Capability wasn’t in dispute today. Supply and jurisdiction were — who can make the memory, whose factory stamped the board, and which rulebook a chatbot answers to.
Brussels used platform law rather than AI law, Taoyuan prosecutors raided a substrate maker over where its boards were made, and the memory shortage surfaced in Apple’s quarterly numbers.
The Commission designated ChatGPT a Very Large Online Search Engine under the Digital Services Act on Aug 31 — the first chatbot classed that way — after the service reported at least 45 million average monthly users in the EU. Reddit and Roblox were designated Very Large Online Platforms the same day, and the Commission’s list now runs to 28 services. The duties are the DSA’s heavy set: systemic-risk assessments covering illegal content, minors, physical and mental wellbeing, fundamental rights, elections and public security; a public advertising repository; semiannual transparency reports; a crisis-response mechanism; and data access for vetted researchers. Compliance falls due four months from notification, in January 2027. The venue is the interesting part. The AI Act’s transparency obligations became enforceable Aug 2, but it’s the DSA — drafted for platforms, years before this — that arrived first with a named deadline attached. Researchers have already flagged the unsettled question of whether “data access” extends to training data or model weights. Nobody has answered it, and the answer is worth more than the designation.
The Information reported Aug 31, citing two people familiar with the work, that ChangXin Memory Technologies has produced high-bandwidth memory in limited volume for the first time. HBM3E sits a generation behind the HBM4 that SK Hynix, Samsung and Micron are mass-producing; CXMT’s yields were running near 25% as of June, against a technical gap estimated at three to five years. Alibaba’s T-Head and Cambricon are testing the parts for products in 2027. US export controls have kept advanced HBM out of Chinese accelerators, and that is the specific bottleneck here — not the logic die, the memory stacked beside it. Keep the scale honest in both directions: small batches at low yield is a milestone, not supply, and a company that raised $8.6B in its Shanghai IPO without earmarking any of it for HBM has now made some anyway. The same firm sued the Defense Department last week over its “Chinese military company” designation. The lawsuit and the fab run on separate clocks.
The Taoyuan District Prosecutors Office searched Unimicron Technology’s headquarters in Guishan and its plant in Zhongli on Friday and questioned 14 executives and employees. Prosecutors allege China-made printed circuit boards were shipped into Taiwan and relabeled as domestic, citing the criminal code on document forgery and false labeling of origin. The PCB division’s general manager was released on NT$15 million bail, about $474,000; four other executives posted between NT$300,000 and NT$12 million and nine people were released without bail. A company spokesman said Unimicron is cooperating with the investigation “regarding disputes over product origins.” Unimicron makes ABF substrates, the packaging layer underneath high-performance computing and AI silicon, so this lands on the supply chain rather than next to it. Export controls and tariffs are both written about where a thing was made, which makes the country-of-origin stamp the enforcement surface for the entire apparatus. This is what auditing one looks like.
Draft IPO documents reported by the Wall Street Journal show SB Energy granting OpenAI warrants the paper valued at about $5.5B to secure it as a data-center tenant — a discount on tenancy paid in stock rather than in rent. Reuters said it could not independently verify the figure, and neither company commented. Nvidia has separately put $1.5B into SB Energy and guaranteed roughly $105B of OpenAI’s Ohio lease; SB Energy is reported to be seeking a debut valuation above $50B. On the income side, OpenAI said Aug 31 that ChatGPT advertising has reached a $1B annual run rate roughly 200 days after US testing began, and opened self-serve ad buying in India, Europe, the Middle East and North Africa, with ads now running in more than 40 countries. It also began letting some large customers pay only when a task actually completes. Three different ways of being paid, all surfacing in the week the company is assembling an offering.
The Information reported OpenAI and rival labs purchasing tens of thousands of Mac minis and Mac Studios, with Anthropic renting Mac capacity through AWS, to train computer-use agents on the same class of machine those agents are meant to operate. Apple’s Mac revenue rose almost 29% to $10.4B in the June quarter, a record for the period, and the most powerful configurations have been sold out for months — the constraint named is the memory shortage running through the rest of this week’s file. The appeal is unglamorous: strong chips, unified memory and cooling that holds under sustained load, which is a different bet from the dedicated GPU that Nvidia’s DGX Spark makes. Around the edges, the open-source Exo project clusters Macs to run large models locally, and Peter Voell, formerly of OpenAI’s infrastructure team, is building an Apple-based cloud called Mount Thor. A company that sells desktops to households is now a line item in frontier training budgets, which isn’t a position Apple went looking for.