The story of modern commodity derivatives really begins in the early 1980s, in the aftermath of the most chaotic decade the energy market had ever seen. The 1973 Arab oil embargo tripled crude prices almost overnight. The 1979 Iranian Revolution sent them parabolic again, pushing real prices to their post-war peak. For most of that decade, pricing power sat in the hands of OPEC and the so-called Seven Sisters, the seven Western oil majors who effectively ran the global petroleum trade through bilateral contracts, posted prices, and handshake deals.
This was, to put it mildly, inefficient. Oil was the single most economically important commodity on the planet, and yet its pricing mechanism was about as sophisticated as a flea market.
As a sidenote, very interesting tales abound of the early pioneers that made a killing in these opaque bazaars, a phenomenon that still carries on to some extent today when liquidity dries up for one reason or another.
In March 1983, the New York Mercantile Exchange, then primarily known for trading potato futures, launched the West Texas Intermediate (WTI) crude oil futures contract. The oil industry sat on the sidelines initially; the Seven Sisters didn’t want transparent pricing because opacity was their competitive advantage. OPEC tried to boycott the contract entirely.
By the late 1980s, trading volumes were north of two million contracts per month. The futures contract survived its first real test in April 1986, when Saudi Oil Minister Sheikh Ahmed Yamani’s netback pricing strategy crushed the crude market and prices plummeted from $34 to $9 in two days. The contract held, providing price discovery when the physical market was in freefall, illustrating exactly the desired function of derivatives in the market.
Futures didn’t just enable hedging, they democratised the price of oil. CFOs could lock in crude costs six months forward, just as airlines could budget fuel expenses with confidence and a hedge fund could express a view on Middle Eastern geopolitics without buying a single barrel. The transparent, composable, permissionless nature of a futures market turned oil from an opaque commodity into one of the most liquid financial instruments on the planet.
The key lesson? Every commodity that matters eventually gets financialised. It happened to Oil. It would then happen for electricity, natural gas and gold. It even happened for orange juice.
The pattern is always the same, or at least rhymes – a resource becomes economically critical, so an already opaque price now becomes volatile and opaque, which leads participants to demand hedging tools to handle this volatility and leads to someone building the exchange infrastructure to capture the margins that were previously captured under the veils of opacity by vendors, and only after all of these market mechanisms take root does a liquid derivatives market emerge that transforms that entire industry.
Compute in 2026 bears many resemblances to oil in 1982. An economically transformative resource, with opaque pricing, fragmented markets and no hedging instruments. All of which creates a spiral towards massive volumes in unmanaged exposure.
The ramifications of the financialisation of oil markets of the 70’s and 80’s have dominated the global political economy and much of global markets ever since, a facet which is rearing its ugly head as I write this. One of the key ideas to emerge from this regime was the Petrodollar.
As the exporting nations that controlled the dominant share of supply in the world’s most valuable commodity accumulated vast pools of USD from the sale of said commodity, institutions were built up (Sovereign Wealth Funds, Treasuries etc.) that garnered substantial global influence.
The Petrodollar reshaped global finance because whoever controls the indispensable commodity controls the capital flows. In an age where intelligence seems inevitable to surpass oil as the most important commodity in the world, the neurodollar – the capital pool accumulating wherever compute is produced, distributed and settled – will do the same.
What I aim to outline below is what the actual structure of this broad new economy of compute markets looks like, who is already helping to facilitate it, how they are approaching it and how it may evolve as the market invariably changes (a property that has been far less of a factor for the resources we pull out of the ground).
The Most Important Market in the World
In 2026, the five largest cloud and AI infrastructure providers (Microsoft, Alphabet, Amazon, Meta, and Oracle) have collectively committed to spending between $660 billion and $690 billion on capital expenditure.
That’s roughly double the $381 billion they spent in 2025, which itself was already a record. The vast majority of this spend is going directly to semiconductors, servers, and data centre infrastructure. Microsoft alone is tracking toward $120 billion or more in its fiscal 2026. Amazon has announced $200 billion. Alphabet is guiding $175–185 billion. Meta is planning $115–135 billion. These aren’t some absurd handpicked projections that I have come across but instead actual guidance from company earnings calls.
