[AI Weekly] Alibaba just outbuilt the export ban
Alibaba spent 53 billion dollars to prove export controls do not work, a coding startup just raised enough to buy a small airline, and humanoid robots are shipping at a rate nobody predicted twelve months ago. Here is the week.
The Big 3
Alibaba says it built the chip export controls were meant to prevent
Alibaba unveiled the Zhenwu V900 on 22 September, calling it the most powerful AI chip in China and roughly three times faster than its predecessor. The number that matters is the clustering claim: up to 500,000 units wired together for frontier training, backed by a spending commitment north of 53 billion dollars over three years and a target of 20 gigawatts of data centre capacity by 2032. The announcement landed days before a US-China meeting where AI competition was on the agenda.
Why it matters: the whole premise of chip export controls is that China cannot get hardware at this scale. A claimed 500,000-unit domestic cluster is the counterargument, made publicly and on purpose.
Read the full breakdown on Future Technology
Cognition raised $2 billion to build AI that writes its own code
Cognition, the company behind the Devin coding agent, closed a $2 billion Series E from a16z, Accel, Founders Fund and General Catalyst this week. The bet is that autonomous coding agents, ones that plan, write, test and open the pull request without a human at the keyboard, are a standalone category worth backing at scale, not a feature the model labs will absorb.
Why it matters: that is more capital in one round than most software companies raise in a decade, and it puts Cognition in direct competition with OpenAI, Anthropic and Google, all of which now ship their own coding agents.
Read more on Future Technology
Humanoid robots are shipping at a rate nobody predicted
Global humanoid robot shipments rose roughly 272 percent year on year in the first half of 2026, to somewhere between 19,000 and 22,000 units. Chinese manufacturers hold 93 to 97 percent of that market, and more than 70 percent of shipments now go to industrial and commercial work rather than demos. XPENG's robotics arm just raised over $900 million at a $6.3 billion valuation, and Figure's robots have crossed 1,000 units in production, including parts sequencing work at BMW's Spartanburg plant.
Why it matters: this stopped being a stage-demo story a while ago. The robots that matter now are the ones clocking in.
Read the full story on Future Technology
Quick Hits
Open weights incoming: StepFun's Step 5 Preview (600B parameters, 27B active) scores seven points behind the closed frontier on reasoning benchmarks, at roughly a dollar per million input tokens, with open weights dropping 15 October.
Regulation split: the Council of Europe's AI treaty, the first legally binding international AI accord, took effect this month, the same week the US pushed Carolina Principles at the G20 arguing against AI-specific rules.
Training data money: Snorkel AI raised $350 million to build out reinforcement learning and evaluation infrastructure, the unglamorous plumbing every frontier lab depends on.
September's model dump: 14 new models shipped from 10 providers this month alone. We put the useful ones in a table so you do not have to track it yourself.
Tool of the Week
LM Studio is the easiest way to run open-weight models like Step 5 or GLM on your own machine, no terminal required. Point-and-click model downloads, a built-in chat interface, and an OpenAI-compatible local API for anything you build on top of it.
Who it is for: anyone who wants to try October's open-weight model drops without waiting on someone else's API queue or sending prompts to a third party.
If you are shopping for hardware to run it on, an RTX 5090 handles most open-weight models comfortably: ASUS ROG Astral RTX 5090 (32GB).
One More Thing
Two continents made opposite bets on AI trust this month. Europe adopted the first legally binding AI treaty. The US argued at the G20 that AI should not be regulated as its own category at all. Both happened in the same week Alibaba proved, in public, that the hardware controls meant to slow the whole race do not actually slow it much. Nobody has agreed on how much to trust these systems, and the systems keep shipping anyway.
Forwarded this? Get your own AI briefing at futuretechnologyhq.com/newsletter.
Until next Wednesday. Nath, Future Technology
Some links in this newsletter may be affiliate links. We only recommend products we genuinely think are worth your time.