Compute Should Trade Like a Commodity
August 2024
Pinned. My standing view on where the compute market goes from here.
The most important input to the next century of growth is priced like a back-room favor. Access to AI compute is held tightly by a few, it costs too much, and there is not enough of it. That is not a temporary supply problem. It is a market-structure problem, and market-structure problems get solved the same way every time: you commoditize the thing, you standardize how it trades, and you let price discovery do the work that opacity used to do.
Intelligence is the product of three inputs: models, compute, and data. Models are already commoditizing fast. A handful of open systems now sit within a few points of the frontier on common reasoning evals at roughly a quarter of the price, and the gap closes every quarter. Data is a parallel arc, later and slower. Compute is the input still trapped in the old world, sold through long-term contracts negotiated in private, with data-transfer fees that have nothing to do with technical merit and everything to do with locking customers in. The way to diffuse intelligence to everyone is not to wait for one lab to be generous. It is to commoditize the complements around compute so the surplus that leaks today to a margin stack and to brokers arbitraging the dark gets returned to the people who build.
History is unambiguous about how this goes. In the 1940s, computing was so expensive that only governments could afford it, housing room-sized machines in hangars and renting them to universities by the hour. What changed was not a single breakthrough. IBM commoditized the hardware while running a very profitable business. Operating systems standardized the platform. TCP/IP gave everyone a neutral protocol to build on, and the internet followed. Oil went through the same passage a century earlier: a chaotic, opaque, relationship-driven trade became a transparent market with reference prices, standardized contracts, and futures, and the entire economy that depended on it grew faster because of it. Compute is sitting exactly where oil sat before the exchanges, and exactly where computing sat in the 1960s. Powerful, essential, and rationed by access rather than by price.
So the work is to build the market that compute never had. Three properties matter, and they are not negotiable. It has to be neutral, so a startup and a hyperscaler meet on the same terms. It has to be transparent, so a buyer can see the real price of an hour of compute the way they can see the price of a barrel. And it has to be efficient, so capacity can be bought, flexed, and resold without a quarter of lead time and a lawyer. Neutral, transparent, efficient. Get those three right and price discovery does the rest: idle capacity comes off the sidelines, prices fall toward their true level, and a developer in Austin or Nairobi or Sofia can reach the same infrastructure as a lab in Menlo Park.
This is also where the geopolitics live, and they are not subtle. Beijing subsidizes its AI champions directly and has already proposed a state-run exchange to give its companies cheap, coordinated access to compute. A state-run market is a market with one customer and one set of interests. The American answer cannot be a more fragmented version of the same opacity. It should be an open market, governed by open standards, that any ally can plug into and any builder can trust. The country that sets the standards for how compute trades will shape who gets to build with AI for a generation. That is a race worth winning on our terms, with markets rather than mandates.
I have spent years close to this market: in the rooms where compute is bought and sold, and in the policy conversations in Washington and beyond about how it should be governed. The pieces that still need building are the ones that turn a trade into a market. The orchestration layer. The ratings. The indices. The rules of the road. Some of these should be commoditized at the category level, governed by open methodology rather than owned by anyone. All of them should exist, and I intend to keep building toward them.
Compute is becoming an asset class. The only open question is whether it trades in the light or in the dark. I am building for the light, and I will keep writing here about how we get there.