If we add in the neoclouds, CoreWeave, which IPO’d in March 2025, reported contracted customer backlog of $55.6 billion in Q325. Once you add in a range of other players such as Lambda Labs, Nebius, Crusoe and a suite of emerging sovereign AI plays, the total capital flowing into AI compute infrastructure starts to approach a trillion dollars annually.
To put this in perspective: the entire global defence industry generates about $600 billion in annual revenue. The pharmaceutical industry $1.5 trillion. The annual GDP of countries like the Netherlands, Switzerland, or Saudi Arabia sits in the $800 billion–$1 trillion range. We are witnessing capital deployment on the scale of national economies being directed at a single class of infrastructure.
The reasonable question given all this information – is this a bubble?
Maybe.
There’s a version of this story where all these data centre buildouts become analogous to the Chinese housing crisis that began in 2020 – monuments to speculative excess built with cheap capital, on the assumption that demand will catch up to supply because it has to.
This Big Short-style hypothetical (or 大卖空, if you prefer) doesn’t seem entirely unreasonable or overstated if the worst case scenario (or a range of middling-to-adverse scenarios) prevails. Morgan Stanley projects that several hyperscalers will see negative free cash flow in 2026; Pivotal Research forecasts Alphabet’s free cash flow will plummet nearly 90% this year. These companies are burning through cash on the assumption that AI revenue will eventually justify the spend. Within the past week alone OpenAI caused some eyebrows to raise when it was revealed that there were offering 17.5% preferred returns to creditors, placing it in the range of many venture-debt funded early-stage startups.
But here’s the thing: even if AI capex is partially a bubble, it’s a bubble in the infrastructure layer of what will almost certainly be the dominant technology of the next several decades. The internet bubble popped, but the fibre optic cables stayed in the ground, and they ended up being wildly valuable. The railway bubble of the 1840s bankrupted half of Britain’s investors, but the railways transformed the global economy. Compute infrastructure, once built, doesn’t disappear. It gets repurposed, repriced, and eventually absorbed. The market for compute is real even if the current pace of investment overshoots.
Which brings us to the really strange part.
There’s no meaningful way to trade, hedge or borrow against compute….yet
Going back to my favourite point from above, we have futures markets for orange juice. You can borrow against your exposure to lean hogs (hell yeah brother). You can trade derivatives on the volatility of volatility. There are financial instruments for almost any economically significant commodity you can name—except the one that’s reshaping the entire global economy.
Try taking phsyical delivery of these bad boys.
Today, if you want exposure to compute, your options are roughly:
1. Buy NVIDIA stock (or other derivative of your choice – photonics are so hot right now). This is a proxy for GPU demand, but not expressing a financial opinion on the value of the underlying compute itself; or
2. Buy shares in a hyperscaler like Microsoft, Meta or Amazon. This buries your exposure under a hundred other business lines, including but not limited to RayBan glasses, third-party logistics and Office 365 (not expressly bad options but also not compute); or
3. Negotiate a private, bespoke contract with a cloud provider. If you are looking for automated fixes to manage your exposure in real-time, good luck dealing with lawyers back-and-forthing OTC contracts.
There’s no way to buy (with substantial size) a futures contract on the price of a GPU-hour nor any direct way to hedge compute budgets against a price spike. Indexes are in place, but they are not explicitly tradeable as of yet. This will obviously change in rapid course, which will be covered in further detail below when I examine the competitive landscape.
The cost structure of an AI startup depends almost entirely on GPU access, a fact that amplifies the closer one gets to the foundation layer. Lab needs thousands of GPU-hours per day to run inference workloads.
A 10% counter-seasonal price spike occurred in December 2025, driven in part by localised supply constraints and demand from agentic AI workloads. If the budget in question was $50mm annually, that spike alone cost $5mm.
The institutional investor class has the same problem in reverse. Pension funds, sovereign wealth funds, and family offices want exposure to the AI infrastructure buildout. But their only options are public equities which come bundled with every other risk factor those companies carry, as stated above.
