Touchmark is making AI tokens a tradable commodity.
A company that already knows it will spend heavily on AI in September can pay for that month now and get their tokens at a discount of up to 30% – or more for larger orders via direct quotes. On Touchmark, a forwards market which opens to commercial buyers and sellers this week, a billion tokens on a named model is a single prepaid contract with a delivery window attached, drawn down through an API when the window arrives.
Touchmark is a two-person company, part of Y Combinator’s Summer 2026 batch. Its premise is that inference is becoming a commodity, and commodities end up with markets.
Touchmark is addressing a significant budgeting challenge many companies are facing. Enterprise AI spending roughly doubled year over year in 2026, to an average of about $1.2 million per organization, and 78 percent of IT leaders reported charges they had not budgeted for. Agents made it worse. A single autonomous workflow can consume many times the tokens of a one-shot query, quietly, over hours. Finance teams comfortable forecasting seat licenses have spent the year discovering that they cannot forecast this.
Touchmark allows a buyer to pick a model, a token volume, a preferred provider, and a delivery window, pays in full up front, and consumes the capacity through an API like any other endpoint. The earlier the purchase, the deeper the discount. Nothing in it asks the buyer to think about anything beyond what it will use and when.
Next to the listed contracts sits a request-for-quote board, live at launch. A buyer posts exactly what it needs, model or models, volume, delivery window, throughput and latency floors, denominated in tokens, dedicated GPUs, or reserved throughput. Providers then compete to quote against it. Supply skews toward open-weight families: Kimi, GLM, Qwen. That follows the market rather than leading it. OpenRouter’s routing data showed Chinese open-weight models crossing into a majority of the tokens it processes by the middle of this year, most of them priced well beneath the U.S. frontier.
Touchmark’s first inference provider, Wafer, is recognized as one of the leading inference providers by throughput. For a provider, the appeal is timing: revenue today against capacity that will not be served for months. Capacity sits underused in one month and runs short in the next, and the industry’s current answer to that gap is a series of private one-off deals.
“Inference is becoming one of the largest costs for companies today, but the market is still extremely fragmented. Pricing for reserved capacity is bilateral and buyers have very little visibility into what a fair forward price actually is,” said Ilia Bolgov, Touchmark’s co-founder and CEO. “As inference spend grows, companies need the same basic instruments every large market eventually develops: price discovery, forward purchasing, and the ability to resell unused capacity. That’s the market we’re building at Touchmark.“
Bolgov came to the problem from the quantitative finance side. He graduated first in his cohort in mathematics at Cardiff University before an MSc in mathematics and finance at Imperial College London, then worked in product at Revolut’s wealth and trading division, building advanced products for advanced traders. That combination of mathematical training and hands-on experience building products inside a fast-growing scale-up shaped his approach to Touchmark: treating market design as both a mathematical problem and a product problem.His co-founder, Roman Yanushevskyi, built the platform. Yanushevskyi took gold at the 2022 International Olympiad in Informatics held in Indonesia, finishing 7th out of 349 competitors from 88 countries. He spent two stints as a quantitative research intern at Citadel Securities on the options team, and before that spent time at Lovable in 2024, during the company’s early days as it grew out of the open-source GPT Engineer project, working on its evaluation and benchmarking pipeline. Lovable has since been valued at $13.3 billion. Roman finished a four-year computer science and mathematics degree at the National University of Singapore in three years.
“Providers selling spot capacity can’t know how much demand they’ll have tomorrow,” Yanushevskyi said. “Little demand means wasted GPU hours and losses. Too much demand leads to unavailability for some customers, who may therefore churn. The current solution is committed-capacity deals. However, all such deals today are bespoke and negotiated case by case. No convenience and no transparency for either side. No one knows the correct price. The finance industry, however, solves all these problems at scale.”
They are early to a concept that has rapidly gained momentum. On May 12, CME Group and Silicon Data said they would bring compute futures to market, and a week later ICE and Ornn announced GPU compute futures referencing Ornn’s Compute Price Index. This Monday, Jensen Huang pitched AI compute to investors as a new asset class. Those instruments track the cost of renting hardware. Touchmark sells the served output of that hardware, in the units a buyer actually consumes. Inference providers, for their part, already offer reserved capacity and committed-use discounts, each on its own paper, to its own customers, at prices no one else gets to see.
The idea is that a standardized contract with providers quoting against each other beats negotiating every deal privately. The obvious risk is the models. Buying six months of capacity on a named model assumes it will still be the right model in six months, in a year when the open-weight leaderboard has turned over repeatedly. Touchmark’s answer is contracts for best-in-class capacity that resolves into whichever model tops a public benchmark at delivery time.
Once something gets traded on a market, it usually ends up with a price everyone can quote. Touchmark has been open about expecting the same for intelligence. Eventually there will be a published price index for it, but that’s a long way off. For now, the only question on Friday’s agenda is what a September token is worth.
