Google and Amazon go after Nvidia's software moat
Nvidia's real lock-in has never been just its chips; it's that most AI code is written in PyTorch, and PyTorch grew up on Nvidia GPUs. This week both big clouds attacked that directly. On Tuesday Google open-sourced TorchTPU, letting existing PyTorch apps run on its Tensor Processing Units with a configuration change rather than a rewrite, and said it worked with Meta, Hugging Face and the vLLM and SGLang makers on compatibility. A day earlier Amazon released TorchNeuron for its Trainium chips and began submitting code to upstream PyTorch that loosens its ties to Nvidia hardware. Nobody will abandon GPUs overnight, but switching clouds just got a lot cheaper, and that pressure lands on Nvidia's pricing. TechCrunch
What else matters
US orders DeepSeek off federal civilian systems
The Department of Homeland Security has ordered federal civilian agencies to remove DeepSeek products, adding to Washington's procurement bar on the Chinese lab. It arrives the same week DeepSeek open-sourced DeepInfer, the Apache-2.0 serving engine it says runs its own production services, so the open-weight supply chain is splitting along political lines. AI Policy Daily
OpenAI says its models can finally handle COBOL
OpenAI published two tests on legacy code. On a COBOL-to-modern-language migration task, GPT-6.1 Sol scored 86.6% at low effort, against 5.4% for GPT-5.6 Sol, and 97.9% at higher effort. These are OpenAI's own benchmarks, and it says real migrations still need expert review, but bank and government back-office rewrites look like a serious early market for coding agents. ToolSift summary
Tenstorrent raises $1.5bn to sell an open alternative to Nvidia
Jim Keller's Tenstorrent closed a $1.5bn Series E at a $6.2bn valuation, with Morgan Stanley, Fidelity and T. Rowe Price joining. It builds RISC-V-based AI servers and cards on an open-source software stack, and pairs neatly with the PyTorch moves above: investors are betting the market for non-Nvidia compute is about to get real. AI Policy Daily
