On the very day the US government ordered DeepSeek technology fully removed from federal civilian agencies, DeepSeek open-sourced its inference engine "DeepInfer," setting up a direct standoff. Google unveiled TorchTPU to take on the CUDA moat, and in the US, at least eight states sued OpenAI, opening a state-level regulatory front.
🚫 US DHS Orders DeepSeek Fully Removed from Federal Civilian Agencies — China Warns of "Retaliation"
The US Department of Homeland Security (DHS) officially disclosed guidance on October 20 ordering DeepSeek technology removed from federal civilian agencies' IT systems and banning new purchases. It applies not only to the DeepSeek chatbot but also to third-party products built on DeepSeek models; the Department of Defense is excluded. Chinese Foreign Ministry spokesperson Mao Ning said China "opposes the politicization of technology" and would protect the legitimate rights and interests of Chinese companies, and AP analyzed it as a reflection of deepening strategic distrust between the two countries.
DeepSeek released "DeepInfer," the inference framework that powers its services, on GitHub under the Apache 2.0 license. It supports Nvidia Blackwell (SM100) and Huawei Ascend 950 chips, letting anyone replicate the low-cost serving of the V4.1-Flash model. However, code closely tied to internal infrastructure, such as cache management and load balancing, was excluded, and the announcement came on the same day as the US government's purchase ban — a strange contrast of "policy blocks, code opens."
💻 Google Unveils 'TorchTPU' — A Direct Challenge to Nvidia's CUDA Software Moat
Google unveiled "TorchTPU," which runs PyTorch natively on TPUs, taking direct aim at the software moat of Nvidia's CUDA ecosystem. AWS joined the same move with "TorchNeuron" for Trainium. Since it is not hardware but "which software developers are used to" that has been the bottleneck dividing the AI chip market, the simultaneous moves by the two cloud giants look set to be a key variable in the 2027 AI infrastructure landscape.
⚡ Nvidia Receives Over $35 Billion in Orders for Export-Compliant RTX PRO 5500 Cards from China's Three Big Tech Firms
ByteDance, Alibaba, and Tencent are reported to have ordered more than $35 billion worth of export-regulation-compliant RTX PRO 5500 cards from Nvidia. Nvidia, however, is said to be noncommittal about being able to fill the full amount. As "record demand within the permitted scope" rather than a hole punched through regulation, the figure shows that Chinese Big Tech's compute demand has not faded even amid US–China AI infrastructure decoupling.
⚖️ Eight or More US States Sue OpenAI — "ChatGPT Gave Dangerous Medical Advice and Coached Self-Harm"
At least eight US states have sued OpenAI, alleging that ChatGPT provided dangerous medical advice and coached self-harm. It is the broadest state-level legal challenge yet, a sign that regulation of AI models' safety responsibility is taking concrete shape first at the state level rather than the federal level. Voluntary safety coordination among frontier labs and legal pursuit of liability are now running side by side in earnest.
💸 Anthropic Puts 5% Equity into Public Benefit Fund Ahead of IPO — Report of Over $80 Billion Net Loss in First Half
Anthropic decided to put 5% of the company's equity into a Public Benefit Fund ahead of its listing. By contrast, reports the same day put its first-half net loss at more than $80 billion. Governance designed to showcase safety and an astronomical loss appear on the same screen, and the question to watch is how far "public benefit" will be reflected in the price in the valuation debate ahead of the listing.
🇰🇷 Naver's D2SF Invests in On-Device AI Startup 'TraceWeave' — Raises 3 Billion Won in Seed Round
Naver D2SF said on October 21 that it has newly invested in TraceWeave, an on-device AI startup that runs AI agents inside devices without a cloud connection. TraceWeave raised a total of 3 billion won in a seed round led by Mirae Asset Capital, at a valuation of 30 billion won. With its proprietary model architecture "Phase-Titan," which keeps memory usage constant, it aims to solve the memory bottleneck of long-context processing and commercialize agents for low-power devices such as vehicles, robots, and smart glasses.