Tech Digest – May 4, 2026
Governance at Two Speeds
UAE Directs 50% of Federal Operations to Run on Agentic AI Within Two Years
The United Arab Emirates has issued a federal directive requiring 50% of government operations to run on agentic AI within two years. The mandate covers administrative processes across federal agencies and represents the most aggressive national AI deployment target set by any government to date.
Note: A 50% target in two years doesn’t just signal ambition — it creates a reference point. Every EU digital strategy document that still says “explore AI opportunities” now reads as a planning gap, not a plan.
Sources: MIT Sloan Management Review ME
South Africa Withdraws National AI Policy After Discovering It Was Written by AI
South Africa’s communications minister withdrew a draft national AI policy after it emerged the document had been generated by AI. The draft contained fictitious academic citations — fabricated references that passed initial review and reached the public consultation stage before the errors were caught.
Note: The failure wasn’t that someone used AI to draft policy. It’s that no one caught the fabricated citations before publication. Any institution using generative AI for document production faces the same risk — and the verification step that catches it cannot itself assume the output is trustworthy.
Sources: TimesLive
Capital & Infrastructure
Hyperscaler AI Capex Reaches $805 Billion in 2026 — Equal to All Other S&P 500 Capital Spending Combined
Morgan Stanley has again raised its forecast for the five hyperscalers (Amazon, Alphabet, Meta, Microsoft, and Oracle), now projecting $805 billion in capital expenditure for 2026 and $1.1 trillion for 2027. The revised figures mean AI infrastructure spending by five companies now roughly equals all non-tech S&P 500 capex combined.
US AI czar David Sacks cited a Morgan Stanley analysis showing AI accounted for 75% of Q1 GDP growth, with a 2.5% capex tailwind this year rising to over 3% next year. Yet much of this compute remains underutilised: xAI is reportedly running just 11% of its 550,000 Nvidia GPUs, compared to Meta and Google’s 43-46% utilisation rates — a reminder that scaling these clusters is an engineering challenge as much as a capital one.
Note: When five companies’ infrastructure budgets match the capital spending of the entire rest of the S&P 500, the digital infrastructure that runs institutional services is being shaped by decisions made in five boardrooms. That concentration of investment is also a concentration of dependency.
Sources: Holger Zschaepitz (Morgan Stanley data), David Sacks, WCCFTech
Japan’s Data Centres Go Vertical: 52-Metre Towers Rise in Tokyo Parking Lots
Japan’s $23 billion data centre market is set to grow 50% by 2030, with providers building 52-metre towers on urban Tokyo parking lots to meet demand where horizontal space is exhausted. The land constraint is pushing infrastructure in unexpected directions elsewhere too: Starcloud, which builds solar-powered data centres in low Earth orbit, is in talks for a $2.2 billion valuation — double its $1.1 billion close just one month ago.
Note: When data centres go vertical in Tokyo and orbital over the Earth, the binding constraint on digital infrastructure isn’t compute or capital — it’s real estate. EU institutions planning sovereign cloud or data localisation should note that the location premium is already here.
Sources: Financial Times, The Information
The AGI Threshold
AI Agent Produces Original Mathematical Proofs That Leading Experts Call Elegant — Brockman Estimates 70-80% to AGI
Harmonic’s formal reasoning agent Aristotle is now solving recently posed research problems in number theory with proofs that leading mathematicians call correct, novel, and elegant. The proofs are formally verified in Lean 4, meaning they are not just convincing — they are machine-checked to be mathematically valid. Additive number theorist Melvyn Nathanson described the agent’s work as containing original ideas of its own.
OpenAI president Greg Brockman, speaking at Sequoia’s AI Ascent conference, estimated that AI is now 70-80% of the way to AGI by his personal definition, adding that AI has gone from writing roughly 20% to 80% of OpenAI’s code in a single month. Sam Altman, meanwhile, said that despite the appeal of cheaper and faster models, making them smarter remains the most important priority — and urged users to prepare for the next major capability leap after GPT-5.5. Demis Hassabis, a former chess prodigy, now plays chess against Gemini to trace its chain-of-thought, noting when the model sees a blunder, searches for better options, and plays the blunder anyway.
Note: An AI that produces original, formally verified mathematics isn’t pattern-matching. It’s reasoning in a domain where correctness is provable, not debatable. “AI can’t really reason” is losing its usefulness as a planning assumption.
Sources: Pietro Monticone (Nathanson quote), Benzinga (Brockman), Sam Altman, Vitrupo (Hassabis)
NIST Evaluates China’s Best AI Model at Eight Months Behind the US Frontier
The US National Institute of Standards and Technology’s Center for AI Standards and Innovation (CAISI) has evaluated DeepSeek V4 Pro — the most capable Chinese AI model to date — as lagging the US frontier by approximately eight months. CAISI’s non-public benchmarks place DeepSeek V4 at the level of GPT-5, which was released in late 2025. Independent analysis suggests the gap may be wider when adjusted for token usage and evaluation freshness. Notably, DeepSeek’s own reported scores exceeded CAISI’s findings.
Note: Eight months is a technical gap. But DeepSeek V4 Pro is also more cost-efficient than US alternatives on most benchmarks. For EU institutions evaluating AI procurement, the question is shifting from “which model is best?” to “which model fits the budget and the risk profile?” — and that calculation may favour the model that’s good enough and cheaper.
Sources: NIST, @scaling01 (gap analysis)
Robots Enter the Workforce
Boston Dynamics Scales from Four Humanoids a Month Toward Tens of Thousands as China Deploys Robots for Retail
Hyundai is pressing Boston Dynamics to scale production of its Atlas humanoid from four units per month to the tens of thousands needed across its carmaking plants, with a new manufacturing facility set to open in the coming months. The push has triggered a c-suite exodus at Boston Dynamics, as the company’s research culture collides with the demands of industrial-scale production.
In China, the transition is already visible. Over the May Day holiday, humanoid robots autonomously ran retail kiosks for tourists, handling customer-facing service roles without human supervision. The deployments are framed not as demos but as operational substitution — filling gaps in a service economy short on labour.
Note: Four humanoids a month is a research lab. Tens of thousands is a supply chain. The organisational pain at Boston Dynamics — leadership departures, forced scaling — is what the transition from prototype to fleet looks like. Hyundai isn’t asking for better robots; it’s asking for more of them, now.
Sources: Semafor, CyberRobo (video)
Synthetic Creativity Goes Commercial
Suno Hits $300 Million Annualised Revenue With Two Million Paying Users for AI-Generated Music
AI music generation platform Suno now has over two million paying subscribers and $300 million in annualised revenue. The company’s growth — from launch to commercial scale — compressed the kind of market adoption that took streaming platforms years into months.
Note: Two million people paying for AI-generated music didn’t happen over years of gradual adoption. It happened in roughly the time it takes most institutions to finish a procurement cycle. The velocity of adoption, more than the product itself, is what makes this a planning signal.
Sources: Forbes
Today’s threads pull in one direction. A government sets a two-year deadline for AI-run operations while another government can’t tell the difference between a human-written policy and a machine fabrication. Five companies spend more on AI infrastructure than the rest of corporate America spends on everything. An AI agent produces mathematics that experts call beautiful. Humanoid robots move from lab curiosity to factory production line. And an AI music platform reaches commercial scale faster than most institutions can draft a requirements document. The gap between those who are moving and those who are still planning to plan is no longer a strategy question. It’s a calendar one.
