The idea that Chinese models cost less to train seems to be based on that one time DeepSeek estimated the training cost for their V3 model at GPU rental rates as $5 million, and comparing this to other companies' entire R&D budgets. Yet DeepSeek raised $7 billion of fresh money last month, enough to train more than 1000 such models. What gives?

- You need to train lots of experimental models to dial in the training process just right for the one model that actually gets released in the end. Fortunately, these can be smaller.

- However, everyone is training much bigger models now, and doing a lot of RL rollouts on top.

- You can't get the GPUs for this piecemeal at rental rates because they need to be wired together using high-bandwidth interconnects.

- Nvidia GPUs are much more expensive in China, and local alternatives are still immature and not as efficient. Some companies have gotten around this using data centers in Singapore, which should tell you that electricity prices are not the primary consideration.

- The one line item where Chinese companies can probably save quite a bit of money is salaries for rank-and-file researchers.

In any case, they need to make back that money somehow. Giving away freebies isn't going to cut it.