A Chinese lab just beat two frontier US models on coding, for a fraction of the price.
Z.AI's Ox Alpha beat Claude Fable 5 and GPT-5.6 Sol on a coding benchmark, then shipped open under an MIT license. OpenAI's first in-house chip beat Nvidia's Blackwell on efficiency. And Anthropic opened a standard that lets any lab wire its own hardware to an AI agent, no partnership required.
Three different chokeholds, on models, on chips, on physical infrastructure, loosened in the same week.
Two stories below have nothing to do with model or chip wars, and they're worth a look anyway.
Sponsored by AI Insider
Premium AI tools without the premium price.
Get access to selected AI & SaaS tools at exclusive Insider pricing.
🎁 New members also get a free tool or exclusive discount.
Join us on WhatsApp →
Frontier Reads
Don't skip Simon Willison's timeline of the Hugging Face attack, reconstructed message by message, showing how OpenAI's agents talked each other past their own hesitation. Read more
IEEE Spectrum's piece on the CPU comeback explains why AWS engineers got a mandate to conserve CPU cycles, the clearest read on why Nvidia built Vera. Read more
Deep Current
Z.AI's Stealth Model Revealed
Z.AI, the Chinese lab formerly known as Zhipu, confirmed it built Ox Alpha, the anonymous model that led coding leaderboards for two weeks before anyone knew who made it. It hit 80% on one benchmark, ahead of Claude Fable 5 at 65% and GPT-5.6 Sol at 52%.
It ships open tonight under an MIT license, at a fraction of what Western labs charge. That combination, frontier performance, open weights, low price, is exactly the pressure point US labs have spent two years trying to avoid.
Claude Can Now Control Hardware
Anthropic opened a research preview of the Model Hardware Standard, a shared spec that lets AI agents safely run lab and factory equipment, microscopes, liquid handlers, robotic arms, cutting integration time from weeks to hours.
At quantum computing startup QuEra, an agent using the standard cut a laser-recovery task from 150 seconds to 6, succeeding 99.3% of the time versus 58% before. Genentech, AWS, and Hugging Face are already building support, a sign labs want AI running their gear directly.
OpenAI's Chip Beat Nvidia's Best
OpenAI unveiled Jalapeño, its first in-house chip, built with Broadcom for inference. On an independent test, it beat Nvidia's Blackwell systems on performance per watt in nearly every scenario, using 1.5 to 1.9x less power for the same work.
It's a real dent in Nvidia's inference margins, its fastest-growing business, and a step toward OpenAI depending less on any one chipmaker. The catch: the comparison used Blackwell, not Nvidia's newer chips, and Jalapeño won't ship at scale until 2027.
In Radar
Apple's new desktop leans into "agentic." The redesigned Mac mini starts at $899, up from $599 a year ago, with Apple now pitching it as a machine built for always-on AI agents, not just apps. It'll be assembled at Apple's new Houston plant. Read more
Robot startups can't hit the 65% bar. The FCC's new rule requires foreign-made robots to be 65% U.S.-made, and founders say no domestic supply chain exists yet. Read more
Most of that power number is fake. Wood Mackenzie expects utilities to build only 28% of the 1,066 gigawatts data centers have requested, the rest is speculative, duplicate filings. Read more
Outside the feed
