Tech Digest – May 28, 2026
The AGI Threshold
DeepMind CEO Calls 2029 Plausible for AGI — AI Mathematicians Already Publishing Peer-Reviewed Proofs
DeepMind CEO Demis Hassabis told Axios that while 2030 remains his baseline for AGI, 2029 is now plausible — and that current agentic deployments are “a little bit like a practice run” for what follows. He warned that economists are “still not taking this seriously enough” and confirmed that all leading labs are focused on recursive self-improvement. The evidence is accumulating independently: Axiom reported that eight papers authored by its AxiomProver system have appeared on arXiv since February, with five accepted at peer-reviewed mathematics journals covering algebraic geometry, number theory, and combinatorics.
Note: When the CEO of the lab that built AlphaFold tightens his timeline and names recursive self-improvement as the shared priority across labs, the statement is both a prediction and a disclosure. Peer-reviewed AI mathematics — theorems passing the same scrutiny as human proofs — removes the question of whether AI can produce novel research. The remaining question is pace.
Compute Sovereignty & Infrastructure
Nvidia Vera ARM CPU Outperforms x86 Leaders — Company Spending $150 Billion a Year on Taiwanese Supply Chain
Nvidia’s Vera CPU, an 88-core ARM64 processor, posted the highest performance scores ever recorded on an ARM chip in initial benchmarks: 10% above AMD’s EPYC 9575F and 1.55x Intel’s Xeon 6980P flagship. The tests were conducted under Nvidia-controlled conditions with pre-production hardware — open benchmarking and efficiency metrics were restricted. Separately, Jensen Huang confirmed Nvidia is spending up to $150 billion annually on its Taiwanese supply chain and scaling local headcount to 4,000.
Note: Even with the benchmark caveats, ARM matching or exceeding x86 at the data centre tier reshapes procurement assumptions. Institutions specifying server hardware now face a market where the dominant AI accelerator vendor also produces the fastest general-purpose CPU — and its entire manufacturing pipeline runs through Taiwan. The architecture shift and the supply chain concentration are two sides of the same de-risking question.
Sources: Phoronix, Nikkei Asia
Lombardy Hikes Data Centre Construction Fees Up to 200% in Green Zones
Lombardy became the first Italian region to impose surcharges on data centre construction: 100% in agricultural areas, 200% in parks and protected green zones. The region accounts for 63% of all data centre permit applications submitted across Italy, which prompted the measure. Operators are being steered toward disused industrial sites instead.
Note: Lombardy didn’t ban data centres — it repriced where they go. As the AI buildout intensifies, expect more European regions to reach for planning fees, zoning surcharges, and energy quotas as steering mechanisms. Any operator scouting European locations needs to price in regulatory surface area alongside land and power.
Sources: Il Sole 24 Ore
EU Proposes Satellite Spectrum Rules — Most Licences Reserved for Local Operators
The European Commission proposed satellite spectrum rules that would allow SpaceX’s Starlink to bid for direct-to-mobile airwaves while reserving the majority of licences for European operators. The proposal lands as satellite connectivity scales globally — American Airlines announced it will equip more than 500 narrow-body aircraft with Starlink.
Note: The spectrum framework encodes the EU’s broader approach to tech sovereignty: let global platforms compete, but structurally favour domestic capacity. For institutions planning rural connectivity or emergency communications, the signal is which satellite providers will have guaranteed spectrum access — and which will operate on borrowed bandwidth.
AI Governance Takes Shape
Illinois Passes First-in-Nation AI Safety Audit Mandate — 110-0
Illinois passed SB 315 with a unanimous 110-0 vote, requiring frontier AI developers with over $500 million in annual revenue to publish catastrophic-risk plans, undergo mandatory annual third-party safety audits, and report incidents within 72 hours. Penalties reach $3 million per violation. The bill includes whistleblower protections. Governor Pritzker has indicated he will sign. OpenAI and Anthropic publicly support the measure.
