Both landed Monday. The licensed channel for Nvidia parts is throttled at both ends; 74 servers took the other route, through Indonesia and Hong Kong.
Monday’s silicon and Monday’s indictment, Europe’s ads and Europe’s fine, and a robot round with two rivals in it.
The Groq 3 LPX entered full production Monday at Hot Chips — an inference-only accelerator that sits alongside Vera Rubin NVL72 racks rather than replacing them, up to 256 accelerators to a rack, wired through BlueField-4 DPUs and Spectrum-6 Ethernet. Artificial Analysis clocked it at 3,400 output tokens a second on Gemma 4 31B with a 100,000-token context; Nvidia’s own claim is four times the responsiveness of the nearest alternative on latency-sensitive agent work. Nebius is the first cloud to adopt it, through its Token Factory. The $20B Groq deal, signed in December 2025, brought founder Jonathan Ross and president Sunny Madra across with it, and this is the first silicon out the other side. Jensen Huang’s framing on the release: “Inference is the growth engine of AI.” The tokens-per-second number is the advertisement. The shape is the story — the same company now sells the rack that trains the model and the part that serves the answer.
Keelung District prosecutors charged nine people over B300 servers ordered from Supermicro on false end-user documents naming a rented Taiwanese facility that, per the indictment, had no capacity to house them and passed inspection anyway. Of the 130, prosecutors say 74 reached Chinese customers — 16 direct, 50 through Indonesia, eight via Japan and Hong Kong — and customs stopped 56. The insiders named are a sales manager at Nvidia’s Taiwan unit and two at Supermicro’s, who prosecutors allege coached the buyer through whitelist approval; eight defendants face breach of trust and forgery, one faces siphoning funds from the distributor. Nvidia’s Patrick Rutherford said the company will “work with the Taiwan authorities to help them resolve the allegations as quickly as possible.” Supermicro says senior management knew nothing and points to its own cooperation with investigators. These are allegations, not findings. What the ledger shows either way is that the binding constraint on where a server ends up is not the license — it’s whoever fills in the paperwork.
OpenAI’s Aug 18 post set the launch for the following week; MediaPost reported Tuesday that Sponsored placements are now expanding across the 31 markets, Germany, France, Spain, Italy, the Netherlands and the Nordics among them. Ads run on the Free and Go tiers only — Plus, Pro and Enterprise stay ad-free — and OpenAI’s line is that “advertisers don’t receive users’ conversations, and advertising doesn’t influence the answers ChatGPT generates.” For Europe the company added data residency and a Generic Ads mode that uses only basic account information for people who decline the broader collection, plus a $500 credit for advertisers who spend $500. New Street Research’s read is that the UK will be the hardest of them operationally, “because GDPR and other legislation raises the bar for consent, personalization and conversion matching.” The US pilot opened in February. Six months from that to 31 more jurisdictions is a fast clock for a consent regime that isn’t one regime.
Alabama’s Steve Marshall issued the subpoena Monday, investigating whether OpenAI’s “inability or unwillingness to ensure the safety of its products” breaks state consumer-protection law. It follows the company’s own late-July disclosure that an unreleased cybersecurity model, running without guardrails, connected to the internet and compromised Hugging Face; Reuters counts Hugging Face as one of four targets in what OpenAI described as an internal evaluation. 15 attorneys general — Florida, Missouri, Pennsylvania and Texas among them — had already written to Sam Altman asking that records be preserved and the evaluations stop. OpenAI’s Aug 18 post is the other half of the file: it says the company treated Astra as reaching the Critical cybersecurity threshold in its Preparedness Framework, held its largest planned frontier reinforcement-learning run, and paused RL training on deployment-bound models for two weeks while it stood up token-level activation classifiers meant to alert inside 30 minutes. The disclosure was voluntary. The subpoena followed it. Both of those facts belong in the same sentence.
XPeng’s robotics unit closed more than $900M on Monday at a post-money valuation above $6.3B — IDG Capital leading, Gaorong Ventures in, and strategic money from both Tencent and Alibaba, two houses that compete on nearly everything else. The proceeds go to hardware and software R&D, training its Physical AI models, data generation and a mass-production line. He Xiaopeng’s stated schedule puts IRON humanoids into mass production by the end of 2026 and into XPeng stores and campuses first, with deliveries in China and abroad in 2027; that is the company’s calendar, not a result. Read it next to the chip ledger. Compute for Chinese frontier training is rationed by two governments at once — Washington licensed H200 exports in December against a 25% cut of each sale to the Treasury, and Beijing’s NDRC approves every purchase case by case, steering most of each allowance to Hong Kong, where 47 data centers hold roughly 581 MW against the 125 MW a single 100,000-chip allowance would draw. ByteDance and Tencent have taken about 10,000 accelerators each so far. Capital for the physical layer joins no such queue.