first compute swaps. Architect has announced exchange-traded perpetual futures on GPU prices, pending regulatory approval. Kalshi has listed prediction contracts on GPU compute pricing.²⁶ I expect OTC bilateral compute hedging within months, and exchange-listed compute futures within twelve to eighteen months.
For the intelligence layer, the timeline is longer because the benchmark infrastructure must be built first. Token-level indices with sufficient provider coverage, capability normalization, and institutional credibility are a twelve- to eighteen-month development effort. Once established, OTC bilateral token swaps could follow within six months, with exchange-listed products eighteen to twenty-four months after that.
Between the two layers, basis instruments will emerge organically as participants trade the compute-to-intelligence relationship. This market does not require its own dedicated infrastructure. It piggybacks on the two benchmarks and the ecosystem of derivatives around each.
Within three to five years, the full ecosystem should be operational: compute futures and options, token forwards and swaps, basis spreads, structured products, and a forward curve at multiple tenors that transforms AI cost planning from guesswork into market-informed budgeting.
The market-building talent is mobilized. The people assembling this infrastructure are not AI researchers. They’re quantitative traders, derivatives market professionals, and commodity exchange architects who understand that the hard problem is not defining what compute or intelligence is worth, but building the plumbing that enables price discovery.²⁷
What CFOs should do now — before the instruments arrive
The companies that benefited most from the maturation of electricity derivatives were not the ones that waited for the market to arrive. They were the ones that built internal readiness while the market was forming.
Establish granular cost attribution. You cannot hedge an exposure you cannot measure. Most enterprise AI cost management today operates at the level of a monthly cloud bill. Effective hedging requires attribution by business unit, application, model provider, capability tier, and use case. The FinOps Foundation's AI working group has published frameworks for generative AI cost tracking.²⁸ The goal is to know your token exposure with the same precision you know your FX exposure.
Decompose your exposure into compute and intelligence components. Understand which portion of your AI cost is driven by raw hardware economics (and would therefore be hedgeable with compute-layer instruments) versus which portion is driven by model-level factors — provider pricing, architectural efficiency, subsidization — that can only be hedged at the token layer. This decomposition will determine which instruments are useful to you when they arrive, and which leave residual basis risk.
Model AI costs as a treasury risk, not a technology budget. Immediately move AI inference cost management from the CIO's office to the CFO's office, or at minimum, establish joint governance. The moment AI spend exceeds the threshold at which FX or commodity exposure would trigger a hedging policy, it should receive the same treatment. For many enterprises, that threshold has already been crossed.
Build scenario models that capture tail risk. Model the impact on your operating margin of a 50% increase in token costs sustained over two quarters. This is likely the kind of event that a subsidy withdrawal, supply shock, or energy cost spike could produce. If the margin impact exceeds your tolerance for unhedged exposure in any other cost category, you have your answer about the urgency of hedging readiness.
Engage with the emerging ecosystem. The compute derivatives market is being built right now, and the participants who engage early will influence contract design, index methodology, and market structure. This is the lesson of every new commodity market. Early adopters shape the instruments to fit their needs. Late adopters inherit instruments designed for someone else.
Begin bilateral discussions with your AI providers about forward pricing. Even without formal derivatives, enterprise agreements with commit-and-discount structures, volume-tiered pricing, and multi-year rate locks represent proto-derivatives that will evolve into standardized instruments. Every bilateral forward pricing agreement executed today generates transaction data that tomorrow's indices will be built on.
Conclusion: the two markets that must be built
The AI inference market will reach $254 billion by 2030.²⁹ Enterprise AI spending is projected at $2.5 trillion in 2026 alone.³⁰ If compute derivatives develop multipliers comparable to oil, where derivatives trade at 10–15x the value of the physical market, the potential market could reach several trillion dollars in notional value within the decade.
But this market will not be a single, monolithic structure. It will be two interconnected markets, linked by basis instruments, serving different participants with different risk profiles.
The compute market: GPU futures, hardware swaps, and DRAM derivatives will serve infrastructure participants like data center operators, cloud providers, lenders, and hardware manufacturers. This market is being built today. Its benchmarks exist. Its first contracts are imminent. It will provide the foundational price discovery layer for the entire ecosystem.
The intelligence market: AI token forwards, inference swaps,and capability-tiered options will serve the participants whose exposure this series has been tracking: the Fortune 500 CFOs managing the fastest-growing operating expense on their income statements. This market requires a fundamentally different benchmark, one that captures the full cost stack from silicon to intelligence, accounts for the efficiency gains and pricing strategies that decouple token costs from hardware costs, and is constructed with sufficient rigor to support hedge accounting under ASC 815 and IFRS 9.
Throughout this series, we have traced a single narrative arc. In Part One, we mapped the terrain of token economics: a market defined by unprecedented deflation, structural fragility, and subsidy dependence. In Part Two, we confronted the consequences: an unhedgeable risk that is reshaping corporate cost structures, creating a new category of balance sheet exposure, and undermining the margin predictability that investors demand. Here, in Part Three, we have laid out the blueprint: the two-layer market architecture that must be built to transform this risk from unmanageable to hedgeable, and identified the critical gap between the compute infrastructure now under construction and the intelligence infrastructure that corporate buyers actually need.
