Anthropic reportedly nears a second $45 billion compute…

What happened: Anthropic reportedly is close to a second compute contract with Nscale, the same size as their August deal.

• The August agreement between the two companies was also valued at $45 billion.
• Bloomberg reports a new contract of the same size is close, though talks are ongoing and details could shift.
• Nscale is seeking a public listing, and investors have pushed back on its proposed $30 billion valuation.

Why it matters: Frontier labs are locking up multi-year GPU capacity from specialist cloud providers at huge scale, and access to compute now decides who can train and serve models. Plan capacity commitments and supplier concentration risk early, even if you are much smaller.

What happened: DeepSeek released an inference engine, the software that runs a model to serve answers, but it supports only one model.

• The only supported model is DeepSeek-V4.1-Flash, and it targets Nvidia SM100 GPUs and Huawei Ascend 950 NPUs.
• It was built mostly on vLLM, with ideas from SGLang and TensorRT-LLM, and ships with a companion token codec called deepseek-recipe.
• A full production service still needs an external KV cache (a memory store for model context), PD disaggregation and load balancing; reported production figures include 155 trillion tokens per day per cluster and a 96% input cache hit rate.

Why it matters: You get a permissively licensed reference for MoE (mixture of experts) serving and speculative decoding, plus a second hardware path beyond Nvidia. The release is a subset of DeepSeek's setup, so benchmark before you assume it matches the hosted API.

What happened: Amazon opened TorchNeuron, an open-source project that lets PyTorch code run on its Trainium chips with minimal changes.

• It runs PyTorch workloads on Trainium accelerators without extensive application rewrites.
• Amazon is targeting the software layer, where porting code is often the main obstacle to switching chips.
• Developers have stayed with Nvidia because the libraries and tools around its GPUs already work.

Why it matters: If the framework works with minimal edits, comparing cloud chips on cost becomes a real option, which can change your unit economics for training and serving. Ask how much of your code would change before you plan a move.