Tech Digest – May 8, 2026
The Compute Race Goes Orbital
Anthropic Moves Into SpaceX’s Colossus, xAI Dissolved — and the Company That Planned for 10x Is Growing 80x
Anthropic has signed a partnership with SpaceX for full access to the Colossus 1 data centre in Memphis, Tennessee — more than 300 MW of capacity and over 220,000 NVIDIA GPUs, operational within weeks. The deal doubles Claude Code rate limits and eliminates peak-hour throttling for paid users. SpaceX confirmed the arrangement extends into an exploration of “multiple gigawatts of orbital AI compute,” positioning space-based data centres as a near-term engineering programme rather than a research concept.
The deal coincides with Elon Musk dissolving xAI as a separate company entirely, folding its products into a new “SpaceXAI” division. Musk acknowledged the original structure “was not built right,” and Anthropic is moving into Colossus 1 just as the freshly absorbed AI lab decamps for Colossus 2. Meanwhile, CEO Dario Amodei revealed Anthropic grew 80x on an annualised basis in Q1 — against a planned 10x — pushing the company to a $30 billion revenue run rate. On-chain pre-IPO trading instruments now imply a $1.2 trillion valuation, up 900% since October 2025 and surpassing OpenAI’s implied secondary-market price for the first time.
Note: When compute demand outgrows terrestrial power, land, and cooling simultaneously, the answer apparently is orbit. That framing may sound hyperbolic, but the underlying logic is not: the company generating $30 billion in annualised revenue from selling cognition just acquired 300 MW of capacity and immediately started planning for gigawatts more. Any institution modelling its own digital infrastructure needs is doing so inside a market shaped by this kind of demand.
Sources: Anthropic, SpaceXAI, CNBC, Fortune, VentureBeat
Silicon & Infrastructure Rebuilt for AI
Motherboard Sales Collapse 25% as Chip Wafers Redirect to AI — While Three Megaprojects Break Ground
PC motherboard sales have dropped more than 25% as chipmakers redirect wafer capacity from consumer silicon to AI accelerators, with Asus alone projected to sell 5 million fewer boards. The squeeze is structural, not cyclical — the same fabrication lines that once made desktop processors now stamp out data centre GPUs at higher margins.
The buildout consuming those wafers is accelerating on multiple fronts. SpaceX’s planned Terafab chip fabrication complex in Texas carries an estimated cost of $55 to $119 billion across phases. Arm doubled its AI-chip revenue guidance to $2 billion for 2027–2028, just one month after its first AI accelerator launched. Nvidia is investing $3.2 billion in three new Corning optical-fibre plants in the US, because copper interconnects have run out of bandwidth for AI-scale networking. And six companies — OpenAI, AMD, Broadcom, Intel, Microsoft, and Nvidia — jointly open-sourced MRC, a multipath networking protocol that keeps GPUs synchronised across cluster failures. South Korea’s stock market overtook Canada’s as the world’s seventh-largest, propelled by AI silicon demand.
Note: If your institution’s IT procurement assumes stable pricing and availability for standard server hardware, check that assumption. The same wafer capacity that made your last refresh affordable is now producing AI accelerators at 10x the margin. Delivery timelines and pricing for conventional compute are already shifting.
Sources: Tom’s Hardware, CNBC (Terafab), Financial Times (Arm), CNBC (Nvidia-Corning), OpenAI (MRC)
AI Capabilities & Agent Infrastructure
Opus 4.7 Tops New Refactoring Leaderboard — Then Reads Its Own Values Before Alignment Training
Claude Opus 4.7 scored 48.57 on Scale Labs’ new Refactoring Leaderboard, beating GPT-5.5 Codex on production-scale codebase restructuring. Separately, Anthropic published Model Spec Midtraining, a technique that lets models study their own values documentation before alignment fine-tuning — essentially reading the syllabus before the exam. For perspective on how far the ladder still extends, the harder ProgramBench — which asks agents to rebuild full codebases from a compiled binary alone — shows Opus 4.7 leading at 3% “almost resolved” and 0% fully solved.
