Touchmark is making AI tokens a tradable commodity.
An organization that already is aware of it would spend closely on AI in September pays for that month now and get their tokens at a reduction of as much as 30% – or extra for bigger orders through direct quotes. On Touchmark, a forwards market which opens to industrial consumers and sellers this week, a billion tokens on a named mannequin is a single pay as you go contract with a supply window connected, drawn down by means of an API when the window arrives.
Touchmark is a two-person firm, a part of Y Combinator’s Summer time 2026 batch. Its premise is that inference is changing into a commodity, and commodities find yourself with markets.
Touchmark is addressing a major budgeting problem many firms are going through. Enterprise AI spending roughly doubled 12 months over 12 months in 2026, to a mean of about $1.2 million per group, and 78 % of IT leaders reported fees that they had not budgeted for. Brokers made it worse. A single autonomous workflow can eat many occasions the tokens of a one-shot question, quietly, over hours. Finance groups snug forecasting seat licenses have spent the 12 months discovering that they can’t forecast this.
Touchmark permits a purchaser to choose a mannequin, a token quantity, a most well-liked supplier, and a supply window, pays in full up entrance, and consumes the capability by means of an API like another endpoint. The sooner the acquisition, the deeper the low cost. Nothing in it asks the customer to consider something past what it would use and when.
Subsequent to the listed contracts sits a request-for-quote board, dwell at launch. A purchaser posts precisely what it wants, mannequin or fashions, quantity, supply window, throughput and latency flooring, denominated in tokens, devoted GPUs, or reserved throughput. Suppliers then compete to cite in opposition to it. Provide skews towards open-weight households: Kimi, GLM, Qwen. That follows the market reasonably than main it. OpenRouter’s routing knowledge confirmed Chinese language open-weight fashions crossing right into a majority of the tokens it processes by the center of this 12 months, most of them priced nicely beneath the U.S. frontier.
Touchmark’s first inference supplier, Wafer, is acknowledged as one of many main inference suppliers by throughput. For a supplier, the enchantment is timing: income right now in opposition to capability that won’t be served for months. Capability sits underused in a single month and runs quick within the subsequent, and the business’s present reply to that hole is a collection of personal one-off offers.
“Inference is changing into one of many largest prices for firms right now, however the market remains to be extraordinarily fragmented. Pricing for reserved capability is bilateral and consumers have little or no visibility into what a good ahead worth really is,” mentioned Ilia Bolgov, Touchmark’s co-founder and CEO. “As inference spend grows, firms want the identical fundamental devices each giant market finally develops: worth discovery, ahead buying, and the flexibility to resell unused capability. That’s the market we’re constructing at Touchmark.“
Bolgov got here to the issue from the quantitative finance facet. He graduated first in his cohort in arithmetic at Cardiff College earlier than an MSc in arithmetic and finance at Imperial Faculty London, then labored in product at Revolut’s wealth and buying and selling division, constructing superior merchandise for superior merchants. That mixture of mathematical coaching and hands-on expertise constructing merchandise inside a fast-growing scale-up formed his method to Touchmark: treating market design as each a mathematical drawback and a product drawback.His co-founder, Roman Yanushevskyi, constructed the platform. Yanushevskyi took gold on the 2022 Worldwide Olympiad in Informatics held in Indonesia, ending seventh out of 349 opponents from 88 nations. He spent two stints as a quantitative analysis intern at Citadel Securities on the choices workforce, and earlier than that frolicked at Lovable in 2024, through the firm’s early days because it grew out of the open-source GPT Engineer undertaking, engaged on its analysis and benchmarking pipeline. Lovable has since been valued at $13.3 billion. Roman completed a four-year laptop science and arithmetic diploma on the Nationwide College of Singapore in three years.
“Suppliers promoting spot capability can’t understand how a lot demand they’ll have tomorrow,” Yanushevskyi mentioned. “Little demand means wasted GPU hours and losses. An excessive amount of demand results in unavailability for some clients, who might subsequently churn. The present resolution is committed-capacity offers. Nevertheless, all such offers right now are bespoke and negotiated case by case. No comfort and no transparency for both facet. Nobody is aware of the proper worth. The finance business, nevertheless, solves all these issues at scale.”
They’re early to an idea that has quickly gained momentum. On Could 12, CME Group and Silicon Information mentioned they’d carry compute futures to market, and per week later ICE and Ornn introduced GPU compute futures referencing Ornn’s Compute Worth Index. This Monday, Jensen Huang pitched AI compute to buyers as a brand new asset class. These devices observe the price of renting {hardware}. Touchmark sells the served output of that {hardware}, within the items a purchaser really consumes. Inference suppliers, for his or her half, already provide reserved capability and committed-use reductions, every by itself paper, to its personal clients, at costs nobody else will get to see.
The concept is {that a} standardized contract with suppliers quoting in opposition to one another beats negotiating each deal privately. The apparent threat is the fashions. Shopping for six months of capability on a named mannequin assumes it would nonetheless be the suitable mannequin in six months, in a 12 months when the open-weight leaderboard has turned over repeatedly. Touchmark’s reply is contracts for best-in-class capability that resolves into whichever mannequin tops a public benchmark at supply time.
As soon as one thing will get traded on a market, it normally finally ends up with a worth everybody can quote. Touchmark has been open about anticipating the identical for intelligence. Ultimately there will likely be a broadcast worth index for it, however that’s a great distance off. For now, the one query on Friday’s agenda is what a September token is value.
