Tech Digest – August 4, 2026
Infrastructure at Continental Scale
$10 Billion for Norwegian Compute, $150 Billion in Chip Financing, and GPUs Heading to Orbit
Caterpillar posted a record $20.5 billion quarter, with power generation revenue up 72% on data centre demand and its order backlog hitting $72 billion — up 92% year-on-year. The physical economy is being reshaped by AI infrastructure requirements, from turbines to the racks they power.
Anthropic signed a $10 billion, six-year compute deal with Nvidia-backed Volta, securing 121 MW of Vera Rubin GPU capacity at Bitdeer’s hydro-powered Tydal site in Norway — delivery beginning by year-end. Google, meanwhile, assembled a $150 billion off-balance-sheet financing programme — drawing in Apollo, Blackstone, and Morgan Stanley — to route its TPU chips to Anthropic via the largest private credit structure ever built around AI hardware. Amazon crossed $3 trillion in market capitalisation on its fastest AWS growth since 2021.
The infrastructure is also leaving the planet. SpaceX and Nvidia announced the Starmind AI1 satellite: Rubin GPUs and Vera CPUs designed for datacenter-class compute in low Earth orbit, with power boosted 67% to 250 kW per satellite and a target constellation of up to one million units.
Note: Caterpillar sells the generators, Google arranges the financing, Anthropic fills the Norwegian racks, and SpaceX flies the overflow to orbit. Any institution planning digital infrastructure on a five-year horizon is operating in a market shaped by these commitments.
Sources: CNBC (Caterpillar), Bloomberg (Anthropic-Volta), Financial Times (Google), SpaceX
Palantir Posts 93% Revenue Growth to $1.94 Billion as Government AI Spending Surges
Palantir reported Q2 revenue of $1.94 billion, up 93% year-on-year, with US government revenue climbing 90% to $809 million and US commercial revenue surging 149%. The company raised its full-year guidance to $8.15 billion, implying 82% annual growth — well above Wall Street consensus. CEO Alex Karp used the moment to accuse leading AI labs of “trying to drug addict us to a future they believe they control,” positioning Palantir as the sovereign alternative.
Note: The 93% isn’t a tech stock vanity metric — it’s a measure of what government AI procurement actually looks like when it accelerates. EU institutions comparing their digital transformation budgets to what US counterparts are channelling through a single vendor may find the gap instructive.
Sources: CNBC, CNBC (Karp)
The Self-Improvement Thesis
AI Agents Optimise Their Own Inference, Beat Human Tuners, and Rebuild Software From Scratch
Asari AI’s self-improving “co-inventor” agents rebuilt the inference stack for DeepSeek v4 Pro and GLM 5.2 on B200 GPUs, lifting throughput up to 16%. Intology’s Locus agent leads PostTrainBench, post-training models unsupervised in ten H100-hours and outperforming human tuners on the harder variant. On MirrorCode — which tests whether agents can rebuild entire software projects from scratch and pass every test — Claude Fable 5 solves 64% versus GPT-5.6 Sol’s 20%.
DeepMind’s chief strategy officer stated explicitly that recursive self-improvement — AI building better AI — is what justifies the unprecedented capital expenditure now flowing into compute infrastructure.
Note: This is the logic behind the hundred-billion-dollar infrastructure bets. Not better chatbots — machines that improve machines. When AI can optimise its own inference and train its own successors, the return on compute compounds. Timelines built on human-paced development are already obsolete.
Sources: Asari AI, Intology, Epoch AI (MirrorCode), The Information (DeepMind)
Mathematics Submissions Spike 20% Above Trend While Budgets Flatline
Monthly maths submissions to arXiv have hit 5,274 — roughly 20% above where 15 years of steady growth would place them — while global mathematics research budgets have been flat or declining. The signal is sharpest through VibeMathed, which tracks AI contributions to open problems: 427 problems tracked, 312 resolved, 205 in July alone — a 193% increase over June, with a third formally verified in Lean.
In wet labs, the same acceleration loop runs with slower atoms. Tianjin University’s REAP system — pairing a hybrid-loss model with robotic experiments — achieved a 57-fold activity gain in cytochrome P450 BM3 in five cycles and 104-fold in Sortase A, published in Nature Communications.
Note: Fifteen years of steady growth in maths submissions, then a sudden 20% break upward — with flat budgets. For anyone who funds research or evaluates research institutions, the productivity baseline just shifted.
Sources: VibeMathed, Nature Communications (REAP)
Geopolitics of Compute
Chinese Memory Enters Major PCs as Huawei Warns of Western Chip Limits and the FCC Drafts a Ban
Three fault lines opened simultaneously. Major PC makers including HP, Asus, and Acer have started using DRAM from China’s CXMT, whose IPO surged 466% on its first trading day and whose profits rose 2,530% year-on-year. CXMT is on track to match Micron’s production capacity in 2026, positioning China as the world’s second-largest DRAM production base.
Huawei’s top semiconductor scientist Liao Heng warned that Western die and HBM scaling is approaching a “cascade failure” point, pitching Huawei’s Tau Scaling Law — which prioritises signal transmission speed over transistor shrinkage — as the alternative path. The first implementation arrives this autumn in Huawei’s next Kirin chips.
Meanwhile, the FCC is drafting a ban on imports of new Chinese data centre optical transceivers, extending the pattern of targeted US restrictions on Chinese drones, routers, and robotics components.
