Who Gets Paid For Model Adaptation

This is really a fight over who gets paid for the hard part of making open models useful in production. Together and Fireworks try to turn customization into software, where a team uploads data, starts jobs from an API or console, and pays by tokens or GPU time. Applied Compute instead sells a staffed project with reserved capacity, where its researchers write evals, train inside the customer’s existing agent harness, then keep serving the resulting model as a recurring workload.1 2 3 4 5

• Together’s model is closest to a cloud utility. It offers self serve fine-tuning, dedicated endpoints, and post-training billed per token or GPU minute, which fits startups and internal ML teams that already know how to run experiments and want to avoid paying for idle clusters.1 3 5
• Fireworks pushes the same utility logic further into production inference. Hebbia used it mainly because new open models showed up fast behind an OpenAI style API, with latency, observability, autoscaling, and concurrency handled by Fireworks instead of the customer’s infrastructure team.1 4 5
• Applied Compute is betting that many enterprises do not just need cheaper GPU access. They need someone to turn a business metric into a training loop. That is why its embedded path includes forward deployed researchers, custom graders, and dedicated Blackwell capacity, and why 40% of customers already use both training and inference.1

Over time these models will converge from both sides. Utility platforms will add more enterprise help, while research led vendors will productize more of their workflow into self serve tools. The durable winners will be the ones that can start as a simple API purchase, then expand into the higher value work of adaptation, deployment, and ongoing model operations.1 4 5