Applied Compute moves into production serving

Applied Compute is moving from project revenue to infrastructure revenue. Training a custom model can be a one time engagement, but serving 1 trillion tokens per day means customers keep sending live production traffic through Applied Compute after the model is built. That turns the company from a specialist that helps tune models into part of the runtime stack, where usage grows with customer product adoption and creates a steadier, larger revenue base.1 2 3

• The clearest signal is product mix. Applied Compute said 40% of customers use both training and inference, and customers ran more than 1,700 AC2 training jobs since July. That shows model building is increasingly connected to hosting and operating those models after launch, instead of ending at handoff.1
• This changes who Applied Compute competes with. Once a model is in production, alternatives are not just custom model shops, but serving platforms like Baseten, Modal, and Replicate that make money every time developers call a model API or run GPU workloads.2 3 4 5
• Production serving also gives Applied Compute more leverage inside customers. In Clay’s case, it post-trained and served a routing model on AC2 that decides which model handles each query. That is deeper than supplying a model checkpoint, it means Applied Compute sits in the request path and can shape cost, latency, and quality every day.1

The next step is a fuller stack where training, routing, evaluation, and inference are sold together as one operating layer for enterprise AI apps. If more customers follow the training to serving path, Applied Compute can compound revenue per account and become harder to replace, because ripping it out would mean retraining models and reworking production traffic at the same time.1 2 3