Open Weights Limit Pricing Power

The core issue is that open weights make the model portable, so the part customers pay for has to be the hard operational work around the model, not the model alone. Tinker runs training on Thinking Machines clusters, handles scheduling and failures, then also offers inference for saved checkpoints. That means revenue depends on charging for GPU heavy training, sampling, and storage faster than those infrastructure costs rise, while proving enough convenience and performance that customers keep using the hosted stack instead of exporting weights elsewhere.1 2 3 4

• Tinker is designed to sell managed post training, not just access to a base model. Users write their own training loop and call four functions for gradients, optimizer steps, sampling, and checkpoints, while Thinking Machines handles the distributed systems layer underneath. That is valuable work, but also an expensive service business built on GPUs.2 3 4
• Portability caps pricing power. Tinker documentation explicitly says trained weights can be downloaded and used with another inference provider. In practice that means a customer can use Tinker for customization, then shift serving to a cheaper host or internal cluster, leaving Thinking Machines to recover economics mainly during the training phase.4
• Competitors are attacking the same budget from different angles. Fireworks combines training with high performance serving and multi LoRA deployment. AWS Bedrock packages model customization inside an existing enterprise cloud relationship. Modal offers compatible post training primitives around a broader compute platform. Those alternatives make it harder to charge large premiums for hosted customization alone.5 6 7 8

The next step is a race to bundle more of the workflow around the weights. The providers that win are likely to be the ones that turn training, evals, inference, and monitoring into one smooth system, so leaving the platform feels slower and riskier than staying, even when the model itself can move.1 3 5 7