GlobalFoundries to run two US chip plants backed by $21B from Japan
Two supply-side stories this hour: Japan bankrolls a pair of American fabs, and a startup argues the guardrails that keep frontier models from helping hackers are now keeping them from helping defenders.
GlobalFoundries will operate two new chip plants in Malta, New York, with Japan providing as much as $21 billion — and Washington taking an equity stake without putting in capital. Each fab carries roughly $10 billion of the tab, with the first 300mm line targeted for 2029 and each one adding about 900,000 wafers a year at full tilt. These are mature-node parts for automotive, industrial and communications customers, not AI accelerators — GlobalFoundries walked away from the leading-edge race years ago, and CEO Tim Breen told Reuters priority goes to US and Japanese buyers, with Renesas and Apple among the first. Per Reuters and the company's own release, the US Commerce Department takes about 4% of GlobalFoundries on top of its existing holding, lifting the government's total to roughly 7%, while Japan funds the two plants outright. The package is part of Japan's $550 billion US investment pledge and comes bundled with a tariff cut from 25% to 15%; the stock rose about 4% on the news.
The AI angle is narrower than the headline number suggests, and worth stating plainly: the two fabs sitting on today's announcement are not AI compute. What is AI-relevant is the $955 million CHIPS award reallocation toward silicon photonics, SiGe and advanced packaging at Malta and Burlington, Vermont — plus GlobalFoundries' separate $2 billion interposer deal aimed at AI packaging. Chip supply chains have been a running theme here — we covered Applied Materials commits $5 billion to India's chip supply chain earlier this week — and this is the same pattern again: governments paying to site capacity inside their own borders, with the money doing the talking and the AI demand story mostly a backdrop. Watching US taxpayers take equity in a foundry in exchange for zero dollars is the part worth remembering.
Artemis Security launched Orion-1, a purpose-built model that investigates and responds to corporate network attacks on its own, and claims it beats frontier generalists at the job. The model pulls logs from authentication systems, cloud setups, endpoints, SaaS apps and network gear, then links separate signals — a login, a fresh cloud permission, a new inbox rule — back to one actor before assigning a confidence score and recommending or triggering a response. On the company's own "Decision-Grade Readiness" benchmark, Artemis reports Orion-1 at 80.8 points on attack reconstruction against 58.3 for the next-best model, GPT-6 Sol, with Grok 4.7 and Kimi K3 also tested. It is in private preview, high-impact actions need human approval by default, and Artemis — out of stealth since April with $70 million raised — has not disclosed whether it fine-tuned on top of another provider's model. Treat that self-run benchmark as a claim, not a result.
The reason the product exists is the more interesting part: guardrails written to stop attackers keep catching defenders. Artemis and The Decoder both note that Claude Opus 5.5 routes most cybersecurity tasks down to the older Opus 4.8, and that when Hugging Face tried to analyze a fully autonomous AI attack on its own infrastructure, commercial frontier models refused the analysis outright — the team fell back to open-weight GLM 5.2 on its own servers. OpenAI hands its less-restricted cyber tier only to verified defenders, and Anthropic does the equivalent through partner security products. The gap between "we can't let this model help a hacker" and "this model can't help the person being hacked" is now a market, and startups are going to sell into it.
What to watch: whether Japan's money gets fabs built on schedule, and whether specialized defense models win adoption or the labs' locked-down frontier models win by default.
Should a model that can hunt attackers on its own need a human sign-off for every containment step — or does the sign-off defeat the point? Tell us in the comments.
