Applied Compute Reaches a $100 Million Run Rate in Under 18 Months, Now Valued at $3.25 Billion

Applied Compute, which trains and runs custom models for companies, said it passed a $100 million run rate in less than a year and a half, as Elad Gil led a round valuing it at $3.25 billion, with Nvidia and Shopify Ventures investing.

Kleiner Perkins also took part. The money goes to compute capacity, hiring and enterprise deployment, and the company has more than doubled the number of enterprises it serves since its last fundraise.

The figure that matters for a buyer is Clay’s. Applied Compute says a custom router it built cut Clay’s production inference costs by up to 47% without sacrificing quality. Up to 47% off the per-token bill on routed workloads, with quality held, is the size of saving that reopens the build-or-buy question against a frontier API. Applied Compute did not say what Clay spent before the router or which model it replaced, so the figure cannot be turned into dollars.

Who is buying?

Microsoft and Shopify are customers, along with Harvey, DoorDash and Cognition. Applied Compute says its customers now span industries including life sciences, cybersecurity, hardware design, legal, financial services and logistics, so the demand is not confined to one industry.

What else has it shown?

The same pattern appears at Harvey, where Applied Compute trained a model for the Review Table product that improved answer and citation quality at less than half the inference cost of a frontier alternative. That result covers one product, not Harvey’s whole workload.

Both examples are company-reported, and Applied Compute did not say how it scored quality or which frontier model it compared against.

Is the usage there?

Applied Compute’s inference service, AC Inference, now serves more than 1 trillion tokens a day. Customers have run more than 1,700 training runs on its AC2 platform, which brings training, inference and continual improvement together, since July.

About 40% of customers use Applied Compute for both training and inference, and the company says that share is growing.

Which of the company’s own workloads could run on a custom model trained on its own data, and what does each now cost on a frontier API?