All of this represents a structural failure amounting to billions of dollars in unmanaged risk and missed opportunity.
Participants in the Compute Markets
AI Startups (Hedging Function). Oftentimes, AI applications & services will price a product for enterprise customers at a fixed fee (e.g. $100/mo for Claude Max 5x), which means margins are directly exposed to GPU price fluctuations. With a futures market in place to fix in predictable COGS, product margins stay intact and pricing & opex plans can be amended accordingly.
Data Centre Operators (Automatic Revenue Stabiliser). Operators of data centres, given their often large scale and capital intensive nature, need predictable revenue to service their debt. In a scenario where compute markets are liquid and efficient, these parties could sell futures on their capacity at a fixed rate, effectively converting volatile spot revenue into a stable income stream. This in turn makes it easier to raise more debt to expand capacity and take their slice of the market from the hyperscalers as cash flows become predictable and bankable. This is the same mechanism that enabled independent power producers in the 1990s electricity market to finance new generation assets.
Commodities Traders (Spreads). The likes of Trafigura, Vitol or Glencore have spent decades trading crude oil, natural gas, metals and untold other commodities that form key inputs to markets in some way or another. Compute looks like a new commodity class with all the characteristics that these players know and love: volatile prices, information asymmetry, complex supply dynamics, regional basis differentials, and an immature market where sophisticated participants can capture spreads.
In any vision for the AI economy, these firms can trade basis (old-gen vs. new-gen depreciation differential, be that H100 – H200 basis trades or GPU – LPU baskets once the index infrastructure is available for the latter), cross-regional arbs (US vs. EMEA pricing differentials driven by energy costs), and calendar spreads (near-term vs. far-term futures reflecting supply expectations). There is a clear opportunity for such firms to provide liquidity that is desperately needed and earn a healthy bid-ask as the market grows towards maturity.
Autonomous Agents , Autonomous Procurement. In a world where AI agents are executing tasks autonomously and paying for their own compute in real-time, agents themselves become market participants. An agent that needs to run a complex reasoning task in four hours could bid for GPU-hours on a spot auction. An agent with a predictable recurring workload could hold futures contracts to lock in its own operating costs. The idea that non-human entities participate natively in financial markets is novel, but it’s an entirely logical extension of the current trajectory of agent-based AI systems.
Existing approaches to Compute Markets
Several companies are already building pieces of this infrastructure, each approaching the problem from a different angle.
Physical Marketplaces
SF Compute is perhaps the closest thing that exists today to a genuine marketplace for GPU capacity. SF Compute operates an order book where buyers can purchase GPU-hours on flexible, short-term contracts. Critically, buyers who’ve purchased capacity they no longer need can resell it to others.
The model is essentially Airbnb for GPUs: SF Compute doesn’t own any hardware, but manages over $100 million worth of GPU infrastructure on behalf of data centre operators. It takes roughly 10% of each transaction. The platform currently lists H100 and H200 GPUs, with B300s coming in Q2 2026, and offers spot rates that frequently undercut the major hyperscalers.
What SF Compute solves is the liquidity problem in the physical compute market. AI companies that signed 12–36 month GPU contracts during the 2023 scarcity period can now offload unused capacity rather than eating the cost. Data centre operators can monetise idle hardware. The platform introduces the beginnings of price discovery through visible bid/ask dynamics on an order book.
What it doesn’t do is provide financial derivatives. You can buy and sell compute time, but you can’t buy a futures contract on where compute prices will be in three months. You can’t hedge your exposure without actually transacting in the physical market.
SF Compute is a spot marketplace, not a financial exchange. It is more akin to buying a truckload of grain from Walmart than from the CME.
Financial Exchanges
Ornn is one player making a serious attempt at building the financial layer. Founded in 2025, Ornn is building what it describes as the first regulated compute futures exchange.
Ornn’s approach starts with the index. The Ornn Compute Price Index (OCPI) tracks live transaction data for GPU rental prices across H100, H200, B200, and RTX 5090 hardware. This is crucial: you can’t build a derivatives market without a credible underlying benchmark, and Ornn’s indices are designed to reflect actual traded prices rather than listed rates or marketing numbers.