Note: The procurement assumption that chip supply chains are global — messy but navigable — is being replaced by something more binary. Institutions planning digital infrastructure aren’t just choosing vendors anymore; they’re choosing sides of a supply chain that’s splitting in real time.
Sources: The Information (CXMT), Bloomberg (Huawei), Reuters (FCC)
US AI Policy Reverses in Weeks After Nvidia-Led Industry Revolt
US officials weighed sanctions and blacklisting against open-source Chinese AI labs, then reversed course after Nvidia CEO Jensen Huang led a coalition of over two dozen companies — including Microsoft, Meta, and Palantir — lobbying against the restrictions that OpenAI and Anthropic had advocated for. The administration chose competitiveness over containment.
Separately, the completed voluntary AI evaluation framework — developed with input from lab staffers — remains undisclosed, with officials stating that “unclassified doesn’t mean we are going to broadcast them to everyone.” Beijing has expressed concern ahead of Xi Jinping’s September visit that the US military AI system Mythos may constitute an offensive cyber weapon.
Note: US policy reversed in weeks, the evaluation framework exists but won’t be shared, and the next pivot may come after a single summit. For EU institutions navigating the transatlantic AI relationship, the signal is that planning around a stable US AI policy posture means planning around something that doesn’t currently exist.
Sources: New York Times, Axios (framework), Bloomberg (Mythos)
Software Economics Repriced
SaaS Valuations Collapse from 18x to 3.4x Revenue as the Cost of Building Software Trends Toward Zero
Public SaaS companies now trade at a median 3.4x revenue, down from 18x at the 2021 peak — an 80% compression confirmed across multiple valuation trackers. One founder captured the shift: “Before AI, the mere fact that you could build the product was a spectacular moat. You needed engineers, months, funding rounds. The difficulty of building software is trending to zero,” leaving distribution as the sole defensible asset.
Elon Musk argued that “source code is on the verge of becoming like assembly,” with AI generating the binary directly. OpenAI’s Thibault Sottiaux predicted that current AI coding tools like Codex “will seem primitive in 2-3 months” as models outgrow laptop-scale workflows.
Note: When software costs approach zero, the moat shifts from engineering capability to distribution and institutional trust. Procurement strategies built on “custom development is expensive, therefore lock in” look different when the cost assumption collapses.
Sources: Aventis Advisors (multiples data), Elon Musk, Thibault Sottiaux (OpenAI)
Institutional Pressure Points
AI Proctoring Loses to AI Cheating at Scale — 58,000 Students Must Retake Mexico’s Largest Entrance Exam
UNAM, Mexico’s largest university, ran its first fully remote entrance exam for 160,000 applicants using AI-powered webcam proctoring and a lockdown browser. The results bore no resemblance to history: 16.3% scored 100 or above on the 120-question test, versus a historical average of 3.5%. Experts estimate roughly half of examinees cheated. The university ordered 58,000 applicants to retake the exam in person and postponed the start of classes.
The governance gap extends beyond education. In the US, tax preparers have been integrating AI tools under a regulatory framework last meaningfully updated in 2013, prompting the IRS to issue emergency guidance in June reminding practitioners that AI “may assist professional judgment but may not replace it.”
Note: The failure mode is instructive: the proctoring AI was outmatched by the test-taking AI. Every institution that runs assessments, certifications, or compliance exams faces the same arms race — and the offence is currently winning.
Sources: Ars Technica, CNN, CNBC (tax preparers)
Autonomous Software Restarts an Idled Copper Mine in Four Months — $310 Million Follows
Mariana Minerals closed a $310 million Series B led by Khosla Ventures and a16z, valuing the company at $1.5 billion. Its flagship achievement: restarting an idled copper mine in southeastern Utah in four months using MarianaOS, an autonomous platform managing mining, refining, and capital project execution under a single AI stack. Traditional restart timelines run 5-10 years. The mine is ramping toward 50,000 metric tonnes of refined copper per year, with a lithium site in East Texas next.
Note: Four months versus five-to-ten years. For EU institutions tracking raw materials sovereignty under the Critical Raw Materials Act, autonomous mining is what AI-compressed timelines look like outside software.
Sources: Fortune, Mariana Minerals (press release)
Stablecoins Reach $300 Billion as IMF Warns Failures Could Outrun Supervisors
Finance’s embrace of blockchain has accelerated: stablecoin market capitalisation sits near $300 billion, tokenised funds have quadrupled, and institutional adoption is widening. The IMF’s financial counsellor Tobias Adrian called this a “structural shift in financial architecture,” warning that automated smart contracts triggering margin calls and liquidations could amplify crises faster than central banks can respond.
Cross-border risks compound the concern: tokenised assets moving instantly across jurisdictions could drive volatile capital flows and rapid currency substitution, eroding monetary sovereignty — particularly in economies with weaker currencies or less-developed financial systems.
Note: The IMF rarely uses language like “faster than supervisors can respond.” For public finance authorities, the message is that the infrastructure needed to manage a crisis in tokenised markets doesn’t exist yet — and the markets aren’t waiting.
Sources: Financial Times, IMF
AI optimising itself, research output decoupling from budgets, software multiples collapsing, exam proctoring outmatched, a copper mine restarted in months, and $150 billion in off-balance-sheet chip financing — the stories span different domains but share one characteristic: the institutional frameworks designed to govern each of them were built for a slower world. The self-improvement loop at the top of today’s digest is the engine behind every other item. It is also the reason the gap between technological pace and institutional pace is not closing.
