Applied Compute: $100M ARR in 18 months, $350M at $3.25B
The funding cycle usually hands you a narrative. On Monday it handed over a number instead: Applied Compute says it crossed a $100 million annualized revenue run rate less than 18 months after it was founded, and it raised $350 million at a $3.25 billion valuation on the strength of it.
Applied Compute closed $350 million at a $3.25 billion valuation, led by Elad Gil, with Nvidia, Blackstone, Kleiner Perkins, Lux Capital, Shopify Ventures and HubSpot Ventures among a long participant list, according to the company's own announcement. SiliconANGLE reported the round the same day. The arithmetic is blunt: $3.25 billion on $100 million of annualized revenue is roughly 32.5 times. That is a growth multiple, not a margin multiple, and it only clears if the curve keeps bending.
The company is not another GPU cloud renting out H100 hours by the minute. Applied Compute builds and runs custom models for enterprises on top of open-weight foundations — banks, law firms, logistics operators and hardware designers that want a model tuned to their own data rather than a general one rented from a frontier lab. Founders Rhythm Garg, Yash Patil and Linden Li came out of OpenAI and Stanford, took $20 million from Benchmark's Victor Lazarte in mid-2025 at a $100 million valuation, and were valued at $1.3 billion by their next round, $80 million led by Kleiner Perkins.
The technical bet is about removing the waiting. Standard reinforcement-learning training runs in two halves that take turns: one generates attempts, the other updates the model's weights, and each sits idle while the other works. Applied Compute's platform runs them asynchronously so both halves work at once, and it uses self-distillation to cut out the separate "teacher" model that normally supervises a training run — two changes that mostly buy speed, which in this business is the same thing as money. A built-in assistant summarizes what a model picked up during a run, work that otherwise means a researcher reading reasoning logs for hours.
Those pieces sit inside AC2, the agent cloud the company launched about three months ago, alongside autoscaling for inference and a router called Shift-1 that sends each request to whichever model handles it most cheaply and splits complex jobs across several models. Since July, customers have run more than 1,700 training runs on the platform, and inference now serves over 1 trillion tokens a day.
The customer list is the actual evidence, and two case studies carry it. For Harvey, the legal AI company, Applied Compute trained a model for its Review Table product that improved answer and citation quality while costing less than half what a frontier alternative charged at inference, per the company's own account. For Clay, it trained and deployed a router that cut production inference costs by up to 47% without a quality loss. Microsoft, Shopify, Mercor, DoorDash, Cognition and Decagon are named as customers, across life sciences, cybersecurity, legal, financial services and logistics. Forty percent of customers now buy both training and inference, which is the number that matters commercially — it is the difference between a project and a dependency.
This is the third price tag in 17 months, and the ladder is the story as much as the round: $100 million, then $1.3 billion, now $3.25 billion. We have watched this pattern repeatedly this year — Factory raised $200M at a $5B valuation, tripling in five months — and the pattern is not that the companies are overpriced. It is that enterprise AI budgets are being reallocated faster than valuation models can track, and the winners of that reallocation get repriced every few months.
The counter-argument is unglamorous and has not gone away: a run rate is not revenue, and revenue quality is not disclosed. The company's own wording is "annualized revenue run rate," which means the latest month multiplied by twelve, not a signed book of annual contracts. A large slice of inference revenue is compute the company buys and resells, so gross margin could sit far below what a 32.5x multiple implies. There is no churn figure, no average contract length, no net revenue retention, and no margin — the three or four numbers a public-market investor would demand first. Add the structural risk: if frontier API prices keep falling, "own your intelligence" gets harder to justify for every customer whose differentiation is thin, and Applied Compute's pitch is partly a hedge against exactly that price decline.
The strategic investors are worth a second look. Nvidia taking a stake in a company whose business is training and serving open-weight models is not charity — every training run on AC2 is demand for accelerators, and a stake in the layer that decides which chips get bought is cheaper than advertising. Blackstone's presence points the other way, toward the infrastructure financing that has become the constraint on every AI company's growth plans.
What to watch: whether the $100 million run rate survives the quarter, since run-rate figures have a habit of being the best month rather than a trend; whether the company ever discloses gross margin; and whether $3.25 billion becomes the comp that the next dozen AI-infrastructure rounds get priced against. If enterprise buyers keep paying a premium to own a slightly smaller model, this valuation looks early rather than rich. If they drift back to renting frontier models as prices fall, it looks like the top.
Would you rather rent a frontier model or own a slightly worse one that costs half as much to run? Write to us via our contact page.
