That sounds incredibly capital intensive though. How are they funding the fab costs?

By leveraging the 'distillation' angle that Y Combinator’s Garry Tan has been pushing hard lately. The thesis is simple: if you can compress a massive frontier model into a smaller, efficient open-weight version that runs on custom hardware, your margin profile looks nothing like an API reseller. You look like a hardware company with high margins.

So essentially they are skipping the training race entirely and winning on inference efficiency.

Exactly. Training is a winner-take-all game where you need fifty thousand H100s sitting in a warehouse.

That makes me think about the recent news regarding Anthropic and distillation campaigns. Are they doing the same thing?

Similar logic, different angle. Distillation requires massive compute to train smaller models efficiently. If you have idle capacity, you can run those distillation campaigns cheaper and faster. It creates a flywheel where your infrastructure efficiency improves your product quality, which drives more users, which justifies keeping the infrastructure running.

It feels like the industry is finally admitting that software eats the world, but hardware feeds the software.

That's a good way to put it. For years, we treated hardware as a commodity.