Tech Digest – June 12, 2026
Capital & Infrastructure at Escape Velocity
SpaceX Prices History’s Largest IPO at $1.8 Trillion — To Build Data Centres in Orbit
SpaceX priced its IPO at $135 per share, valuing the company near $1.8 trillion — the largest public offering in history. More than $70 billion in retail orders poured in before trading began. Elon Musk stated the capital is for the “massive capital endeavour” of orbital data centres: AI compute facilities powered by round-the-clock solar, beyond Earth’s energy constraints and permitting battles. The demand behind the bet is measurable: Epoch AI found that the largest single data centre’s compute capacity has doubled every seven months since 2024.
The offering instantly reshapes the venture landscape. Founders Fund’s $600 million early bet is now worth over $50 billion; Andreessen Horowitz’s position exceeds $10 billion. Musk’s estimated net worth sits at $970.5 billion — one rally away from the world’s first trillionaire.
Note: When the largest IPO in history exists to escape Earth’s power grid, it quantifies a scale of compute demand that most infrastructure plans haven’t yet factored. The power and permitting bottlenecks driving this orbital bet are the same ones shaping data centre availability — and costs — on the ground.
Sources: Bloomberg, CNBC, Epoch AI
Chip Supply Diversifies as AI Reshuffles the Industrial Hierarchy
Memory chipmaker Kioxia overtook Toyota as Japan’s most valuable company, its stock up 600% this year on AI-driven NAND flash demand. The moment is symbolic: the industrial economy’s flagship displaced — even briefly — by a component supplier for AI infrastructure.
Meanwhile, Nvidia — locked out of China’s GPU market by export controls — is pitching Arm-based Vera CPUs to Chinese clients, a workaround that trades compute architecture for market access. And Google is courting Samsung to fabricate a future AI chip, moving beyond exclusive reliance on TSMC as capacity tightens globally.
Note: A memory company worth more than Toyota, an Nvidia CPU workaround for a GPU ban, and a Google fab deal that sidesteps TSMC — three moves in one day, all pointing the same direction. The semiconductor supply chain is mid-reorganisation. Procurement lead times and vendor optionality will be volatile for years.
Sources: Bloomberg, Reuters, The Information
Fusion Validated, Flash Charging Deployed, Solar Passes Coal
Commonwealth Fusion Systems published five peer-reviewed papers in the Journal of Plasma Physics validating the physics of its ARC fusion power plant — designed to produce 1.1 GW of fusion power, converting to 400 MW of continuous net electricity via high-temperature superconducting magnets and 15-minute fusion pulses. In parallel, BYD is deploying 1.5 MW Flash Chargers across the UK and Europe, adding 400 km of EV range in five minutes — commercial infrastructure, not a prototype.
And for the first time, US solar generation outproduced coal — a milestone projected for later this decade.
Note: Any one of these would be routine energy news alone. Together they compress three infrastructure timelines — baseload generation, transport charging, and grid mix — that most institutional energy plans haven’t updated for.
Sources: Ars Technica, The Verge, The Guardian
The Capability-Cost Squeeze
AI Reasoning Horizons Double Annually While Costs Race Toward Zero
A new study found that frontier AI models’ effective reasoning time horizon — the complexity of problems they can solve without chain-of-thought prompting — has doubled every year for six consecutive years. GPT-5.5 now reasons past the equivalent of three human-minutes of problem complexity, with projections reaching 25 minutes by 2030. At the same time, Claude Fable 5 sits well ahead of the previous token-efficiency frontier, delivering more capability per unit cost than any predecessor.
The commercial response is a price war. The Wall Street Journal reports OpenAI is weighing drastic price cuts in anticipation of a “war for users” with Anthropic — racing the cost of frontier intelligence toward zero. Separately, the AI Alliance’s Project Tapestry now lets nations co-train a shared frontier model while keeping data sovereign, opening multilateral access to capability that was recently exclusive to a handful of companies.
Note: Capability doubling annually while costs collapse is not a product cycle — it’s a structural shift in what can be automated and at what price. Any cost-benefit analysis for AI-adjacent projects written more than twelve months ago is already stale.
Sources: arXiv, Wall Street Journal, AI Alliance
Cybersecurity’s AI Arms Race
AI Finds More Bugs, Creates Better Exploits, and Gets Gamed by Malware Authors
Microsoft’s latest Patch Tuesday was notably larger than usual — the company attributes the increase to AI-powered code scanning surfacing vulnerabilities in even its most mature codebases. On the offensive side, Epoch AI assessed the Mythos model family as a “genuine leap in exploit development,” performing seven months ahead of trend projections for AI-assisted cyberattacks.
Malware authors are adapting. Security researcher John Scott-Railton of Citizen Lab flagged a new adversarial tactic: embedding text about nuclear and biological weapons inside spyware code to trigger AI security scanners’ safety refusals — effectively blinding the tools designed to catch them.
