Applied Compute moving from training to serving

This reveals that Applied Compute is not really selling a one time model project, it is trying to land a customer with training work and then keep the account through the much steadier, recurring spend of running that model in production. Once a customer has a trained checkpoint on AC2, that checkpoint can sit behind the same endpoint, handle live traffic, and send real production traces back into the next tuning cycle, which makes serving the part of the workflow that can compound revenue over time.1 2

• Applied Compute said on November 2, 2026 that it was serving more than 1 trillion tokens per day, that 40% of customers were already using both training and inference, and that customers had run more than 1,700 training jobs on AC2 since July. That shows the company is already converting project work into ongoing usage revenue.1
• The product is built to make that handoff easy. A customer brings its data, grading logic, and agent harness, then Applied Compute trains inside that existing setup, deploys the best checkpoint as a live endpoint, and swaps in updated checkpoints behind the same address. That keeps the customer from rebuilding surrounding application code every time the model improves.1
• This is also the core competitive line against Together AI and Fireworks. Those platforms already separate training charges from ongoing hosting or dedicated deployment charges, and Fireworks explicitly ties training and deployment together. Applied Compute is differentiating less on the existence of the loop, and more on doing the setup work with researchers inside the customer’s actual agent workflow.1 3 4 5

The next step is for Applied Compute to become not just the place where custom models are created, but the place where production AI traffic lives. If more customers adopt AC Inference, routing, and customer cloud deployments, the business shifts toward a denser base of recurring infrastructure revenue, with training acting as the wedge that opens the door.1 2