Taiwan charged nine over smuggled Nvidia servers
Taiwanese prosecutors indicted nine people, including an Nvidia Taiwan employee, over 130 B300 servers papered to stay on the island, and Alabama's attorney general subpoenaed OpenAI over the experimental model that spent days inside Hugging Face.
Taiwan indicted nine over 130 smuggled Nvidia B300 servers Prosecutors charged nine people on Monday, among them an employee of Nvidia’s Taiwan unit and two of Super Micro’s, over end-user paperwork claiming 130 B300 servers would stay on the island. Seventy-four reached Chinese customers directly and through Japan, Indonesia and Hong Kong before customs stopped the remaining 56 on their way to Japan; the charges include breach of trust and document forgery. Source
Unitree shares have fallen 45 per cent since their debut The humanoid maker steadied on Tuesday after three straight sessions of losses took it 45 per cent below the peak it hit on last Wednesday’s listing, cutting roughly $30bn from a market value that touched $66bn on day one. Its own prospectus shows adjusted net profit fell 53 per cent to 40m yuan in the first quarter of 2026. Source
Alabama subpoenaed OpenAI over the model that hacked Hugging Face Attorney General Steve Marshall said on Monday he had issued a subpoena to OpenAI, naming Sam Altman, to establish whether July’s lab escape — an experimental model gaining unauthorised access to several networks and spending days inside Hugging Face — breached Alabama’s Deceptive Trade Practices Act. It follows a multi-state coalition letter earlier this month demanding OpenAI stop the tests that produced the hack until it can show they are controlled, and Marshall said the leak proved “Alabamians’ and Americans’ worst fears about artificial intelligence are not just theoretical”. Source
Chinese state hackers doubled their attack volume using DeepSeek Taiwanese research firm TeamT5 told Bloomberg on Monday that state-affiliated groups have more than doubled the attacks they run since handing reconnaissance and exploit-writing to DeepSeek, with chief analyst Charles Li putting the choice down to weak guardrails and low cost rather than capability. It names Grimfengxi as generating exploit code, Huapi as hitting a Taiwanese company’s email system and Teleboyi as mapping about 1,000 IP addresses, and says a fourth group used Claude to move laterally through a Taiwanese tech firm. Source
Nvidia put its Groq 3 LPX inference chip into full production Announced at Hot Chips on Monday, the accelerator extends the Vera Rubin platform and recorded 3,400 output tokens a second on Gemma 4 31B at 100,000-token context in Artificial Analysis testing; Nebius is the first cloud to take it, on its Token Factory inference platform. Nvidia’s $20bn deal last year took Groq’s founder and senior staff, and the release notes the Groq name is used under licence from Groq, Inc., which plans to be among the chip’s earliest adopters. Source
Meta’s first paid AI product runs on Anthropic’s models The Information reported on Monday that Hatch, a consumer agent platform Meta plans to launch within weeks at up to $199 a month, currently runs on Claude Opus 4.6 and Sonnet 4.6, with a move to Meta’s in-house Muse Spark planned. Meta built a sandbox simulating DoorDash, Etsy, Reddit, Yelp and Outlook to train it. Source
OpenAI put the five-hour cap back on ChatGPT Plus OpenAI said on Monday that Codex and ChatGPT Work would return to a five-hour usage limit for Plus subscribers from Tuesday, after several weeks in which only a weekly cap applied. Pro tiers stay exempt for now, and Enterprise and Edu run on separate credit systems; engineering lead Thibault Sottiaux put the change down to smoothing compute load and to casual users burning a week’s allowance in one sitting. Source
Thomson Reuters built its own frontier model for legal work The company said on Monday that “Thomson”, trained from an open-source base on decades of Westlaw, Practical Law, Checkpoint and Reuters material for about $40M in compute and talent, is its first proprietary large language model. It ships first inside CoCounsel Legal’s tabular analysis feature, with a small open-weight version going to Hugging Face for academic use, and CoCounsel stays multi-model. Source