AI compute is sold in at least three different units. Until the market settles on one, it cannot price the asset consistently.
Every commodity market depends on a unit that participants do not argue about. A barrel of oil, a bushel of wheat, a megawatt-hour, a tonne of copper. Each was contested before it was standard: what counts as a barrel, at what temperature, delivered where, at what grade. The unit determines the contract, the index and the forward curve.
Compute has not settled on one. A buyer today might purchase an hour on a class of accelerator, a reservation on a cluster for a quarter, or a million tokens of output from a model they do not control. Those are three different goods with three different pricing logics.
Accelerator-hours are not fungible
The obvious unit is the accelerator-hour, and it is the one that fails fungibility. Two identical chips deliver different useful work depending on interconnect, available memory bandwidth, network contention, and whether the job can be interrupted. A chip on a fast interconnect running an uninterruptible job is a different good from the same chip, preemptible, in another region.
This is a grading problem. Crude is a family of grades with a benchmark and published differentials. Copper trades to a purity standard. Compute needs a reference specification precise enough that two counterparties can agree what was delivered, plus a schedule of adjustments for deviations from it. That standard does not exist yet.
Location matters differently here. Compute is not transportable. You move the workload rather than the hardware, and that is bounded by latency, data gravity and data residency rules. It makes compute closer to electricity than to oil. Power markets resolved the same problem with locational pricing, over several decades.
Compute cannot be stored
An hour of unused capacity is not inventory carried into tomorrow. It is gone. This property shapes the market more than any other.
There is no cash-and-carry arbitrage, so the forward curve is not tied to spot by storage economics. The forward price is a forecast rather than a financing calculation. A firm making markets here warehouses obligations rather than inventory, which is a different risk with a different failure mode. Power and freight both developed deep forward markets for the same reason.
Compute is non-storable, non-transportable and imperfectly fungible, with volatile demand. Markets with those properties develop forwards.
Tokens are a derived unit
Tokens are what the end buyer pays for and the unit most often quoted publicly. They are also the furthest from the underlying scarcity.
A token price is a compute price divided by how efficiently a model converts hardware into output, plus margin. That middle term is not stable. Efficiency has improved through better architectures, better serving, quantisation, caching and specialised silicon, so the cost of producing a token has fallen for reasons unrelated to the supply of chips.
A long-dated token forward is therefore a position on the rate of efficiency improvement. The seller is short future efficiency gains. That is a legitimate risk to hold, but it is not the risk most buyers think they are transferring. Tokens are also model-specific and quality-differentiated, which resists standardisation.
The clearer way to hold both units together is the refining structure. Capacity sits upstream: chips, clusters, power, reservations. Output sits downstream: tokens served at a given latency and quality. The conversion margin between them moves on model efficiency and serving competition rather than chip supply, which makes it a separate market with separate drivers.
What we work on
The directional question, whether capacity gets cheaper or more expensive, is crowded and mostly a view on capital expenditure schedules. The structural questions are more useful: what specification is precise enough to trade against, what delivery means when the underlying is a reservation on another firm's cluster, who carries an outage, and how to settle a contract on a good that cannot be stored or shipped.
Those get answered by trading in modest size with real counterparties. Market conventions accumulate from terms that survived. The unit will be settled by whoever negotiates it enough times.
Hidden Lattice Capital · Field Notes
← All insights