A (very loose) proxy for this index has already been introduced via Ornn’s Kalshi markets for the Price of H100 compute hours a week forward.
On top of the index, Ornn plans to offer cash-settled compute swaps and futures. In a compute swap, two counterparties agree on a fixed price per GPU-hour for a defined period. At settlement, the difference between the agreed price and the average index price over the period is settled in cash. The key point worth emphasising is that the financial hedge is decoupled from the physical procurement.
You don’t need to change where you buy your compute, but rather just lock in an effective price through a separate cash-settled contract. In January 2026, Ornn partnered with Architect Finance (AX Exchange) to launch what they describe as the financial industry’s first exchange-traded perpetual futures contracts on compute, pending regulatory approval.
Compute Exchange, launched in January 2025, approaches the same problem from the physical trading side first. Compute Exchange operates as an auction-based marketplace for buying and selling GPU capacity via a trading venue where GPUs actually change hands through transparent, real-time auctions with no NDAs and side-by-side provider comparisons.
The strategic play seems straightforward and rather analogous to the commodities market evolutions of yesteryear – Compute Exchange builds liquidity and price discovery in the physical market, partners like Silicon Data capture that pricing data into institutional-grade indices, and futures contracts eventually get layered on top, something that Compute Exchange could obviously capture end-to-end but that also offers benefits to other players in the market.
The approaches of Ornn and Compute Exchange put forward an interesting chicken-or-egg design question. For a new commodities market, do you need a liquid spot market before you can build derivatives (Compute Exchange), or can the derivatives market itself create the price discovery that the spot market lacks (Ornn)? Oil had a functioning spot market at Cushing before NYMEX launched futures.
Electricity had wholesale spot markets created by regulatory mandate before futures emerged. Compute is trying to do both simultaneously, which is either visionary or premature depending on your disposition. This is something worth following as the market evolves.
Onchain Economic Layers for Compute Ownership
The tokenised compute ownership model that has been expressed by the developments of some adolescent players in the market is possibly the purest representation of the neurodollar idea posited in the introduction, in that they represent a means-of-exchange / unit-of-account denomination backed by the megacommodity du jour.
One of these, GAIB, is constructing what it calls the ‘economic layer’ for AI infrastructure onchain. It tokenises GPU assets and their revenue streams, creating yield-bearing instruments backed by real-world compute demand.
GAIB reached $200 million in total value locked at peak, deployed over $50 million in GPU and robotics tokenisation deals in 2025, and announced a $30 million GPU tokenisation partnership with Siam.AI.
GAIB’s USP is that it turns illiquid GPU infrastructure assets into tradeable, fractionalisable onchain instruments. A data centre operator in Southeast Asia can access capital from a DeFi investor in Europe without going through a bank. Investors can gain direct exposure to AI compute yields without buying hardware or equities.
The limitation is that GAIB is primarily a financing and yield product, not a price-discovery or hedging tool. It gives you exposure to compute infrastructure returns, but it doesn’t let you hedge against GPU price volatility or take a directional view on where compute pricing goes next quarter.
USD.AI takes a more aggressive approach to the same underlying thesis. The distinction between USD.AI and GAIB is important to understand, though they both look similar at face value. GAIB is a yield product – you deposit capital, it gets deployed into compute infrastructure, you earn a return. It’s passive exposure to the AI buildout. USD.AI is a credit product — it creates a lending market where GPU hardware is collateral and the stablecoin is the liability side of a balance sheet backed by physical compute assets.
USD.AI introduces a stablecoin (USDai) collateralised directly by physical NVIDIA GPUs housed in insured data centres. The protocol issues loans to AI companies using GPU hardware as collateral, reportedly cutting approval times by over 90% compared to traditional lenders. Borrowers get capital to buy GPUs; depositors earn yield from the rental income those GPUs generate, with the protocol targeting returns between 13% and 17%.