The compute layer is being built. The intelligence layer is the opportunity.
We’ve closed each of the previous two pieces with the analogy that the AI token is the new oil barrel. The forward curve is being drawn. And the most important question for every participant in the AI economy (buyer, seller, lender or intermediary) is not whether this market will emerge, but who will build the intelligence layer that sits above the hardware and below the enterprise. That is where the value will concentrate. That is where the hedging problem is actually solved. And that is where the next great financial market will be made.
Sources
- Ornn, ornnai.com; Data Center Dynamics, "Kalshi users able to gamble on Nvidia GPU compute prices, based on Ornn's compute derivatives platform," March 2026.
- Dave Friedman, "Compute is the Commodity No One Knows How to Price," Substack, February 4, 2026; Silicon Data SDH100RT index on Bloomberg, May 2025.
- PR Newswire, "Architect Financial Technologies Partners with Compute Index Provider Ornn to Launch Exchange-Traded Futures on GPU and RAM Prices," January 21, 2026.
- Introl Blog, "DeepSeek-V3.2 Matches GPT-5 at 10x Lower Cost," 2026; Fortune, "China's DeepSeek just dropped a new GPT-5 rival," August 21, 2025.
- Introl Blog, "Cost Per Token Analysis: Optimizing GPU Infrastructure," 2025.
- eMarketer, "OpenAI's forecast $143 billion cash outflow raises stakes," 2025; Yahoo Finance, "OpenAI's own forecast predicts $14 billion loss in 2026," 2025.
- Fortune, "China's DeepSeek just dropped a new GPT-5 rival — optimized for Chinese chips, priced to undercut OpenAI," August 21, 2025.
- MetaCTO, "Claude API Pricing 2026: Full Anthropic Cost Breakdown," 2026; Awesome Agents, "LLM API Pricing Comparison — March 2026."
- AI News Hub, "Nvidia to Google TPU Migration 2025: The $6.32B Inference Cost Crisis," 2025; HPCwire, "Google Announces Sixth-generation AI Chip, a TPU Called Trillium," May 17, 2024.
- The Innermost Loop, "The First Tradable Compute Price Index," April 2026.
- The Innermost Loop, ibid. ("Search 'ORNNH100' on the Bloomberg Terminal").
- Ornn Research, "Compute Futures," 2026. Available at ornnai.com/research/compute-futures.
- IOSCO, "Principles for Financial Benchmarks," Final Report, July 2013.
- Dave Friedman, "Compute is the Commodity No One Knows How to Price," Substack, February 4, 2026; Sacra, "CoreWeave revenue, valuation & funding," 2025.
- Dave Friedman, ibid.
- PR Newswire, ibid.; The Block, "Former FTX US president Brett Harrison's Architect expands crypto-style perpetual futures into AI compute markets," January 21, 2026.
- Ornn Research, ibid. ("Asian-style settlement aligns the financial payoff of the contract with this economic reality").
- Commodity Exchange Act, 7 U.S.C. § 1 et seq.; CFTC Regulation 40.2(a).
- CFTC, "Chairman Michael S. Selig Launches Innovation Task Force," March 2026; Gibson Dunn, "Derivatives, Legislative and Regulatory Weekly Update," January 9, 2026.
- FASB ASC 815 (Derivatives and Hedging); IFRS 9 (Financial Instruments), Section 6.
- DRW CEO remarks, cited in Dave Friedman, ibid., and multiple publications, 2025.
- PR Newswire, ibid.; Morningstar, "Architect Financial Technologies Partners with Compute Index Provider Ornn," January 21, 2026.
- The Block, ibid. ("AI labs and GPU-heavy companies need price certainty... We've already executed our first compute swap with more in pipeline").
- Barrieu, P. and Scaillet, O., "A primer on weather derivatives," 2010; Carbon Credits, "Weathering the Storm: The Rise of $25B Weather Derivatives Market," 2024.
- European Commission, EU ETS Phase 1 documentation, 2005; Wikipedia, "European Union Emissions Trading System."
- Data Center Dynamics, ibid.; Kalshi/Ornn partnership announcement, March 2026.
- Pulse 2.0, "Ornn: $5.7 Million Seed Funding Raised For Launching Compute Futures Exchange," October 2025; PitchBook, "Ornn Compute Exchange 2026 Company Profile."
- FinOps Foundation, "How to Forecast AI Services Costs in Cloud," 2025; "How to Build a Generative AI Cost and Usage Tracker," 2025.
- Market estimates, multiple sources; compound annual growth rate ~17.5% from $97.24B (2024). See Medium (Ahmed's Tech Brief), "AI Inflation: How Compute Costs Are Reshaping Tech's Margins," November 25, 2025.
- Process Excellence Network, "Global AI spending will total $2.5 trillion in 2026, says Gartner," 2026.