Three stories worth your coffee today. The federal government finally kicked DeepSeek off its computers, a draft AI safety bill written by Anthropic leaked, and Amazon and Google both took a swing at Nvidia in the same 48 hours. Let's go.
01 of 03
DHS orders DeepSeek ripped out of every civilian agency
The Department of Homeland Security published an order in the Federal Register on Tuesday, October 20, that does two things. First, it bars federal civilian agencies from buying any IT products or services from Hangzhou DeepSeek Artificial Intelligence Co. or its affiliates. Second, it tells agencies to remove all DeepSeek products already on their systems.
The part that matters most is the fine print. The order also covers "third-party products and services developed using DeepSeek implementations." So it is not just the DeepSeek chat app. If a contractor built a tool on top of DeepSeek's open models, that tool has to go too.
The legal hook is the Federal Acquisition Supply Chain Security Act of 2018, the same law built for pulling risky vendors out of government supply chains. It came after a recommendation from the Federal Acquisition Security Council, and the CISA team handling it is listed as the contact. DHS's stated worry is data: sensitive information from government users could be reachable by Chinese authorities under Chinese law. A December 2025 law had already kept DeepSeek out of intelligence and national security systems. This stretches that to the whole civilian side.
Beijing was not thrilled. Foreign Ministry spokesperson Mao Ning said China opposes "overstretching the concept of national security and politicizing technology and trade issues."
Remember last month, when the FBI, NSA, and CISA named DeepSeek, Moonshot AI, Alibaba, MiniMax, StepFun, and Z.AI for distilling American models? I said an alert is not a sanction. Well, here is a sanction, at least for one of the six. It is a procurement ban, not an export control, but it has teeth, and the third-party clause is the sharpest one. My bet is a lot of agency IT shops are about to find DeepSeek buried in tools they did not know were built on it. What I want to know is whether the other five names get the same treatment, because right now DeepSeek is the only one paying for it.
02 of 03
Anthropic's draft AI safety bill: 50 deaths or $1 billion
Politico got its hands on a confidential draft bill from Anthropic, and an Anthropic spokesperson confirmed it. The company says the draft reflects its "Advanced AI Framework" and that it has been pitching these ideas to federal and state lawmakers since June. It has not been introduced in Congress.
Here is how it would work. Independent evaluators register with the Commerce Department and periodically check frontier labs and their models. Labs can pay those evaluators, as long as the "payment is not contingent on the evaluations' results." If an evaluator finds a "foreseeable, imminent risk" of catastrophe, they have 72 hours to tell Commerce. Commerce can then ask a judge in Washington, D.C. to halt the model's release.
The bar for catastrophe is specific. More than 50 people killed or seriously injured, or more than $1 billion in property damage. And it only counts if the harm comes from one of three things: the model helping create or release a biological weapon, the model evading its developer's controls, or a cyberattack. The draft also says no federal agency can halt a model on its own say so. Commerce is the only one that can even ask, and a judge has to sign off.
Critics showed up fast. Heidy Khlaaf, chief AI scientist at the AI Now Institute, called it "autocratic," arguing it lets the company set its own harm thresholds and ignores the harms AI is causing right now. Mistral CEO Arthur Mensch said American labs use doomsday talk to close off the market. On the other side, Nvidia's Jensen Huang has been saying the industry does not need new laws at all.
Here is where I land. I would rather have a lab put a real number on paper than keep waving its hands about "safety." Fifty people and a billion dollars is at least something you can argue with. But the lab gets to pay its own graders, and the bill mostly reads as a list of things the government is not allowed to do. That is a lot of guardrails on the regulator and not many on the company. Before this goes anywhere, I want to see somebody other than the people being regulated take a red pen to it.
03 of 03
Amazon and Google both go after Nvidia's real moat
Nvidia's advantage was never just the chips. It is that most AI code is written in PyTorch, and PyTorch has always run best on Nvidia. Some accounts put PyTorch at 85% of AI apps. This week two of Nvidia's biggest customers tried to break that link.
Amazon went first on Monday, announcing TorchNeuron, a native PyTorch backend for its Trainium2 and Trainium3 chips. It is open source under Apache 2.0. You flip one configuration setting to point your code at Trainium, no rewrite. Amazon says Cartesia ran its standard inference pipeline on it with minimal code changes, Hugging Face added native Trainium support to its tools, and the Lieber Institute for Brain Development is using it for drug discovery research. Amazon is also sending code upstream to the main PyTorch project to make it less tied to Nvidia by default.
Google answered Tuesday by open sourcing TorchTPU, which does the same trick for its Tensor Processing Units. Google says it worked with Meta, Hugging Face, Inferact (the folks behind vLLM), and RadixArk (behind SGLang) to make sure their tools play nice. It also put out a TPU agent suite that helps developers migrate apps and fix the errors that pop up along the way.
What I keep coming back to is that "you can switch" and "you will switch" are two very different sentences. Developers stick with Nvidia because everything just works, and one config flag does not erase years of muscle memory. But this is the first time I have seen both big clouds go after the software lock in the same week instead of just bragging about chip specs. If even a slice of inference work moves, Nvidia's pricing power takes a dent. Keep an eye on the cloud bills, not the press releases.