Note: The EU has the AI Act. Illinois now has enforceable external audits with real penalties. The pattern is converging across jurisdictions: self-reporting gives way to mandated third-party verification. YouTube’s parallel announcement — automatically detecting and labelling AI-generated content rather than relying on creator disclosure — shows the same shift from voluntary to structural accountability, this time at the platform layer.
Robinhood Opens Trading Platform to Autonomous AI Agents
Robinhood launched agentic trading in beta, allowing customers to connect AI agents via MCP (Model Context Protocol) to a dedicated brokerage account. Agents can analyse portfolios, place stock trades, and execute strategies autonomously. The system works with Claude, ChatGPT, Codex, and Cursor. Users receive notifications of agent-executed trades. Options, crypto, and prediction markets are planned next.
Note: This is the first major retail brokerage to make its transaction stack agent-accessible. The regulatory framework for autonomous financial agents doesn’t exist yet — not in the US, not under MiFID II. When this pattern reaches EU brokerages, financial regulators will need to decide whether an agent placing trades constitutes algorithmic trading, investment advice, or something that doesn’t fit either category.
Sources: Robinhood, TechCrunch
Workforce Reconfigurations
NYT Union Challenges AI Performance Monitoring — OpenAI Foundation Commits $250 Million for Worker Transition
Unionised technology workers at the New York Times filed a complaint alleging the newspaper is using AI to monitor employee performance in breach of their collective bargaining agreement — one of the first major union challenges to algorithmic management in a newsroom. Separately, the OpenAI Foundation committed $250 million to forecasting AI’s economic impact and funding programmes to help workers through the transition.
Note: Labour is pushing back on one end while capital funds transition research on the other. The NYT case will test whether existing employment contracts can constrain algorithmic supervision — a question that extends well beyond journalism. Every organisation deploying AI-driven performance tools should be tracking this, because the legal boundaries of algorithmic management are being drawn now.
Sources: The Verge, OpenAI Foundation
Smaller Consultancies Report Up to 50% Growth as AI Levels the Playing Field
A Financial Times analysis found smaller management consultancies reporting revenue growth of up to 50%, with partners attributing the acceleration to AI tools that allow lean teams to deliver work previously requiring large-firm headcount. The shift is pulling clients toward specialised, senior-led alternatives.
Note: The procurement implication is direct: the vendor pool just got deeper. Institutions accustomed to shortlisting only large firms may find that a three-person consultancy with strong AI tooling delivers comparable output at a fraction of the cost. Evaluation criteria weighted toward firm size are increasingly misaligned with delivery reality.
Sources: Financial Times
Cybersecurity Under Pressure
AI Bug-Finders Push Linux Vulnerability Rate to 13 Per Day — Rust Positioned as the Fix
At Rust Week in Utrecht, Linux stable kernel maintainer Greg Kroah-Hartman revealed that AI-powered vulnerability scanners are surfacing new classes of kernel bugs — including previously unknown categories he named Dirty Frag, Copy Fail, and Fragnesia — pushing CVE issuance to roughly 13 per day. Kroah-Hartman positioned Rust adoption within the kernel as the structural response to code that C can no longer safely maintain.
Note: The vulnerability discovery rate isn’t just rising — it’s being accelerated by AI tools that find bugs faster than maintainers can patch them. Any institution running Linux infrastructure faces compressing patch cycles and an expanding attack surface. Rust addresses the root cause for new code, but decades of existing C remain exposed.
Sources: ZDNet
Today’s developments compress on one question: how fast can institutions adapt when the systems they depend on are being rebuilt simultaneously? AGI timelines tighten while AI systems publish original mathematics. ARM chips overtake x86 while the supply chain concentrates in a single geography. Regulators in Illinois draft audit mandates while a brokerage hands trading to autonomous agents. And the consultancy market fractures — not because large firms are failing, but because small ones no longer need their scale. The consistent signal across every category: the organisations adapting fastest aren’t necessarily the largest. They’re the ones that stopped treating the current pace as temporary.
