DeepSeek's Huawei Training Bet, Google's €403M Ireland Fine, and Alibaba's New Qwen Chief
Five items this morning, and four of them are really one item wearing different clothes: who gets to decide how fast AI moves, and on whose hardware.
China’s AI Plan Now Runs on Huawei’s Delivery Calendar
Liang Wenfeng rarely says anything in public, so the framing matters. Training DeepSeek’s models on Huawei silicon is, in his words, one of the company’s biggest bets. He called it a bet, which is a stronger word than a company hedging against tighter export controls would reach for. Huawei is expected to hand over training chips in the fourth quarter or early 2027, which puts the whole thing on a schedule that nobody outside Shenzhen can verify.
Read that timing carefully. It is a promise, not a shipment. Chinese labs have spent two years running inference on domestic accelerators while quietly keeping the heavy training runs on hardware that arrived before the rules tightened, or through channels nobody wants to describe on the record. Moving training itself onto Huawei parts is a different kind of commitment. The software stack has to hold up over weeks of continuous work, the interconnect has to keep thousands of chips talking, and the failure rate has to stay boring. Any of those going wrong costs a model generation.
If it works, the export control regime loses most of its remaining teeth. If it slips two quarters, DeepSeek spends 2027 shipping refinements of what it already has while others ship something new. That is a real fork in the road, and Liang is telling everyone which side he picked.
The Economists Xi Brought Along Are Nervous
The second China story this morning cuts against the confidence of the first. As Xi Jinping arrives in Washington, Chinese economists are warning that Beijing’s concentration on AI is soaking up attention and capital while the broader economy sits in genuine trouble. Property is still unresolved. Local government balance sheets are still a mess. Youth employment numbers are still the kind of statistic that gets published irregularly.
This is the oldest tension in industrial policy. A state can pick a winner and pour everything into it, and sometimes that works spectacularly. It also means that when the sector underdelivers, there is no second engine warming up. Beijing has made AI the answer to a question that is partly about growth and mostly about national standing. Those two things do not respond to the same medicine, and the economists raising their hands seem to know it.
Put the two stories side by side and you get the actual picture. The whole program is leaning on a chip delivery that has not happened yet, in an economy that needs the program to pay off quickly.
Ireland Fines Google €403M and Starts a Six-Month Clock
The Irish Data Protection Commission has fined Google €403 million over how it handled location data, following complaints from European consumer rights groups. Six months to bring the processing into line.
The number is real money and also not. Google will write the cheque without a board meeting. The interesting half is the compliance order, because that is the part that changes product behaviour rather than a line in a filing. European regulators worked out some time ago that fines alone get absorbed as a cost of doing business, so the enforcement now comes with instructions attached and a deadline for following them.
Location data keeps coming back because it is the most commercially useful thing a phone knows and the hardest to justify collecting at the scale advertising wants. Every settlement narrows the gap between what the consent screen says and what the pipeline actually does. Slowly, and one jurisdiction at a time, but it narrows.
Alibaba Finally Names Someone to Run Qwen
Dayiheng Liu, a senior researcher, now heads the Qwen project. After the reorganisation rounds Alibaba has been through this year, the news value is less about the person and more about the fact that there is a single name on the door at all.
Qwen is one of the few open-weight model families with genuine global adoption, and it has been running through a period where nobody could say with confidence who owned the roadmap. Researchers leave when that goes on too long. Naming a technical lead rather than an executive suggests Alibaba wants the model line steered by people who train models, which is the correct instinct and not always the one large companies follow.
A Ban on Recursive Self-Improvement
There is a case circulating this morning for a regulatory ban on recursive self-improvement, built on recent hacking incidents and the industry’s own warnings about existential risk. The reasoning is straightforward: competitive pressure means no lab will slow down voluntarily, so the brake has to come from outside.
The hard part is the drafting. Systems that write code, evaluate the results, and use those results to write better code already exist in ordinary development work. A ban that catches the dangerous version also catches a large amount of routine engineering, unless the line is drawn with a precision that legislation rarely achieves. Write it loosely and it bans everything or nothing.
Still, the piece arrives on the same day that a Chinese lab announces a hardware bet designed to keep training at full speed regardless of what anyone in Washington or Brussels decides. Whatever rule gets written applies to the labs that answer to the people writing it. That has always been the flaw in the argument, and nobody has solved it yet.