One of the best books about trading is The Coffee Trader, David Liss’ account of the development of the coffee market in the seventeenth century Netherlands. Protagonist Miguel Lienzo recognizes coffee’s potential and hatches a scheme to crash the East India Company supply and then buy it up cheap before the market wakes up to exploding European demand. He moves his plan forward with a series of secret auction schemes and shaky partnerships across the taverns of Amsterdam. The legal ethics and bet sizing are a bit hairy. But a modern trader can see in Miguel the seeds of a new world, the trader’s world.

One nice moment in the book is an account of the benefits of caffeine. “Coffee,” a promotional pamphlet proclaims, “has all but destroyed the plague in England. It preserves health in general and makes those who drink it hearty and fat; it helps the digestion and cures consumption and other maladies of the lung. It is wonderful for fluxes, even the bloody flux, and has been known to cure jaundice and every kind of inflammation.”

Respect the copy, but the world didn’t get addicted to coffee because of its health benefits. The pamphlet saves its honest pitch for last: “Besides all that, it imparts astonishing powers of reason and concentration. In the years to come the man who does not drink coffee may never hope to compete with the man who avails himself of its secrets.”

Today, more than was the case with coffee, the man who does not use computation cannot compete with the man who avails himself of its secrets. No hedge fund can price second order effects faster than a 200-megawatt Jane Street data center. No lone genius can produce a better language model than Anthropic.

It’s not just that companies need compute to serve software applications. Many of our biggest companies, with highly scalable business models in things like search and learning, actually have “hyperscale” demand for compute that appears limitless.

Whereas the seventeenth-century market for coffee beans was measured in perhaps the equivalent of millions of modern-day dollars, the computing and cloud market today measures in the hundreds of billions a year, with trillions more committed to building it out. And whereas the coffee trade involved a vast colorful tapestry of characters from Turkey to Ethiopia to Indonesia to Amsterdam, the compute market concentrates into a shockingly small number of multi-trillion-dollar actors.

Magic Beans

Coffee is a simple commodity: coffee beans are grown in farms in places like Colombia or Tanzania, graded, bagged, and exported to global markets. The grading allows a buyer in, say, Hamburg, to price a lot of coffee he’s never seen before. Compute, meanwhile, is harder to even define. The thing that’s sold is not a commodity but a sellable capacity for work done on computers, purchased by the hour. When a startup “buys compute,” it is renting time on racks of specialized computers in massive data centers. You cannot put compute in a bag, store it for later, or sell it in a farmer’s market. No grade exists that tells a buyer what he’s getting.

Why is grading impossible? Compute is extremely nonfungible. Even identical hardware running identical software can produce wildly different performance results depending on a whole variety of subtle differences in how chips might be networked, powered, cooled, and used. Depending on a thousand hidden choices made by your counterparty, the same standard Nvidia H100 — the $30,000 gold standard chip that powers almost all of the AI consumed in America today— from Company A might be 38% more powerful for your workload than the same exact computer from Company B. Think less coffee beans and more Bertie Bott’s Every Flavour Beans. That is nonfungibility.

But nonfungibility is not in itself a fundamental barrier to commodity trading. As Dave Friedman argues in a Substack piece, “Can Compute Commoditize if it’s not Fungible?,” lots of commodities that people trade every day are nonfungible. Electricity is nonfungible at the local level. A watt-hour in a data-center town is different from a watt-hour in New York City. And both are different from a watt-hour next to a solar farm at noon on a sunny day.

Electricity markets solve this with a local reference price. Power is usually indexed at a “node” — roughly, a town or an area served by a plant. Big trading hubs branch into these nodes. Then there is other information that determines what power is worth in one place at one point in time: time, load, generation, transmission constraints, losses. To the extent that this other information is known or predictable, traders are willing to add “basis” to the reference price and trade. Factor in congestion and loss on a $10 reference and you get to $9.83, or $10.17, or whatever the case may be. And just as a crowd at a county fair, averaged together, can guess the weight of a cow almost perfectly, a richly traded market becomes very good at predicting and collapsing all those mysterious long-tail unknowns lurking in the GPUs. For example, an operator with higher incidence of thermal failures would have that factored into its output price like beans from a leaky warehouse. Over time, spreads narrow, and basis risk dissolves into prices.

