Breadth or Depth: Is Efficient Compute Market Even Possible?
Compute market participants and their incentives.
I’ve been battling with the problem of benchmarks and unit standardisation since I first dived into research on compute markets. After all, a perishable, non-fungible good with highly unpredictable usage doesn’t sound like a fit for a large, liquid market.
But I want to see it happening. I want to see new asset classes, new markets and new sources of wealth creation.
So do many others. Hence, there are some great ideas for how to develop a usable benchmark. The latest that gave me lots of food for thought is the one from Ariel, who made a compelling case that compute needs a grading system rather than a universal unit.
But his article inspired me to take a step back and start from the basics to better understand what market participants actually need.
A grading taxonomy is necessary. But it is not the same thing as a market. You can have perfectly calibrated scales and still have an empty bazaar. The question of what makes a compute market big - deep, liquid, active, worth building exchanges around - has less to do with the elegance of the taxonomy than with a simpler question that commodity market designers have been asking for a century: who shows up, and what do they want?
So this is a thought exercise. Start with the participants, work through their incentives, and see what falls out the other end.
Meet The Crew
A mature commodity market has a predictable structure. Compute is new, but one can extract a basic classification. Six groups matter.
The Big Boys - frontier AI labs and hyperscalers, consuming compute at civilisation-scale budgets. There are few but highly sophisticated and they operate at massive dollar scale. The contracts they sign are currently among the dominant anchors of the market.
The Ironmongers - data centres and neocloud providers who own the bare metal. They’ve taken on serious debt to build capacity and are the natural sellers of future compute. Their entire business model stands on the ability to utilise their machines to the fullest possible scale, they like long-term contracts with creditworthy clients.
The Consumers - a sprawling middle of the market - firms that consume AI compute. This is the most diverse lot, from enterprises with serious consumption levels (but not at frontier-lab scale) to AI startups and agent workflow runners. They are large in aggregate and modest per transaction.
The Speculators - as old as the commerce but highly sophisticated these days. Basically, the traders who have no interest in compute itself and every interest in its price. They bring no chips to the table, only liquidity.
The Financiers - banks and project finance funds that don’t trade directly but shape everything that does, because they decide whether datacenter buildouts get funded. They don’t really care about the price of compute per se, but rather if the operator of the facility they lend for can hedge their revenue. If the answer is no, they don’t lend, and the infrastructure build out slows.
The Plumbing. Exchanges, indexes, brokers, clearinghouses. They don’t have incentives of their own so much as they have metabolic requirements: they need enough standardisation to settle trades and enough volume to survive.
Six groups. Same underlying asset. Wildly different desires.
What Each Group Actually Wants
I have a strong belief that the best way to understand the world is to understand the interests and incentives. This is also a basis for market design. Each participant has a specific need, and those needs are often not compatible with each other in any obvious way.
The Big Boys want long-term contracts with predictable prices for reliable capability. They know exactly what they want from a provider and want to engage in the technical nitty-gritty to ensure their compute needs are covered. They want precision and bilateral OTC agreements at confidential, case-by-case negotiated prices serve them best.
The Ironmongers want to recover capex on expensive GPUs fast before they depreciate, hence they want to sell the maximum capacity, ideally on long-term contracts to creditworthy clients. They also want to get the maximum price for their offering, the premium operator with newest chips and great up-time does not want to be compared to a shoddy provider operating from a converted garage. They want the market to price their value properly, showing their differentiation.
The Consumers want a combination of fair price and quality they can trust. They’re more price sensitive and want to optimise for the best performance per dollar spent. Their usage is more unpredictable and more bursty, so they appreciate flexibility and are likely to shop for providers (once there are more reliable actors on the market). For that, they want easy comparison, such as a benchmark number they can budget against and occasionally hedge. They want simplicity and clarity.
The Speculators want depth. They will trade almost anything, but they will trade it seriously only if the order book is thick enough that they can move size without moving price. Fragmented markets with thirty thinly-traded contracts are their personal hell. What they optimise for is liquidity.
The Financiers want credible instruments. They want to lend to operators with bankable cashflows that can hedge their exposure to reduce the risk. That requires trusted benchmarks, a reliable forward curve to allow them to underwrite debt against future revenue.