Note: Defensive AI finds more holes. Offensive AI exploits them faster. And adversaries have already learned to weaponise AI’s own guardrails against it. Cybersecurity budgets built on pre-AI threat models are funding yesterday’s defences.
Sources: Microsoft, Epoch AI, John Scott-Railton (Citizen Lab)
Governance & the Autonomy Line
Fully Autonomous Drones Kill Soldiers With No Human in the Loop
Alexander Kokhanovsky, CEO of Ukrainian drone manufacturer Aero Center, confirmed to New Scientist that fully autonomous drones have killed soldiers on the battlefield — a first. The one-off test involved ten AI-controlled quadcopters programmed to fly 3 to 5 km to the front line near Bakhmut, then activate “Terminator mode”: an AI system that searched for and engaged targets with no human authorisation. Surveillance drones sent afterward found dead Russian soldiers in the engagement area.
Note: The line between human decision and machine decision in lethal force just moved from theoretical to confirmed. Every defence ministry, ethics board, and international law framework built around “meaningful human control” now confronts a precedent, not a hypothetical.
Sources: New Scientist
German Court Rules Google’s AI Overviews Are Its Own Words — And Liable for Them
The Regional Court of Munich issued a temporary injunction declaring Google’s AI Overviews to be Google’s own speech — not neutral search results. The ruling found that AI-generated summaries produce “independent, new, and substantive statements” for which Google bears direct liability. The case involved two Munich publishers whose names Google’s AI wrongly linked to scams and “dubious business practices,” fabricating connections that appeared in none of the cited sources. Google has confirmed it will appeal.
Note: First ruling to answer the question every institution deploying AI-generated content will eventually face: when the AI gets it wrong, whose words are they? Munich’s answer — the company that built it. Any public body publishing AI-generated citizen-facing content now has a live liability precedent to consider.
Sources: The Decoder, The Next Web
Amodei Calls for FAA-Style AI Testing — Days After His Own Company’s Governance Lapse
Anthropic CEO Dario Amodei published “Policy on the AI Exponential,” arguing that AI’s pace now outstrips political institutions’ ability to respond and calling for mandatory third-party testing of frontier models above a compute threshold — with government authority to block unsafe releases. The audits would cover cybersecurity, biological weapons, loss of control, and automated AI research. Anthropic pledged $350 million alongside the essay: a $200 million economic research fund and a $150 million fellowship programme.
The timing underscores the urgency. Days earlier, researchers discovered a clause in Anthropic’s own Fable 5 system card that would have permitted poisoning of AI research. Anthropic reversed the policy within 48 hours and apologised — but the incident illustrates how quickly governance gaps can emerge even at safety-focused companies.
Note: The CEO of a leading AI company is asking governments to have the power to stop him. The practical proposal — mandatory testing with blocking authority — is closer to aviation regulation than anything the tech industry has previously accepted. The EU’s AI Act already mandates conformity assessments for high-risk systems. Amodei’s argument is that even that framework may be too slow for what’s coming.
Sources: Dario Amodei, Wired
AI Agents Enter Commerce
Visa Inside ChatGPT, Coinbase for Agents — AI Gets Payment Rails
Visa has integrated checkout capabilities directly inside ChatGPT, allowing users to complete purchases from any merchant without leaving the conversation. Separately, Coinbase launched “Coinbase for Agents” — infrastructure enabling AI agents to execute cryptocurrency trades within user-defined limits, no human approval required per transaction.
Note: Two of the world’s largest payment networks just gave AI agents the ability to spend money. The governance questions are no longer theoretical: who authorises the purchase, who bears liability for the error, and what audit trail exists when a machine executes a financial transaction on someone’s behalf?
Investing in Human Capacity
Anthropic’s $150M Claude Corps and Google’s 300,000 Trade Workers — Bottleneck Bets on People
Anthropic launched Claude Corps, a $150 million fellowship embedding 1,000 early-career workers in nonprofits to help organisations adopt AI tools — $85,000 annual salary, no degree required, applications open now. Separately, Google announced funding for 300,000 trade workers to build the physical infrastructure AI demands: data centres, power systems, cooling networks. Both companies are also expanding in London as the AI industry’s geographic footprint widens.
The investments arrive as “Europe 2031,” a coalition of AI researchers and entrepreneurs, warned that nations risk missing the competitiveness window if workforce development and AI adoption don’t accelerate now.
Note: One programme supplies the people who teach organisations to use AI. The other supplies the people who build the buildings it runs in. Both are bottleneck bets — the software won’t matter without trained operators, and the models won’t matter without physical infrastructure. The Europe 2031 warning is aimed squarely at this gap.
Sources: Anthropic, Axios, Europe 2031
The day’s clearest signal isn’t any single headline — it’s the distance between them. The largest IPO in history funds data centres in orbit. A Munich court rules the AI they’ll run is the company’s own speech. Autonomous drones kill without orders and chip companies eclipse car manufacturers. The infrastructure of intelligence — physical, financial, legal, lethal — is being built faster than the institutions meant to govern it.
