An agora is a marketplace that is also a public square. This one is an interactive model of the market being built around AI compute: who sells it, why the price is so hard to pin down, and what a swap or a future actually does for the people exposed to it.
Tap each layer. Compute changes hands four times before it runs a model, and every hand-off adds a reason the price can't be quoted like a barrel of oil.
A simple model of what an H100-class GPU-hour rents for over three years. Move demand, new supply and the pace of chip turnover, then see whether a small operator who borrowed to buy servers survives.
An operator agrees a fixed price per GPU-hour. Each month it's settled against spot: if spot is below the fixed price the operator is paid the difference; if spot is higher, the operator pays. This is the over-the-counter structure a few startups are already writing, in the tens of millions of dollars so far.
A startup needs a steady stream of GPU-hours. It can buy at spot every month, or lock a price with a future tied to a published index. The catch is basis risk: the index tracks one chip in one kind of setting, and your actual rental may not move with it.
The article's sequence, from WhatsApp brokers to clearing houses. Compute is somewhere around the middle rungs, with the top ones waiting on regulators.
Every firm has a "compute guy" buying second-hand capacity through an opaque network of middlemen on messaging apps. Prices are private, counterparties unverified.
Where most small buyers still areVenues such as Compute Exchange list GPUs by location, rental term, price and availability. Transparency improves; the product is still bespoke.
Live nowStartups publish rental-price indices for popular chips. An index is what a future settles against, so whoever defines it quietly defines the commodity.
Live now, few and youngBilateral swaps and forwards (Anera's pilot trade, Luxor's per-server-hour swaps). Flexible, but each contract carries the counterparty's credit risk.
Early, tens of millionsCME Group and ICE plan compute futures with index partners; Architect already lists them from Bermuda. Standardised, margin-backed, publicly priced.
Awaiting US regulatory approvalA central counterparty guarantees both sides of every trade and collects margin daily, which is what lets strangers trade without trusting each other. Jump's investment chief says this is "a long way" off.
Years to decades, by insiders' own estimateSeven Western majors set prices through long-term, bilateral arrangements. The market was large but not open; outsiders took whatever terms they could get.
A handful of hyperscalers and frontier labs make deals in the tens of billions. Thousands of smaller firms have no comparable access and no way to hedge.
Nationalisation broke the cartel's hold; Brent and WTI futures gave everyone a public price and a way to hedge. Price discovery moved from boardrooms to screens.
Governments eye the labs; exchanges and startups build indices and contracts. Whether the analogy holds depends on whether compute can be standardised enough to settle a contract on.
Markets don't just price things; they create new dependencies and new places for things to break. Five questions worth asking as this one forms.
Futures need a reference asset. "One H100-hour" has to mean a specific chip, interconnect, location and service level. Index designers are making standardisation decisions that will shape what gets built and where.
Until there is a clearing house, every swap is only as good as its counterparty. Several of the first movers are small startups; the credit risk is real and currently unpriced.
If an organisation hedges mission-critical compute against an index that doesn't track its actual supply, it may believe it is protected when it isn't. Hedging policy becomes part of continuity planning.
Many of the builders, and some neoclouds, came from crypto mining and trading. That brings real expertise in turning compute into a tradable asset, and a track record regulators will scrutinise.
Small operators carry double-digit rates and chips that lose value faster than the loans amortise. Derivatives can smooth cashflow; they can't fix a balance sheet. Watch where the stress shows up first.
Computers that train and run AI are being rented out like hotel rooms. Right now there's no price board: big players cut private deals, and everyone else haggles through middlemen in group chats.
Bottom line: the instruments are arriving before the referee, and the unit they settle on is still being defined. That is where the governance work is.