Applied Compute raises $350 million at a $3.25 billion valuation to scale custom AI

Applied Compute has raised $350 million at a $3.25 billion valuation to build out its business helping companies train and run tailored AI. The Elad Gil-led financing comes as the startup says it reached a $100 million annualized revenue run rate less than 18 months after its founding, pairing fresh capital with a company-reported growth milestone.

Participants included Nvidia, Kleiner Perkins, Lux Capital, Blackstone, Shopify Ventures and HubSpot Ventures. Latham & Watkins, which advised Applied Compute on the transaction, confirmed the financing terms in its November 3 announcement. The law firm represented the startup; Gil led the investment round.

Revenue pace and a wider customer base

The company disclosed the funding on November 2 alongside its operating figures. Its own announcement describes the $100 million milestone as an annualized revenue run rate, rather than revenue collected over a completed year. That wording also differs from SiliconANGLE’s description of the figure as annualized recurring revenue: Applied Compute’s announcement does not label it recurring.

Applied Compute says it has more than doubled the number of enterprises it serves since its previous fundraise. It names Microsoft, Shopify, Mercor, Harvey, DoorDash, Cognition, Clay and Decagon as customers. The disclosed customer base spans life sciences, cybersecurity, hardware design, legal work, financial services and logistics—not just companies selling AI software.

Tailored models, not just rented computing

Applied Compute’s pitch starts with open-weight models, which companies can build on using their own data and expertise. It then combines model training with deployment and ongoing improvement. AC2, its Agent Cloud, was introduced three months before the financing announcement. It is an existing platform supporting the funding case, not a new product launched with this round.

The service includes researchers who help customers with tasks such as designing quality evaluations. SiliconANGLE describes a software stack covering three related jobs: improving models, operating them once trained, and choosing which model should handle a request. Those jobs connect the development work to the cost of serving customers.

• Training: The platform uses asynchronous reinforcement learning, allowing the generation of training data and adjustments to a model’s internal settings to happen at the same time. Neither step has to wait for the other to finish.
• Running models: Its inference service—the system that runs trained models to produce answers—automatically adjusts hardware allocation as traffic changes. An online learning feature uses incoming prompts to periodically refine the model.
• Routing requests: Shift-1 directs coding-agent requests to the model it selects as most cost-efficient for the task. It can also divide a complex task among multiple models, rather than assign the whole job to one.

Two customer examples in the funding announcement illustrate different routes to savings: replacing a frontier alternative with a tailored model, and choosing among models more efficiently. These are Applied Compute’s performance claims, not independently verified comparisons or savings guaranteed for every customer.

Funding more capacity and enterprise deployments

The company says customers have completed more than 1,700 training runs on AC2 since July, while AC Inference serves more than 1 trillion tokens a day. It also says 40% of customers use both training and inference, showing how much of its customer base already buys the two services together.

Applied Compute plans to spend the new funding on expanded computing capacity, a larger team and faster enterprise deployment. Its team has already tripled since January, and researchers and engineers account for 85% of staff, according to the announcement. The financing backs further growth in both the infrastructure and the people who help customers use it.