The Plumbing simply wants activity. Exchanges live on transaction fees; indexes live on data licensing. Both need participants to actually show up. From their perspective, the more volume, the better.
Breadth or Depth?
The immediate tension that’s visible from this little mapping exercise is the one between granular, precise pricing preferred by the main and biggest participants of the market, and the need for simplification that is in the interest of newer participants. One could argue that the current state of the market addresses the needs of the Big Boys really well, the price is not discovered but directly set by them, driven by their own needs and availability of compute infrastructure. The bilateral deals are convenient, address their needs for precision of the order, and can hide a serious competitive advantage.
That is, however, a scarcity-driven view. What turns a functional market into a big market is the arrival of participants who have no physical stake - speculators, market makers, hedge funds, banks. They multiply the transaction count by an order of magnitude or more. They tighten spreads. They provide the liquidity that lets natural buyers and sellers execute at fair prices. In most mature commodity markets, financial volume is often several multiples of physical volume. The stronger and more credible the financial infrastructure, the easier it is to lend for and invest in the further development of the physical infrastructure. It is an argument for the growth of the pie that in the end benefits all the participants, even the Big Boys.
So it pays off to create the market that is a tradeoff between breadth and depth. Breadth is how many distinct products are listed. Depth is how much trades in any one of them. Many commodity markets combine both: sovereign debt markets have thousands of distinct securities but concentrate liquidity in a handful of benchmark issues. Oil has hundreds of recognised grades but two dominant pricing references. Agricultural markets have regional variations layered over a small number of globally-quoted contracts. The breadth exists to accommodate real-world heterogeneity. The depth exists at a few specific points where everyone has decided to meet.
This is where Ariel’s idea about a grading system based on compute quality shines. The grading taxonomy provides the breadth. The question is what sits at the depth point, and whether the market can coordinate on a small enough number of flagship contracts to actually get deep.
The Three-Tier Structure That Falls Out
To reach the compromise between breadth and depth the market will likely settle into tiers that most directly address the participants’ needs.
At the top sits a flagship tradable contract, likely settled against a benchmark index that serves as the primary reference for price discovery.
Something like: a standard GPU-hour of the dominant production chip, with the specific networking, with a grade for data centre quality and reliability. It will likely be the specification most dominant in the real market at the time (the consequence being that the benchmark would actually rotate in chip development cycles).
This is what speculators are likely to trade. This is what the published price series reports. This is what enterprises reference for their internal budgeting and what financiers point to when they underwrite operator debt. This is what media reports when it says “the compute prices are skyrocketing”.
Beneath that sit spread products. This is where the grading shines. The prices reflect specific simplified capabilities. A premium product for top-tier capability; a discount product for capacity that falls below the benchmark floor but is still useful for particular workloads. They are less liquid than the flagship, they have fewer participants who mostly care about the price of specific workload/capability and they allow the high-quality, reliable providers to show their differentiation.
Underneath that lives the bilateral OTC layer. Capability-exact physical contracts between specific counterparties, using the full grading taxonomy for dispute resolution and attestation but not trading on exchange.
The OTC layer handles the deals that need precision; the benchmark provides a common pricing reference, while the flagship exchange-traded contract could concentrate liquidity and contribute to price discovery.
The three layers satisfy different participants’ needs, allowing for the market that’s both liquid and precise. They’re all neatly related to each other because they’re a part of the same taxonomy, one that reflects the actual commodity specification with a degree of detail that allows for differentiation of workflows and compute needs, serving as a credible basis for price discovery.
Wen flagship index?
The beauty of markets is that they are self-organising, so the above-described structure is more an observation of market dynamics than a prescription. It is already happening in real time with numerous indexes tracking the GPU per hour prices, although with much simpler taxonomy.
As I mentioned in my very first article about benchmarks, the one that will win will be the most useful, the most robust and trustworthy. Then I focused on the methodology of data collection and IOSCO principles, now it is clear to me that it also requires a proper specification and grading system to better reflect the heterogeneity of the product.
The race for it is still on and I believe we are still a few years before a commonly recognised benchmark index that’s widely tradable. And I will be here for it.
If this piece resonated with you, subscribe for more takes on intricacies of compute market structure and attempts to track and define how this space will evolve.