Tech Digest – March 22, 2026
Compute at Civilizational Scale
Musk Announces TERAFAB — A Vertically Integrated Chip Foundry Targeting a Terawatt of Compute Per Year
SpaceX and Tesla jointly unveiled TERAFAB, a vertically integrated semiconductor fabrication plant in Austin, Texas, targeting over a terawatt of annual compute output — roughly 80% for space applications and 20% for terrestrial use. The facility will fabricate 2-nanometre chips in two categories: edge silicon for Tesla’s robotaxis and Optimus robots, and high-power chips for SpaceX orbital AI satellites and xAI. With 100,000 wafer starts per month planned and every stage from mask design to testing under one roof, the $20–25 billion project anticipates small-batch production in 2026 and volume output in 2027.
Note: The institutional signal isn’t the factory — it’s what it means for chip supply. A new foundry producing 2-nm chips outside the TSMC/Samsung duopoly changes the procurement map for anyone buying compute-dependent technology over the next decade.
Sources: Bloomberg, SpaceX, Elon Musk
SoftBank Plans a $500 Billion, 10-Gigawatt Data Centre Campus on a Former Uranium Enrichment Site
SoftBank Group is developing an AI data centre campus in Piketon, Ohio, on 3,700 acres of a former U.S. Department of Energy uranium enrichment complex. Masayoshi Son described it as a $500 billion project. The first phase — 800 megawatts, costing $30–40 billion — is expected by early 2028, with full 10-gigawatt capacity targeted by the end of the decade. Gas turbines capable of generating 9.2 GW have already been sourced. The investment is part of a broader $550 billion Japan-U.S. commitment linked to trade negotiations.
Note: Ten gigawatts is roughly what Belgium consumes. When a single data centre campus requires the electrical output of a small European country, every institutional assumption about energy planning, grid capacity, and digital infrastructure costs needs revisiting.
Defence AI & Its Weak Links
Pentagon Formalises Palantir’s Maven AI as a Core Military System — While a Fitness App Exposes a French Aircraft Carrier
The U.S. Department of Defense will designate Palantir’s Maven AI as an official program of record by September 2026, locking in multi-year funding and expanding the platform from $480 million in 2024 to $13 billion. Maven now serves over 20,000 active users across all military branches, using AI to analyse battlefield data from satellites, drones, and intelligence feeds to identify targets. Meanwhile, Le Monde demonstrated the other edge of pervasive data: journalists located France’s Charles de Gaulle aircraft carrier in real time — 100 km off the Turkish coast near Cyprus — after a naval officer logged a 7 km run on the flight deck using Strava. Satellite imagery confirmed the carrier’s position within minutes of the upload.
Note: A $13 billion AI targeting system and a free fitness app operating on the same data infrastructure. One finds targets; the other makes you one.
Sources: Reuters, Le Monde, Tom’s Hardware
Super Micro Co-Founder Charged With Smuggling $2.5 Billion in Nvidia AI Chips to China
Yih-Shyan “Wally” Liaw, co-founder and board member of Super Micro Computer, was arrested for allegedly orchestrating a scheme to divert $2.5 billion worth of Nvidia AI servers to China in violation of U.S. export controls. Servers were assembled in the U.S., shipped to Super Micro’s Taiwan facility, re-routed through a Southeast Asian intermediary, repackaged in unmarked boxes, and delivered to Chinese buyers. Super Micro shares dropped 33% on the news. The company itself is not named in the indictment.
Note: Export controls are only as strong as the supply chain is transparent. For any institution procuring AI hardware through intermediaries, this case is a due-diligence wake-up call.
The Web After Humans
Cloudflare CEO: Bot Traffic Will Surpass Human Traffic Online by 2027
Matthew Prince, CEO of Cloudflare, predicted at SXSW that AI bot traffic will exceed human traffic on the internet by 2027. Before the generative AI era, bots accounted for roughly 20% of web traffic. The shift is driven by AI agents that visit orders of magnitude more pages than humans — where a person might check five websites when shopping, an AI agent may visit 5,000. Cloudflare has documented blocking 416 billion AI bot requests since mid-2025.
Note: When most of your website visitors aren’t human, your analytics, accessibility compliance, security model, and even your definition of “engagement” stop meaning what they used to.
Sources: TechCrunch
OpenAI Plans a Desktop ‘Superapp,’ WordPress Opens to AI Agents, Google Replaces News Headlines with AI Text
Three platform shifts in one week. OpenAI is building a desktop application that merges ChatGPT, its Codex coding tool, and a built-in browser into a single “Superapp.” WordPress.com now allows AI agents to draft, edit, and publish posts autonomously — no human in the loop required. And Google Search has begun replacing news outlet headlines with AI-generated text in some results, turning the index itself into a generative layer that rewrites what publishers wrote.