The reference for GPUs could be something like the price of an H100-hour in a certain region, predicted via an AI trained on the entire history of transaction data available to the reference provider. Nobody needs all GPU-hours to be created equal; the market will learn what the difference is worth, then quote it.

Why would this be exciting? Commodity trading markets would allow people to do all kinds of interesting things that are prohibitively hard or impossible now. A startup could lock in next year’s compute at today’s price, the same way Southwest locked in cheap jet fuel before oil prices spiked. A neocloud that borrowed half a billion dollars to buy GPUs is making two bets: that it runs a good cloud, and that GPU prices will stay high. A market would let it pay someone to take the second bet off its hands, so its survival depends only on what it actually controls — the quality of its cloud. And a trader who thinks compute is mispriced could bet on that opinion without buying racks of computers. Today, none of this is possible.

Markets also make the system smarter. A mature market like oil can make sure no gas pump goes dry without any international coordination around a big event like the closing of the Strait of Hormuz. A compute trading market could do the same: route every bit of the scarce silicon our Taiwanese benefactors allot us towards its most economically useful work, priced and competitively weighted to perfection.

And the information generated by all these open prices and signals — in an otherwise-opaque tech stack — could produce a Hayekian explosion of open data, the mirror image of the advertising-driven Big Data explosion in the 2010s. That decade brought the signals previously hidden in the world into the clouds: every click and purchase and location ping was harvested and centralized, thus financing most current data center sites and catalyzing the success of the transformer and large language models. This decade could bring the knowledge currently hidden in the clouds back out to the open markets, where anyone can read the prices.

A Dutch time traveler who looked at our data-center-covered nation and heard about our mighty capitalist ways might have guessed that trading would be the way things developed from the start. He might have guessed we’d live in a society of compute bartering and bit auctions; after all, Amsterdam pioneered the stock exchange and systems where the trading of companies remained tightly bound to the trading of commodities. He might fail to understand how knowledge could get so locked up in clouds in the first place. But that trading world is not the world we live in.

Compute today trades something like coffee before the market. To buy and rent Nvidia H100s — a basic chip that powers almost all of the AI consumed in America today — is to play a game of relationships and handshakes and allotments and bilateral agreements. Nvidia gives people allotments of computers, with big blocks reserved for favored clouds and neoclouds. Exact deal terms and basis points don’t matter much; all that matters is your counterparty and how a deal ties into the great game of big tech strategy.

So, there is nothing to trade. There are no H100-hours to trade in America today. There are no goodput futures. No token indices. (More on these later.) There is no commoditized place for a humble army of quantbots to rent the spot compute they need for their agent sandboxes. It’s not even clear what the tradable unit would be. Future deliveries of hardware boxes? Rental hours on those boxes? Higher abstractions, like guarantees of useful multiplications that power model training? Lower ones, like ultra-localized watt-hours? Congestion? Memory? None of it exists at scale. There are just the big cloud companies, renting compute alongside a dizzying suite of value-added services, intermediate networking and routing, and other add-ons.

Why? The reason has a lot to do with the complexity of the technology and the tiny number of companies that have ever been able to scalably deliver state-of-the-art computing.

Imperfect Markets

Originally there was IBM, the typewriter company. IBM operated the room-size mainframes that first commercialized universal computation, doing useful math for companies, like Boeing or Cargill. Then Intel came and miniaturized those computers into chips small enough to cover the earth in them, enabling a Cambrian explosion of software written on top of the bare metal, and eventually the Internet. Intel’s cofounder Gordon Moore discovered the existence of scaling laws in compute that came to bear his name: the more silicon we build, the better compute gets; the more attention and data and money flows to semiconductors, the more silicon we build. Tech and demand climb together, with no plateau in sight.

