Welcome to March 22, 2026

The Singularity is building its own foundry. Elon Musk unveiled TERAFAB, a joint SpaceX/Tesla project aiming to produce over a terawatt of compute per year, 80% for space and 20% for Earth. SpaceX calls it “the next step towards becoming a galactic civilization,” designed to “close the gap between today’s chip production and the future’s demand.” Near Tesla’s Austin gigafactory, it will fabricate 2-nm chips in two flavors: edge silicon for Tesla’s robotaxis and Optimus robots, and high-power chips for SpaceX and xAI, including for a new mini AI data center satellite at 100 kW. Its planned recursive design loop will keep masks, fabrication, testing, and iteration in a single building. Tesla declared that “to understand the universe, you must explore the universe,” reading less like a slogan and more like a procurement order for the impending Dyson Swarm.

The models fueling that compute are getting lighter and longer-range. Chinese lab Evermind AI launched Memory Sparse Attention, an architecture with under 9% degradation scaling from 16k to 100 million tokens, decoupling memory from reasoning. OpenAI is “throwing everything” into building a fully automated AI researcher, targeting a research intern by September and a multi-agent system by 2028. Even the tooling layer is compressing: Cursor shipped Composer 2, billing it as frontier-level coding at a fraction of the cost.

The agents are inheriting the web. Cloudflare’s CEO predicts bot traffic will surpass human traffic online by 2027. The bots are learning to blend in. Browser Use found its agent was the stealthiest, accessing websites 81% of the time. Democratization is accelerating. OpenClaw proved fully autonomous AI can run at home without the big labs, and in China, schoolkids and retirees alike are raising “lobsters” as the craze goes mainstream.

That democratization is cooking mathematics. Researchers built an agent that generated 665 novel research problems in differential geometry, many unknown to experts. Terry Tao agrees, noting that even high schoolers can now make real contributions to frontier math thanks to AI tools.

The applications layer above is consolidating to match. OpenAI plans a desktop “Superapp” merging ChatGPT, Codex, and its browser. WordPress.com now lets AI agents draft, edit, and publish posts. Google Search is replacing news headlines with AI-generated text, turning the index itself into a generative layer.

The physical infrastructure is keeping pace. SoftBank is developing a $500 billion, 10-gigawatt data center campus in Ohio, built on a decommissioned uranium enrichment plant, and powered by natural gas. Not all infrastructure moves are legal, however. The co-founder of Super Micro Computer was charged with diverting $2.5 billion in Nvidia AI chips to China.

Atoms are catching up to bits. China now fields 140 humanoid robotics companies. Arc Institute introduced BioReason-Pro, a model that predicts function for the 99.9% of proteins lacking experimental annotations. Nectome has preserved a pig’s brain with minimal damage, now offering the technique to the terminally ill. AI optimization is extending to livestock and land. Halter’s AI-powered solar collars now let ranchers herd cattle via vibrations from an app, valued at $2 billion. Making land post-scarce, Coastal Assembly grew over 90 feet of new beach in six months at a Maldives resort using AI-optimized underwater structures to redirect sediment.

The same superintelligence that grows beaches and herds cattle is also selecting targets. Palantir’s Maven AI, which has carried out thousands of targeted strikes against Iran, will become an official program of record across the U.S. military. Pervasive data cuts both ways, though. Le Monde located France’s Charles de Gaulle aircraft carrier in real time through a sailor’s Strava profile.

The economy is metabolizing disruption at machine speed. Ironically, OpenAI plans to nearly double its headcount to 8,000 to sell the tools that replace humans. Capital is voting with its feet. Massachusetts residents fled with $4.2 billion in income after a millionaire surtax took effect. Walmart has patented ML-driven dynamic pricing. Europe, struggling to compete despite homegrown labs like Mistral, is trying to tax what it can’t build. Mistral’s own CEO argues AI companies should pay a content levy. Domestically, the White House released a national AI policy framework to preempt a 50-state patchwork, betting unified rules will outrun fragmented ones. At Meta and OpenAI, employees now compete on “tokenmaxxing” leaderboards, spending thousands a month automating work. Jensen Huang wants to go further, proposing AI tokens as a salary supplement and envisioning hundreds of thousands of Nvidia AI agents.

Every token consumed down here is a down payment on the terawatt up there.

(Disclosure: I have a financial interest in Coastal Assembly.)