The next AI revenue layer is customization, not inference
Ramin Hasani argues that strong open source base models are eroding the value of raw inference, and that a new category of companies will capture enterprise AI spending at the customization layer. The bet is specific enough to track.
The dominant revenue model in AI today is selling inference: frontier labs and inference providers host models and charge per token consumed. Ramin Hasani, of Liquid AI, thinks that model is already losing its grip. His argument is that as strong open source base models proliferate, the marginal value of raw inference falls, and the money migrates to the layer above it: customization.
Hasani names the companies he sees building on that thesis. Rei, Core Automation, Thinking Machines, and Trajectory are among the names he puts in this category. Each, by his framing, is pursuing what he calls customization tokens: the compute and data work that adapts a base model for specific enterprise use. That is a materially different business than selling access to a hosted model. It is also a more defensible one, in the sense that a well-customized model is harder for an enterprise to swap out than a commodity inference endpoint.
The underlying logic turns on a supply shift. When base open source models were weak, enterprises had strong reasons to pay frontier prices for frontier capability. As open source quality rises, that calculus changes. An enterprise that can start from a capable open model and pay a customization company to adapt it for a specific domain faces a lower switching cost than one locked into a proprietary API. Hasani’s prediction is that this dynamic redirects a meaningful share of AI spending toward the customization layer.
A lot of people are turning into customization tokens because now there's a lot of good base open source models as well.Ramin Hasani
The call has a structural implication worth stating plainly. If customization tokens become the primary monetization surface, the companies that win will need differentiated expertise in training pipelines, domain-specific data, and fine-tuning work, rather than in model architecture or inference infrastructure. That is a different kind of moat, and it favors a different kind of team.
There is a real counter-argument that Hasani’s framing does not resolve. The same open source proliferation that creates demand for customization also makes it easier for enterprises to run customization work in-house. A capable engineering team with access to strong base models can, in principle, handle adaptation and fine-tuning without hiring a specialist vendor. The companies Hasani names will need to demonstrate that their customization tokens deliver returns that a do-it-yourself approach cannot match. That is not a given, and the competitive pressure is not only from other customization vendors but from the enterprises themselves.
The horizon on this call is not pinned to a specific year, which makes it harder to grade cleanly. What can be tracked is whether the companies Hasani identifies build durable revenue, whether enterprise AI procurement shifts visibly toward training-time services rather than inference-time consumption, and whether open source base model quality continues to close the gap with proprietary frontier models. Those are observable outcomes. If customization-focused vendors like the ones Hasani names are generating substantial enterprise revenue by the time open source quality reaches rough parity with today’s frontier, the call has landed. If inference providers manage to maintain pricing power despite the open source pressure, it has not. The bet is specific enough that the evidence, when it arrives, will be hard to misread.