Today the Intel role — the company whose silicon the whole industry runs on — is played by Nvidia, which designs the vast majority of AI chips. And the IBM role is played by big tech companies like Google and Amazon, which make money by filling data centers with computers and using them to run the products and services we rely on every day. These companies rent out the rest as “cloud” compute, so businesses can avoid operating their own hardware. Right now, when a startup needs to buy compute, they rent pre-set-up boxes that they access as virtual machines or even abstract that layer entirely and just make calls to metered cloud APIs.

The “cloud” shift back to IBM-style centralized computing facilities was pioneered by Amazon in the mid-2000s. Amazon Web Services (AWS) launched in 2006, but the cloud conquered in the 2010s. Amazon’s constellation of business units found it easier to share one set of state-of-the-art computers and systems rather than deal with the nightmare of each running their own. Amazon recognized before anyone else that what worked internally would extend to every business on earth, as it has. Today, when we use a computer, we make outsourced calls to a hyper-vertically-integrated tech stack owned by a set of companies whose CEOs would fit at one poker table. We don’t touch bare metal. We touch search, streamed video, and a thousand other convenient abstractions. It’s monstrously complex, and the companies that can do it reap monstrous margins.

So-called “neoclouds” like CoreWeave and Nebius offer GPUs on demand as an alternative to the big clouds — the “hyperscalers” (AWS, Microsoft Azure, and Google Cloud). But, fundamentally, the neoclouds don’t change the cloud business model. They are far more eager to differentiate — moving up the stack into runtimes (the software environment that executes code on silicon chips) and AI serving (running a trained model as a live service) — than to commoditize, renting out bare metal by the hour. They want to be, and seem on pace to be, just another flavor of big cloud all the way down. White-hot inference providers such as Modal and Baseten, which are smart AI workload runtimes, have been able to secure some strategic bare metal leases from neoclouds, but generally neoclouds consider selling bare metal to be low-margin, undifferentiated, and best left to a messy world of existing and emerging data centers too un-credit-worthy or otherwise not quite right for the clouds.

The cloud obsession permeates the entire tech stack. Everyone wants to work with, lend to, and buy from Google. The hyperscalers borrow at lower interest rates, extract far better terms than any other players could get, and extend those benefits to their ecosystems. Private developers and speculators build for clouds. Any data center with a five-year Google commitment is going to get funded by everyone with a pulse in less than 48 hours.The big tech companies are price makers, and everyone else is a price taker.

Even huge-but-not-ginormous tech companies face a different financial reality. A recent nine-figure data center project — built by Crusoe, filled with AMD chips, and financed by investors including Goldman Sachs — required AMD to personally guarantee it would use the chips if nobody else would. Nvidia would never need that. Its demand is guaranteed.

Brannin McBee, a cofounder of the Nvidia-focused premium neocloud CoreWeave, recently indicated on the Bloomberg podcast Odd Lots that he sees no evidence demand will shift anytime soon: “We really don’t see demand on a material basis for anything but that Nvidia compute, and that’s what we are building today.”

CoreWeave specs its units to DGX, Nvidia’s official reference design for networking and cooling its chips, and claims performance advantages over non-AI-specialized clouds using a suite of neo-metrics for AI throughput. The primitive units for these metrics include floating-point multiplications, dollars, time, bit packet throughput, and watts. MFU, Model FLOPs Utilization, is the benchmark that measures utilization relative to physically possible peak performance. Goodput is a throughput measurement for what was useful for training models, now adapted for inference applications as well. Getting back to the nonfungibility question, every computer will score differently on these metrics, even on the same exact hardware, and on a thousand other metrics we haven’t invented yet.