Both approaches validate the same underlying thesis: that GPU infrastructure generates predictable enough cash flows to support financial products, and that onchain rails can deliver that exposure more efficiently than traditional capital markets. Where they differ is in who they’re for. GAIB appeals to DeFi investors who want yield exposure to AI infrastructure. USD.AI appeals to AI companies who need capital and to stablecoin holders who want higher returns than Treasury-backed alternatives can offer.
Rather than providing hedging or price discovery, these protocols form complementary layers of what will eventually be a complete financial stack for compute that encapsulates all of price discovery (indexes), hedging (derivatives and forwards) and credit and capital formation (via a variety of means, both novel and well established like private lenders such as Magnetar Capital).
Data
No derivatives market can function without credible, transparent benchmark pricing. Silicon Data has positioned itself as the index provider that transforms these data streams into more comprehensive bases for broader scale compute financialisation. Its flagship product, the Silicon H100 Rental Index (SDH100RT), launched in May 2025, tracks daily GPU rental prices across hyperscalers and neocloud providers, covering over 80% of the available H100 rental market. Note: Silicon Data is developed by the same team as Compute Exchange.
Silicon Data now publishes daily on Bloomberg terminals, providing the kind of institutional-grade pricing infrastructure that futures and derivatives markets require to function.
An A100 index is in final testing, and expanded coverage across B200 and other SKUs is expected through 2026. The easy analogy is to think of Silicon Data as the S&P or MSCI of compute.
Risks still abound…
What If AI Capex Is a Bubble?
The most existential risk to compute derivatives markets is that the underlying demand collapses or was never quite there in the first place. If AI turns out to be a technology whose capabilities plateau, whose economic returns disappoint, and whose capital spending proves to have been wildly speculative then the entire thesis unwinds. Data centres become stranded assets, GPU prices fall to commodity scrap value and trading volumes turn to tumbleweeds.
There are reasons to think compute is different to other precedent industries where supply buildouts ran well ahead of demand. AI is a general-purpose technology still in its early deployment phase. Even skeptical analysts agree that AI will generate significant economic value the debate is over the magnitude and timing, not the direction. More practically, even in a downturn scenario, compute doesn’t become worthless. GPUs can be repurposed for scientific computing, climate modelling, drug discovery, rendering, and a dozen other workloads. The floor is higher than zero, even in the worst case.
Regulatory Hurdles
Compute derivatives face a genuinely novel regulatory question that doesn’t have an obvious precedent. Under the Commodity Exchange Act, the CFTC has jurisdiction over futures, swaps, and options on commodities. However, a “commodity” per the CEA’s language is defined expansively enough to include virtually anything that isn’t an individual security. GPU compute-hours almost certainly qualify as a commodity under this definition, which means that any exchange offering standardised compute futures or swaps to US customers (currently by far the dominant market at least from a hedging needs perspective) needs some form of CFTC licence or exemption (for example, Ornn currently operates under a de minimis exemption for swap dealing).
Technological Disruption
What if quantum computing achieves useful fault-tolerance and renders GPU-based compute partially obsolete? What if photonic chips (Lightmatter) or neuromorphic processors (Intel Loihi) create entirely new categories of compute with different pricing dynamics? This is a genuine risk but also an immense opportunity. A well-designed compute derivatives infrastructure can accommodate new modalities by adding new index products and contract specifications. The exchange that builds flexibility into its architecture can capture the entire spectrum of compute evolution rather than being tied to a single hardware generation. Where do we go from here?
From Speculation to Full-Service Banking
Perpetual swaps and futures are just the beginning. The natural evolution of a compute financial market follows the same trajectory as every other commodity market, whereby forward contract prices form the basis for spot trading benchmarks, which then sets a baseline for futures and options prices, which in turn can be packaged in sophisticated structured products, insurance and lending instruments.
As with any other commodity, the impact and necessity of such products to the stakeholders in the market is quite visible. Insurance products can protect data centre operators against GPU depreciation, whereby they pay a premium today, and if the resale value of hardware drops below a specified floor, the policy pays out (this was introduced by Ornn in a blog post that I now cannot source).
