Google released open-source TorchTPU, letting developers run existing PyTorch applications on Google TPUs and Google Cloud after a configuration change, without rewriting code.
Some estimates put PyTorch in 85% of AI applications, and Nvidia GPUs have historically shaped its features. Google’s TPU agent suite helps developers migrate apps and fix errors; an updated LiteRT works better with PyTorch applications. Amazon released TorchNeuron Monday to port PyTorch apps to Trainium and AWS without rewrites.
Google Cloud will offer Nvidia Vera Rubin A5X instances next month and plans up to 960,000 Rubin GPUs across multiple sites. Google is lowering the software barrier to TPUs while still offering Nvidia infrastructure.
Why it matters: Google and Amazon can now court the developers behind most AI applications without asking them to rebuild their software, putting pressure on Nvidia’s default position even if developers do not move en masse.
Canadian organizations can now access seven OpenAI models through Amazon Bedrock from AWS’s Canada (Central) and Canada West (Calgary) Regions. The lineup includes GPT-6.1 Sol, GPT-6 Astra, Sol and Luna, plus GPT-5.6 Sol, Terra and Luna.
Bedrock cross-Region inference sends requests from an application’s source Region to a supported destination Region, giving customers access to a broader capacity pool without managing routing themselves. The models accept text and image inputs, support 1-million-token context windows, reasoning, tool calling and prompt caching.
Customers can use US geographic profiles, which keep processing within Canada and the US, or global profiles, which draw on supported commercial AWS Regions. They can invoke the models through OpenAI’s Responses and Chat Completions APIs or Bedrock Converse.
Why it matters: Canadian AWS teams can now trade off North American processing controls against a broader capacity pool during demand spikes without building their own cross-Region routing.
NVIDIA introduced Confidential Bare Metal, a deployment option that lets frontier-model builders run sensitive workloads on dedicated third-party infrastructure while retaining control of their models, data and host software. It joins Confidential Containers and Confidential Virtual Machines, with support extending across Grace Blackwell NVL72 and Rubin systems.
Separately, ALTK-Evolve raised GPT-4.1 agents’ strict Pass^4 reliability on Sierra’s τ²-bench from 57.5% to 67.5% in retail and from 35% to 45% in airline tasks. It distills successful and failed attempts into behavioral guidelines and appends them to the prompt—without updating model weights or using human annotation.
Why it matters: Cohere can extend Model Vault’s hardware-rooted protection to GPUs on third-party infrastructure, while customer-service teams can improve GPT-4.1 reliability without updating model weights.
Anthropic’s Claude largely autonomously designed protein binders for scorpion neurotoxins that kill more than 3,250 people annually. Inside Claude Science, it generated 318,484 designs targeting five toxins from four scorpion species, then selected 180 for laboratory testing. Researchers added 31 more.
Of the 211 designs tested, 17 met the criteria for a binder. The two strongest bound extremely tightly to Cn2, a toxin from the Mexican scorpion, near the site targeted by an antibody known to neutralize Cn2 in mice. Two independent labs confirmed the results, but researchers have not tested whether the molecules work as an antivenom. Claude worked for 24 hours after an initial prompt before the team provided more instructions.
Why it matters: For Anthropic and venom researchers, Claude produced 17 experimentally validated binders, but the decisive question—whether binding to Cn2 translates into antivenom activity—still has no answer.
Anthropic is circulating a draft bill that could block an AI model’s release when evaluators find a “foreseeable, imminent” risk of more than 50 deaths or $1 billion in property damage. The listed triggers are biological weapons, a model evading developer controls, or a cyberattack.
Independent evaluators registered with the Commerce Department would have 72 hours to report catastrophic risk. Commerce would then have to ask a Washington, D.C., judge to halt distribution; federal agencies could not stop a release on their own.
Anthropic says the proposal follows its Advanced AI Framework, which it has advocated to federal and state lawmakers since June. The bill goes further than Nvidia CEO Jensen Huang’s call for no new AI laws.
Why it matters: For Anthropic and rival AI labs, the draft trades unilateral federal intervention for court review while leaving a credibility problem: labs could pay the evaluators judging their models, provided payment is not tied to the findings.
DHS ordered federal civilian agencies to stop buying DeepSeek products and remove them from existing IT systems. The order also covers affiliated companies and third-party services built with DeepSeek technology, though agencies can request exceptions.
The order sets no removal deadline, identifies no specific security concerns, and leaves unclear whether AI gateways such as OpenRouter qualify as services “built with” DeepSeek technology. It extends a December 2025 law that restricted DeepSeek in intelligence and national-security systems.
DHS warned that sensitive information from U.S. government users could potentially reach Chinese authorities under Chinese law. China condemned the move, while the U.S. government previously considered sanctions, a trade blacklist, and rules limiting which Chinese open-weight models federal agencies could use before pulling back.
Why it matters: For OpenRouter and other federal IT suppliers, the undefined “built with” standard creates a procurement risk: agencies may reject services that use DeepSeek implementations even when those services are not DeepSeek’s own products.
Germany’s cabinet has moved to ease data-center heat rules, dropping a requirement that new facilities reuse waste heat. Under the amended Energy Efficiency Act, operators would instead assess whether recycling is technically and economically feasible; if not, they receive an exemption. Parliament must ratify the change.
Developers blamed the 2023 law for making projects nearly impossible. Stack Infrastructure canceled a $3 billion Hesse facility, while state-owned Komm.ONE withdrew from a Baden-Württemberg project and pointed to the act’s burdens.
Anthropic is also short on cloud capacity from AWS, Google, and Microsoft, which increasingly view it as a competitor. A leaked S-1 says the Claude maker could spend more than $750 billion on cloud contracts while leasing data centers and using third-party operators it considers less reliable than hyperscalers.
Why it matters: The amendment may help German developers pursue projects, but a small-town council’s rejection of a $2.5 billion facility shows climate opposition can still block data-center construction.