The real issue with nonfungibility is not that it is an insurmountable barrier to trading; rather, it is that no cloud company wants to participate in a commodity trading market because to do so would be to equate their offering with being commoditized. Joining a commodity market is a public confession that their product — compute — is somewhat interchangeable with everyone else’s. The absolute most obvious use for compute futures would be for neoclouds to sell futures as a hedge against their structurally mega-long position on compute. A neocloud is, in effect, a leveraged bet that GPU prices stay high: if the price of compute goes up, their core business prints; if compute goes down, their downside is capped at a level that underwriters can accept. But neoclouds do not want to associate with any kind of commodity exchange, as each views their compute as differentiated in some unique, magical way that today’s crude metrics cannot capture. The other piece is that the big neoclouds already have good hedging via handshake bilateral agreements with other compute companies, customers, and financial entities. Major hyperscaler-neocloud agreements include abilities to assign obligations in case of default, and agreements often include worst-case-scenario-capping guarantees of volume or upfront cash.

No big company really wants a compute trading market. The clouds don’t want a commoditized alternative that could undercut their pricing. OpenAI and Anthropic are just as structurally opposed as the clouds because of their evolution into clouds themselves, and they need their internal compute economics kept secret because their tech stack choices can give away billion-dollar research insights or sensitive experiments or gambles. Nvidia doesn’t want to upset the apple cart. And AI infra companies want to handle their levels of the tech stack internally, to maintain control over efficiency, their moat, and pricing power.

The natural buyers of compute futures — startups and inference providers that don’t operate data centers — also don’t necessarily have what it takes to participate. They don’t want to deal with bare metal or complex infra, so they would likely go through intermediate clouds or providers, and the friction of a middleman may blunt the efficiency that makes trading worth it. Not every commodity is suitable for trading. History is littered with ones like lithium, where unpredictability and transaction friction prevented the magic of markets from getting off the ground.

The list of obstacles to compute trading doesn’t stop there. Physical settlement is ugly as delivering compute means provisioning verified machines, amplifying the nonfungibility problem. Standardization will be a knife fight. Security and fraud questions will bite. Brokerage costs and basis spreads would be wide enough to scare small children. It’s possible, even likely, that compute trading never amounts to much more than a crypto-like speculative curiosity in the shadow of the big clouds.

To get over these obstacles, you would likely need a rebel alliance of smaller chip companies (where smaller could mean AMD which sits at $800B+ ), startups, AI consumers, finance players, and — crucially — at least one giant tech company willing to flip on the oligopoly, commoditizing its complements or otherwise improving their odds of making it to the final table. Catalysts could include Chinese chip dominance, a neocloud loan default, a major AI offtake customer blowing up, the dawn of locally runnable small models, government breakups of major clouds, or any other cataclysmic shock to the system in the next few years.

The Rebel Exchange

Exchange startups are trying to catalyze the rebel charge by providing a first wave of index reference prices and market-clearing data.

Compute Exchange, headed by Bloomberg LP alum Carmen Li and backed by Don Wilson of DRW, is one of these companies. It uses historical AI modeling and data on historic transactions to predict and track prices for different types of compute, and has a partnership with CME to offer compute-hour futures. Ornn, a New York City startup led by Kush Bavaria and Wayne Nelms, is doing the same thing, partnering with CME’s rival exchange, ICE. These companies are competing to be the Bloomberg of GPU pricing: the data providers whose numbers underpin every transaction. Quietly, a third company, Compute Desk, has actually won the race to the actual Bloomberg terminal — it supplied its GPU pricing data for use on Bloomberg. It serves an exchange called the American Innovation Exchange, operated by the derivatives platform Architect, which reportedly has two market makers willing to participate and multiple compute providers looking to sell idle compute into the market. And Kalshi, the prediction market better known for its sports and politics offerings, has launched predictions for the future price of renting certain Nvidia chips, creating the first mainstream prediction for future compute rental prices over time.

These companies believe that they have what it takes to produce reliable enough, and eventually granular enough, references. Then, markets will handle the rest. Carmen Li calls this a “Carfax for GPUs.” The exchanges provide data on GPUs, often bundled with verification and fraud prevention for advertised specs. Li argues that the media confusion about the residual values of GPUs – do they depreciate in four years, or two, or seven? – is an artifact of our bad information ecosystem. Markets, on the other hand, can process the question with more efficiency and less volatility. Her plan is that compute markets will start small, prove their value, gain adoption, and challenge the cloud monopolies.