Note: The three dominant channels through which institutions reach citizens — search, publishing, and productivity tools — are all being rewritten in the same week. Communications strategies built around any of these platforms are now operating on shifting ground.
Sources: WSJ, TechCrunch, The Verge
Research Enters the Machine
OpenAI Targets a Fully Automated AI Researcher by 2028 — While AI Already Generates Hundreds of Novel Math Problems
OpenAI has declared building an autonomous AI researcher its “North Star,” with an automated research intern targeted for September 2026 and a full multi-agent system by March 2028. The effort, led by chief scientist Jakub Pachocki, spans reasoning, agents, and interpretability. The timeline is not aspirational — the tools are already producing results. A separate team built an agent that generated 665 novel research problems in differential geometry, many unknown to experts. Terry Tao, widely regarded as the world’s leading mathematician, observed that even high school students can now make real contributions to frontier mathematics using AI tools.
Note: When a Fields Medal winner says the floor has risen, the floor has risen. Research funding models, university admissions criteria, and institutional R&D staffing all face the same question: what does “original contribution” mean when the tools can generate hundreds of novel problems overnight?
Sources: MIT Technology Review, arXiv, Andrew Curran / Terry Tao
Regulating What’s Already Moving
White House Releases National AI Policy Framework to Preempt a 50-State Patchwork
The Trump administration published a four-page National AI Policy Framework calling on Congress to legislate across seven areas: child safety, community protections, intellectual property, free speech, innovation, workforce development, and federal preemption of state AI laws. OSTP Director Michael Kratsios described it as ensuring “one national AI framework, not a 50-state patchwork.” The document is not binding but pairs federal preemption with bipartisan state-level priorities like utility ratepayer protections — framed as a deliberate “give and take.”
Note: The EU has the AI Act. The U.S. now has a framework. China has sector-specific rules. Three regulatory regimes, three philosophies — and any institution operating across borders or procuring from global vendors now navigates all three simultaneously.
Sources: White House, CNN, CNBC
Mistral’s CEO Proposes a Revenue-Based Content Levy for All AI Companies Operating in Europe
Arthur Mensch, CEO of Mistral — Europe’s largest AI company — proposed in the Financial Times that all AI vendors operating in Europe pay a revenue-based levy for training on publicly available content. The suggested rate of 1–1.5% of revenue would apply equally to foreign companies, creating what Mensch calls a level playing field within the European market. Funds would support culture and new content creation in Europe, addressing what Mensch describes as a “fragmented legal environment” that disadvantages European developers compared to US and Chinese competitors operating under permissive copyright rules.
Note: A European AI CEO proposing a tax on his own industry is unusual enough to read twice. The subtext: Mistral can absorb 1.5%. OpenAI and Google, training on European content while operating under permissive U.S. copyright rules, would pay disproportionately more. It’s an industrial policy proposal dressed as cultural stewardship.
Sources: Financial Times, AI Business
OpenAI Plans to Nearly Double Headcount to 8,000 — While Its Employees Compete to Automate Their Own Jobs
OpenAI plans to grow from approximately 4,200 to 8,000 employees — hiring aggressively to sell the tools that automate human work. Internally, the irony compounds: employees at Meta and OpenAI now compete on “tokenmaxxing” leaderboards, spending thousands of dollars per month on AI tools to automate their own tasks. Nvidia CEO Jensen Huang has taken the logic a step further, proposing AI compute tokens as a salary supplement and envisioning hundreds of thousands of Nvidia AI agents operating as a virtual workforce.
Note: The company building the replacement hires more humans. The humans it hires compete to replace themselves faster. And the chip company supplying the compute wants to pay people in tokens. If this reads like satire, it’s because the labour market is moving faster than the narratives we have for it.
Sources: Financial Times, New York Times, CNBC
The Robotics Race Is Already Running
China Ships 90% of the World’s Humanoid Robots — With 140 Companies and $26 Billion in Government Backing
China’s Ministry of Industry and Information Technology reports over 140 domestic humanoid robot manufacturers, with more than 330 models released. Chinese companies produced an estimated 87–90% of the 13,000 humanoid robots shipped globally in 2025, led by Shanghai-based AgiBot with 5,168 units and 39% global market share. Beijing and other cities have established government investment funds exceeding $26 billion to finance the sector, and China has now published its first national standards for humanoid robots to support industry scale-up.
Note: The EU’s robotics strategy reads differently when one country already controls 90% of global humanoid shipments and is setting the national standards others will either adopt or compete against.
Sources: The Guardian, TechCrunch, Robotics and Automation News