Compute-backed lending, where an AI company’s contracted future compute capacity can serve as collateral for a loan, with covenants that the borrower may have to hedge their GPU-hr price exposure in order to make the lender more comfortable. Credit facilities could even feasibly be denominated in GPU-hours rather than dollars, where the unit of account is compute itself.
The ultimate vision is something like a full-service bank for AI infrastructure: an institution where data centre operators can hedge revenue, AI companies can hedge costs, investors can earn yield, lenders can underwrite infrastructure, and the entire ecosystem operates on a transparent, composable financial layer.
Compute abundance and avoiding ‘Peak Intelligence’
One of the most fascinating questions in this space is whether compute faces a supply constraint analogous to ‘peak oil’, i.e. a point at which physical limits on chip manufacturing, energy supply, or cooling infrastructure cap the total amount of compute available globally.
Data centres already consume roughly 1–2% of global electricity, and that share is growing rapidly. Meta’s planned 5GW facility in Louisiana would consume more electricity than many small countries. The industry is racing to secure nuclear power, geothermal, and other baseload energy sources to feed its appetite.
As you may have heard in any commentary surrounding the impending SpaceX IPO, multiple ventures are exploring orbital data centres.
These facilities would operate in space where cooling is essentially free (the vacuum of space is nature’s most efficient heat sink), solar energy is uninterrupted and abundant and the physical footprint constraints that throttle terrestrial buildout simply don’t apply. It sounds like science fiction until you remember that the economics of launch have collapsed by over 90% in the last decade thanks to SpaceX, and that cooling alone accounts for roughly 30-40% of terrestrial data centre operating costs.
If you’re building a compute index that needs to price capacity across every available source, orbital compute is the kind of supply shock that turns a forward curve upside down. The traders who see it coming before the index reflects it will make a fortune…if they have a means to trade it.
Agent-Native Finance
Today, agents already purchase compute through human-intermediated procurement processes, whereby an engineer provisions cloud instances, a finance team approves the budget and eventually a contract gets signed. But as agents become more autonomous, the logical endpoint is agents that manage their own compute budgets.
An agent that runs complex reasoning tasks could maintain a ‘compute wallet’ with a budget of GPU-hours acquired through spot auctions or futures contracts, and spend down that wallet as it executes tasks. An agent running a recurring workload could buy forward contracts to lock in its operating costs, exactly as a human CFO would. In a world where agents have access to a compute wallet as outlined above, it makes sense to have underutilised portions of that budget ‘lent’ elsewhere to other agents for some kind of return from an active, high-utilisation, short-duration agent that uses those tokens/compute hours to conduct productive work.
If you build the financial infrastructure so that agents are first-class citizens with programmatic access to order books, margin systems, and settlement mechanisms you create a financial system where autonomous software entities participate natively in markets The exchange that nails the agent-native API and offers terms so compelling that agents wouldn’t optimise to secure compute anywhere else becomes the default financial infrastructure for the AI economy.
Compute as Reserve Asset
If compute is the fundamental resource of the AI economy, could it become a store of value, a unit of account, and a medium of exchange – in other words, money?
Consider a stablecoin pegged not to the dollar but to a basket of compute credits – a token that always redeems for one standardised GPU-hour of compute. For agents operating in an AI-native economy, this might be more useful than a dollar-pegged stablecoin, because their costs are denominated in compute, not dollars. A compute-pegged stablecoin would offer natural inflation protection (as hardware improves, the same token buys more useful work), built-in demand (every AI workload needs compute), and native integration with AI-first financial infrastructure.
Closing Thoughts
AI startups are looking down the barrel of massive compute budgets that swing with a price they cannot predict, hedge, or control. They are running a
business on top of the most important commodity of the twenty-first century without any of the sophisticated risk management tooling that its forebears in the hard commodities sectors were blessed with.
As with any major commodity of any meaningful importance, markets for compute are transforming from a flight of fancy into an inevitability as we speak. Given the problem outlined above, the establishment of markets for this commodity and their potential to solidify the commodity as a synthetic store-of-value itself can provide more transparent truth to a chaotic market and lead to more sustainable development in the sector.