On July 14, Kalshi CEO Tarek Mansour announced Kalshi’s offering of GPU rental price prediction with a similar message: “Compute is still not uniform enough, but this is a step towards standardization as forward curves will help us see the rise and fall of different model prices and how they correlate.” Mansour pointed out that derivatives could grow to 10-20x the size of the existing compute market. Early betting volume has been thin, with a daily volume measured in the tens of thousands of dollars — as opposed to the millions or tens of millions associated with bets on the World Cup or NBA rumors. But the news went viral. Perhaps more importantly, it marked the first time hundreds of thousands of people learned about compute trading.

When Brett Harrison, the CEO of Architect, launched a compute trading group on Telegram on June 28 of this year, it quickly filled with dozens of more seasoned traders and quants and speculators. Architect offers GPU futures and contracts, like Ornn or Compute Exchange, based on Compute Desk’s deep transaction data. But Architect also offers contracts on things like energy, appealing to a market of traders who are interested in combining energy and compute and other derivatives into highly opinionated, uncorrelated trades.

Andrawes Bahou, the CEO of Swiss startup Compute Desk, points out that serious compute-intensive companies are already setting up internal risk desks for compute — the same way that Volkswagen has a metals desk and Nestlé has a cocoa desk to trade, hedge, and secure its core inputs.

Bahou, whose team includes a mix of ex-Big-Tech, quants, and improbably one of the discoverers of the Higgs Boson at CERN, described what a compute risk desk would look like: “Companies exposed, not to cocoa, or copper, but to compute, will have a compute desk. As a matter of fact, Meta announced a compute desk six months ago, in light of the recent, but not surprising, news that they are now selling compute. For much smaller companies with hedging needs, but without a dedicated team for that, a financial institution will serve as their outsourced desk. Compute Desk (the company), is, well, just that: the institution where data centers, neoclouds, inference companies, and hyperscalers go to trade or hedge compute.”

In a mature market, the roles of various players are distinct: there are data platforms, neutral pricing indexes, brokers, and clearinghouses. In compute, the lines are fuzzy. Ornn appears to want to play both pricing index and broker at once — as a broker-index. The same is true of Carmen Li and Compute Exchange. The catch for these startups is that existing brokers and neoclouds don’t want to give data on real transactions to upstart competitor brokerages. The ecosystem badly needs someone to provide better pricing data. But the only way to serve that role is to give up on the juicy spreads and hypergrowth potential of brokerage.

Bahou explained the catch-22: “Listed price data, the type you can get from AWS APIs, or off the website of CoreWeave, is practically useless. More than 95% of compute trades bilaterally, in private deals, which are steeply discounted from the published prices. There are a lot of broker quotes shared in private channels, but they are also not useful. What is important is verified data from real executed transactions. A real trade tape. The tension is that nobody wants to share the most private data in their business, especially not with their competitors.”

Nobody is forcing the issue, either. Every middleman with secret sauce that sits between the bare metal and the customer has little to gain from generic price indices and much to lose from disruption to the market ecosystem.

There isn’t even serious pressure on the kinds of smaller players who transact marginal, non-cloud H100s today. Decentralized GPU matching platforms like Vast.ai and Akash already operate as functional spot markets for compute: They enable real-time matching between buyers and independent GPU providers, publish live prices, and require no relationship with hyperscalers for access. For bursty inference (spiky, unpredictable workloads), experimentation, and smaller training workloads, they are often good enough — and meaningfully more accessible than the clouds.

