Airlines buy next year's jet fuel at today's price. Bakers do it with wheat. Nobody does it with AI tokens, and token bills are now big enough that somebody should.

What they built

Touchmark runs a market for future inference capacity. You buy AI model tokens ahead of time, at 10 to 30 per cent below on-demand pricing, with the price fixed today. If your plans change and you do not use them, you can sell what is left to another buyer.

Ilia Bolgov, co-founder and chief executive, read mathematics and finance at Imperial College London and was a product manager at Revolut in its wealth and trading division. Roman Yanushevskyi, co-founder and chief technology officer, studied computer science and mathematics at the National University of Singapore, won a gold medal at the International Olympiad in Informatics, did two quantitative research internships at Citadel Securities, and was an AI engineering intern at Lovable. The company is based in San Francisco.

Plenty of companies sit between you and compute. OpenRouter routes a request to whichever model answers it best. Together AI, Fireworks AI, Baseten and Modal sell inference itself. SF Compute built a market for raw GPU time, and CoreWeave and Lambda rent the machines. Stoa is assembling a reference price for the hardware and Princeps is writing insurance for it. Every one of those sells you capacity now. Touchmark sells you a claim on capacity later, which is a different kind of thing entirely.

The point: every financial instrument we have is about to be pointed at AI

AI is changing what a company's finances look like. A line that did not exist three years ago is now one of the larger ones, and it moves. So it makes complete sense that the old financial vehicles are being picked up and applied to the new components.

Forwards, futures, options, insurance, leasing, securitisation, indices, hedges, spot and settlement. Every single one of those was invented for a physical thing that was expensive and unpredictable, and every one of them has an obvious AI version that nobody has built yet.

This is a golden window, and it looks a great deal like the early crypto years, because the financial work can currently be done by a purely technical startup.

That is the part worth stopping on. A forward contract on jet fuel is a regulated derivative. You need licences, clearing, capital requirements and lawyers, and two people in their twenties cannot simply open one. A forward contract on API tokens is, for now, a commercial agreement about buying software. Same instrument, same economics, and an entirely different amount of permission required to build it.

So the opportunity is open to builders rather than to banks, and it is open right now.

How long the window stays open

It is worth saying plainly what "for now" means, and the crypto comparison is the reason.

Early crypto also felt like a space without rules, and the rules arrived later and were applied backwards. Plenty of people who thought they were shipping a purely technical product discovered they had been running a financial venue the whole time. The rules did not fail to exist, they simply had not turned up yet.

The honest version is that the window is exactly as long as it takes the regulators to notice, and nobody gets to know how long that is. The design decisions that matter most are the ones that determine which side of the line you land on when they do arrive. Selling a customer capacity they intend to use is buying software. Running a venue where strangers trade contracts with each other to make money starts to look like something else.

That is the real risk here, and it is a larger one than whether token prices go up or down.

Try every instrument, and most will die quietly

This part is opinion rather than anything the company has said.

The exercise is almost mechanical, and that is what makes it a good one. Take each financial instrument that exists, point it at an AI component, and see whether anything happens.

Futures on inference. Options on training capacity. Insurance on model uptime. Leasing for GPU fleets. Securitising data centre revenue. A published index for the price of a million tokens. Escrow for work an agent completed. Factoring for AI-native firms waiting to be paid. Settlement between agents transacting with each other. Credit underwritten against a company's model spend.

Most of these will die without making a ripple, because the need was not really there and a spreadsheet was already good enough. A few will be enormous. The useful thing is that you cannot reliably tell which from the outside, and building one is cheap compared with the upside if it lands.

Either way it is worth the gamble for a founder shooting at the moon, and there are more of these ideas than there are people currently trying them.

Why it is bigger than it sounds

The short description is a discount on tokens. What it belongs to is the arrival of finance in the AI stack.

  • The cost is now large enough to hedge. Token spend has become a real line item that moves, which is the condition under which a forward market has always appeared.
  • A technical startup can do this today. The same instrument in fuel or wheat needs licences and capital, and that asymmetry is the whole opportunity.
  • The resale clause is what makes it a market. Without it you have a discount voucher. With it you have something that can develop a price other people quote.

What to watch

The measure is whether anybody resells. A contract that is only ever bought and consumed is a bulk discount, and the second a stranger buys one from another stranger it is a market.

The second is who the sellers are. Selling forward capacity means someone is short, and whether that is Touchmark, the model providers, or other customers decides what kind of company this becomes.

The third is the first letter from a regulator, to anyone in this category. That will not be the end of it, but it will be the moment the window stops being wide open, and everything built before then gets judged by rules written afterwards.