Applied Compute raises $350M at $3.25B valuation after reaching $100M ARR

Applied Compute said it raised $350 million at a $3.25 billion valuation. It also said it reached $100 million in annualized revenue run rate less than 18 months after being founded. The company announced the deal on Monday, November 2, and its own post says Elad Gil led the round.

The post lists Kleiner Perkins, Lux Capital, Hanabi Capital, NVIDIA, Blackstone, Shopify Ventures and HubSpot Ventures among participants, followed by a longer list of other investors. SiliconANGLE, which covered the deal, described Nvidia, Shopify Ventures and HubSpot Ventures as participants alongside Gil and said more than a dozen others also took part. The company's product is a cloud where customers train their own large language models and then run them.

The technical pitch rests on how training is scheduled. SiliconANGLE explained that a typical LLM training stack has two parts. One generates multiple copies of the model working on different tasks and turns the results into training data. The other uses that data to adjust the model's weights. Those parts usually run synchronously, so each waits for the other to finish. Applied Compute's platform uses asynchronous reinforcement learning, which lets the two run at the same time. It also supports self-distillation, which removes the need for teacher models, the specialized networks researchers use to supervise training runs.

On the inference side, SiliconANGLE describes an autoscaling engine that adjusts how much hardware a model gets based on user traffic. It also describes an online RL feature that uses the prompts a model receives to fine-tune its weights periodically. A model router called Shift-1 sends each request to whichever LLM can handle it most cheaply, and can split complex tasks across several models. SiliconANGLE said that is aimed at coding agents. It also reports that the company's researchers can help customers with work such as building quality evaluations for new models.

The usage figures come from the company. Its post says customers have run more than 1,700 training runs on its platform since July. It says its AC Inference product serves more than 1 trillion tokens a day, and that 40% of customers use it for both training and inference. The team has tripled since January, and 85% of staff are researchers and engineers. The two case studies are also company-reported. For Harvey, it says it 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. For Clay, it says a custom router reduced production inference costs by up to 47% without sacrificing quality. The post does not name the frontier model used as the comparison or explain how quality was measured. It names Microsoft, Shopify, Mercor, Harvey, DoorDash, Cognition, Clay and Decagon as customers, but does not say how the $100 million run rate is split among them or how the company calculates run rate. SiliconANGLE used the term annualized recurring revenue for the same figure.

The company says the funding will go toward expanding compute capacity, growing its team and accelerating enterprise deployment. In a blog post, founders Rhythm Garg, Yash Patil and Linden Li wrote that "Training, inference, and routing are the first pieces of a much larger stack," and described building the infrastructure companies will need, "from the models themselves to the systems that run, monitor and secure them."