But any Dutch time traveler worth his salt — a man who sat frozen while the glaciers melted and awoke to see our modern world — would know that true markets can do better than Craigslist-like matching platforms. Performance varies wildly across hosts, hardware verification is weak, routing and matching is opaque, reliability is terrible, and the risks and signals that tomorrow’s companies will need to control are hidden far more than revealed. Matching platforms are one part of the messy current ecosystem, and the extent to which they succeed in the medium term is the extent to which they become full-fledged markets.

Equally interesting are the middlemen and developers slouching towards markets — the tidepool of firms that currently connect capital and computers. These are companies that pair private credit with data center projects and build them, or which broker Nvidia B200 deals, or that rent GPUs short-term to individuals and small businesses. They are dealmakers who are making it happen, even without hyperscale connections, even in a wild west era of scarce information and massive spreads.

One of these companies is Compute Labs, which finances compute projects: it helps small clouds and data center developers access capital and actually get their projects built. When I called Compute Labs’ head of capital markets, Warren Hosseinion, he explained how he makes deals happen even without guaranteed hyperscaler offtake involved. It’s full-contact work: fundraising, technical derisking, and concocting highly bespoke mixtures of debt and equity suited to one-off situations. Compute Labs is making money, but it’s been a messy ride, with detours including a valiant but stalled attempt to pair GPUs with crypto instruments. Crypto backgrounds are common among the current crop of dealmakers: USD.AI, a startup which underwrites GPUs as collateral for loans and maintains a GPU registry to prevent double pledging, also began with ideas around tokenizing GPUs. A given day for Hosseinion might include calls with an insurer in Zurich, a creditor in Miami, and a shareholder in a Bitcoin hoodie. A forward curve of prices for compute deliveries at various points in the future would lower costs of capital for new-entrant data center developers and make this process more scalable.

Another interesting company in this space is Hydra, which operates the Hydra Host cloud and Brokkr, an intelligent brokering platform.

Hydra’s director, Kai Golden, acts as perhaps the preeminent bridge between high finance and AI-native neo-inference-providers like Hydra. At the recent DealCatalyst private credit conference in New York, he provided the metaphor of the day when he compared general-purpose accelerators to Legos and specialized chips to scalpels. Whereas general GPU chips can handle most AI workloads, other chips such as TPUs have special features or tradeoffs that limit them to certain cloud tech stacks or workloads. Golden also prompted a new row in a lot of Excels when he said that seven years’ depreciation for current Nvidia GPUs might be closer to reality than the canonical four.

Hydra represents a class of do–it-all companies that can operate a layer of the tech stack between the equipment and the user’s final token consumption, providing good customers access to marginal computation and getting in the token stream. It’s part dealmaker, part market-cloud, part network, part router. Whereas inference runtimes like Modal and Baseten can run these kinds of efficiency-maxxing roles at the level of running code workflows, elbow-grease operators like Kai are figuring out how to make it work at cruder scales. Hydra is young, but Kai told me it has already processed nine figures in transactions.

It’s hard to predict the future. Perhaps tomorrow a rebel alliance — a long tail of chip challengers like AMD, Etched, and Cerebras — will announce a suite of neo-metrics and trading commitments to catalyze better fungibility and go after Nvidia’s near-monopoly pricing power. Compute trading booms into a multi-trillion-dollar derivatives market overnight. Perhaps nothing happens, and compute trading remains a marginal thing that big companies don’t really want. Our wandering seventeenth century coffee trader decides he has better short term options in power or midstream gas or slinging data to frontier labs.

I suspect that, even if the coffee trader abandoned compute, he would be sad to do so. The worlds where compute markets take off are better worlds than ones where decisions about compute remain the private reserve of cloud titans. Every software founder panicking about Anthropic, every investor looking for more granular ways to express a differentiated opinion on the future, every risk taker who wants to insure against something specific, would do better in a world of open trading. A free-markets fan can go further: Any lover of open societies, of freedom from imperial tyranny, should root for a world of open prices, where every bit of information, down to the smallest packet, has options and contractual protections and a chance to prove its worth.

Compute traders, good luck. Keep your brew strong, your leverage reasonable, and your exchange funds separate from your Bahamas investment arms.