dax says matching DeepSeek inference is a $100M problem
He also says any provider advertising a 99% cache rate is wrapping DeepSeek, because DeepSeek is the only one hitting that number.
dax says his team has been working for months to reproduce DeepSeek's quality of inference, and that it is not a matter of effort or tuning. He puts the cost of doing it at a $100M budget plus the right connections, plus getting the right ten people in the world to help you. His conclusion about everyone else advertising the same thing is blunt, anyone claiming they can do this is lying.
if you ever see anyone claiming they can do this, they are lying
That sits on top of a second claim from him about how you would spot the gap from the outside. dax says he keeps seeing inference providers claiming 99% cache rates, and that DeepSeek is the only provider in the world hitting that number right now. If a provider is quoting it, his read is that they are wrapping DeepSeek. He adds the part that actually bites, that they later switch to a different provider and keep claiming the same number.
the only provider in the world that hits that right now is deepseek. so this means they are wrapping deepseek
What this is and is not
This is one operator's account of one long attempt, not a benchmark and not a teardown of a named competitor. dax does not name any of the providers he is describing, and he does not show numbers from his own attempt. Treat it as a report from someone who has been paying the bill to try, which is a different thing from a measurement anyone else can reproduce.
For working engineers, the useful part is what a cache hit rate is actually describing. It is the share of your prompt tokens served from a previously computed prefix rather than recomputed, which is why it shows up in pricing pages, a high rate makes long repeated system prompts and agent loops much cheaper. dax's point is that the number is a property of the stack that serves the request, so a reseller can quote it honestly on day one and dishonestly on day thirty without changing the marketing page.
If you are choosing an inference provider on price and cache marketing for agent workloads, the takeaway is to verify against your own traffic rather than the number on the page. Measure your own latency and token spend on your own prompts over time, and watch for the moment the numbers move without an announcement. Nobody in the sources has published a provider by provider comparison, so there is no list to check against yet.