Two unrelated DeepSeek stories crossed the wire this week. One is about software: DeepSeek and Huawei released open-source programming tools for Huawei's Ascend AI chips, including the DeepGEMM-Ascend compute library and DeepEP-Ascend communication library, plus native Ascend 950 support for TileLang. The stated goal is to give developers a simpler programming model than Nvidia's CUDA and reduce reliance on the Nvidia ecosystem. Huawei says it fully supported the development.
The other story is about money and operations. A call center consultancy rented a four-GPU Nvidia H200 server to test the claim that DeepSeek is '80x cheaper' than Claude. It found that at on-demand rates the box would cost about $13,200 per month, more than double the consultancy's September Claude subscription bill. The '80x' figure comes from comparing DeepSeek's API list prices with Claude's, not from renting hardware. The test also hit security snags: reviewers found ways agent code could escape the sandbox, so DeepSeek never wrote code and instead ran as read-only reviewer agents.
In short, the sources don't conflict so much as cover different ground. Both point to DeepSeek pushing for cheaper or alternative AI compute, but the rental test is a reminder that low API prices don't automatically translate to low self-hosting costs. There's no direct comparison between the Ascend tools and the H200 test; they're separate developments.