Note: The alignment innovation may matter more than the benchmark. Models that internalise their own value specifications before fine-tuning is a qualitatively different approach to safety — less “bolted-on guardrails,” more “studied the ethics curriculum.” Whether it scales is the question; that it’s being tried at this stage of capability is worth watching.
Sources: TestingCatalog (Scale Labs), Anthropic Alignment, ProgramBench
Claude’s Agents Now “Dream” — Reviewing Sessions and Building Shared Memory Overnight
Anthropic launched “dreaming” in Claude Managed Agents: a scheduled overnight process in which agents review their session histories and curate shared memories across teams. The feature turns deployed agents from stateless tools into systems that accumulate institutional knowledge — learning which queries recur, which answers worked, and what context matters across an organisation.
Note: This is the difference between deploying a tool and deploying a colleague who remembers last week. For any organisation running AI agents at scale, the question shifts from “can it answer?” to “what has it learned?” — and who controls that accumulated knowledge.
Sources: Claude Blog
Digital Sovereignty
EU Weighs Restricting US Cloud for Sensitive Government Data — While Chrome Quietly Installs 4 GB of AI on Every Desktop
The European Commission is preparing its “Tech Sovereignty Package,” expected May 27, which would restrict US cloud platforms from processing sensitive government data in sectors including healthcare, finance, and judicial systems. The package includes the Cloud and AI Development Act (CADA) and Chips Act 2.0, targeting sovereign alternatives for public-sector compute. The restrictions are motivated in part by the US CLOUD Act, which allows American law enforcement to request data from US-based providers regardless of where it is physically stored.
The timing is pointed. Google has begun quietly deploying Gemini Nano — a 4 GB on-device AI model — to every Chrome desktop installation with available storage. No explicit opt-in. For any institution running a Chrome fleet, that is 4 GB per endpoint of a third-party AI model installed without procurement approval.
Note: One arm of the EU is drafting sovereignty legislation while a US platform silently installs AI on every browser it controls. These are two sides of the same dependency. IT departments managing institutional Chrome deployments should audit what just landed on their machines.
Sources: CNBC, Computing.co.uk, 9to5Google
Geopolitics & Governance
Washington and Beijing Weigh Official AI Guardrails Ahead of Trump-Xi Summit
The US and China are pursuing formal guardrails on AI development ahead of next week’s Trump-Xi summit, according to the Wall Street Journal. The discussions aim to prevent the AI competition from escalating into a broader crisis, with both sides exploring frameworks for responsible military AI use, compute export controls, and mutual transparency on frontier model capabilities.
Note: The EU’s AI Act was first-mover regulation. A US-China bilateral framework would be first-mover arms control. Brussels should be at that table — not because it builds frontier models, but because it regulates the market where those models deploy.
Sources: Wall Street Journal
Energy Transition
Texas Passes California in Utility-Scale Solar Capacity
Texas has overtaken California as the leading US state for utility-scale solar generation, producing 58,634 GWh from large-scale installations in 2025 versus California’s 53,713 GWh. Texas also leads in wind and energy storage. More than 40% of all new US utility-scale solar capacity planned for 2026 is in Texas — much of it driven by data centre power demand.
Sources: Inside Climate News, Governing, EIA
A single day’s news: one company acquires 300 MW of compute and starts planning for orbit. The chip supply chain pivots away from consumer hardware to feed it. The models it runs learn to review their own values and dream overnight. And the EU begins drafting the sovereignty rules that will determine whether European institutions ride this wave or watch it from a regulated distance. The infrastructure of intelligence is being rebuilt at every layer simultaneously — silicon, fibre, data centres, orbit, regulation. The institutions that see this as one interconnected transition, rather than a series of separate tech stories, are the ones planning for the right decade.
