DeepSeek-Quality Inference Is Hard — and "99% Cache Rates" Are a Red Flag

Key Info

A builder at Anomaly says matching DeepSeek's inference quality has taken his team months and is extremely hard — roughly a $100M budget, the right connections, and about 10 world-class experts. He also warns that inference providers advertising "99% cache rates" are almost certainly wrapping DeepSeek, the only provider currently hitting that number, and may later quietly switch to another backend while keeping the claim.

Highlights

• Reproducing DeepSeek's inference quality is "difficult": months of work, ~$100M in budget, key connections, and ~10 top experts — so anyone claiming easy replication is likely lying.
• A "99% cache rate" claim is a red flag: DeepSeek is said to be the only provider that actually achieves it today, meaning such providers are probably reselling DeepSeek.
• If a provider later swaps backends, they may still advertise the same cache performance, and users often can't easily verify the real cache hit rate.