The week the AI race became a chip race
This week's headlines were less about smarter models than about the physical and political scaffolding underneath them — foundries, packaging deals, power grids, and the money and law racing to keep up. Two counter-currents stood out: compute consolidating into a geopolitical contest, and intelligence quietly shrinking to run on phones and in browsers.
The race became a chip race
The most consequential moves happened in fabs, not model cards. ASML posted €9.3bn in quarterly sales and flagged an Intel Panther Lake deal; Intel became the first to ship logic chips made with High-NA EUV tools and committed another $5bn to its Ireland fab; and Google reportedly chose Intel's packaging over TSMC for its next TPU — a rare crack in TSMC's dominance. The geopolitics hardened in parallel: China's CXMT closed in on Micron's DRAM capacity, DeepSeek moved to design its own chips around export controls, and the US eased AI-chip rules for the UAE. The throughline is that whoever controls advanced packaging and fab capacity increasingly controls AI, and every major player is now hedging its supply chain.
Small is the new frontier
Against the compute arms race ran a quieter counter-current: intelligence getting smaller and more portable. A 27B reasoning model was compressed to fit on an iPhone, Thinking Machines shipped its first open model (Inkling), and Google released LiteRT.js to run models directly in the browser, while Hugging Face added a native-speed vLLM backend and on-device AI gained traction globally. The strategic point is that the industry's default of renting frontier models over an API is no longer the only path — open weights and on-device inference are becoming a real alternative, with different economics and far more control for whoever adopts them.
Agents are mostly plumbing right now
The agent story matured in an unglamorous direction. A survey found most enterprise 'AI agents' are still effectively chatbots, and the week's real progress was in infrastructure rather than autonomy: Oak emerged from stealth with $60M for AI-agent security, Vint Cerf floated a standard for identifying AI agents online, OpenAI's Codex began encrypting instructions between agents, and Google's Gemini API added background tasks and remote MCP. The signal is that the hard, valuable work in agents right now is identity, security, and protocols — the plumbing — not raw capability.
Power is becoming the ceiling
Energy kept surfacing as the binding constraint. Data centers consumed 23% of Ireland's electricity in 2025, and data-center power demand was reported to be straining the US manufacturing push. As models get cheaper to train and run, the limiting factor is shifting from algorithms to megawatts — a constraint capital can't instantly solve, and one that increasingly dictates where, and whether, new AI infrastructure gets built at all.
Capital floods in; the mess piles up
The money was staggering and the governance questions were not resolved. AI helped drive a record $392bn in North American startup funding in the first half of 2026, SambaNova raised $1bn at an $11bn valuation, and Indian coding startup Emergent hit unicorn status. Yet the same week brought xAI suing a user over Grok-generated CSAM, a hack revealing that Suno scraped YouTube, Deezer and Genius for training data, and OpenAI employees funding a Super PAC to oppose their own company's political arm. Capital is compounding faster than the legal, ethical, and even internal-governance answers — a gap worth watching.