Tech Digest – August 25, 2026
Infrastructure Meets Democracy
Three-Quarters of Americans Now Oppose Data Centres — Texas Governor Calls the Backlash Self-Inflicted
A Fox News poll of 1,003 registered voters finds Americans reject local data centre construction 70 to 30, with roughly eight in ten preferring slower buildout over competing with China on AI. Senator Bernie Sanders, who proposed a construction moratorium five months ago with Representative Alexandria Ocasio-Cortez, says 75% now back the policy he was called “extremist” for floating. Governor Greg Abbott, who once marketed Texas as the “epicentre of AI development,” told ABC’s This Week that the industry “basically dug their own grave.” He has ordered regulators to audit roughly 250 to 300 projects seeking grid connections.
The federal response is split. President Trump insists that refusing data centres is “making a mistake” because the industry “could be bigger than oil.” Meanwhile, the grid itself is backlogged: ConEd’s interconnection queue for battery storage — the projects meant to stabilise the grid — has grown 300% in two years for want of transformers. One exit route is literal altitude. SpaceXAI is adopting NVIDIA’s Vera CPU for its gigawatt-scale Grok infrastructure and planning a Starmind satellite based on a Vera Rubin NVL72 rack — compute that orbits past any zoning board.
Note: Dr. Alex Wissner-Gross coined “thread lines” for the queues this creates — where the 20th century rationed calories, the 21st rations intelligence. If the world’s largest AI infrastructure pipeline stalls on voter opposition, compute scarcity ripples downstream to every institution buying cloud services. The bottleneck isn’t silicon. It’s zoning boards.
Sources: Fox News, Axios, Axios, NVIDIA, LA Times
Autonomous Systems Cross New Lines
AI-Piloted Drone Kills Three Civilians in Ukraine — Uncrewed Warship Fires Live Missiles at Sea
Ukrainian investigators report that a Russian Molniya drone carrying an NVIDIA Jetson Orin module struck a Zaporizhzhia petrol station on 6 July, killing 19-year-old Tetiana Bubynets, 41-year-old Oleksiy Svirin, and 48-year-old Roman Karpiy. The wreckage carried no radio antennas; onboard software selected the target autonomously in its terminal approach. Forensic analysis of the unencrypted code confirmed the targeting logic. Kateryna Bondar of the Center for Strategic and International Studies calls it the first documented case of civilian deaths from a Russian drone using fully autonomous targeting. The Jetson Orin board — a $199 consumer module — was stamped “Made in China.”
Separately, Saildrone fired two Lockheed Martin JAGM missiles from a 20-metre uncrewed vessel during the US Navy’s RIMPAC exercise off Hawaii — the first weapons launch from one of its wind-powered ships. A human made the shoot decision via the Navy’s ARES weapons control system, with targeting data from the USS Theodore Roosevelt carrier strike group. In Hangzhou, a pilotless lifebuoy flew to a drowning swimmer, deactivated its rotors so the person could grip the frame, and returned to shore autonomously. Same class of silicon. Different objective function.
Note: The Jetson Orin is designed for hobbyist robots and smart cameras. The gap between autonomous rescue and autonomous killing is a software configuration, not a hardware threshold — and the hardware costs €200.
Sources: New York Times, Saildrone, CGTN
Adoption Accelerates, Unevenly
AI Power Users Aren’t Coders — Legal Adoption Up 108x Since February
Data from a16z shows the fastest-growing Codex adopters since February are in legal (108x), sales and recruiting (41x each), marketing (26x), and healthcare (24x). Gavin Baker of Atreides Management reports his fund’s internal AI spend will be roughly 100x higher in August than in March, still doubling monthly — what he calls “compute inequality” in action. Sam Altman, meanwhile, admitted on the Founders podcast that he misjudged timelines: “The economy just has so much inertia,” he said, noting he expected far more disruption after GPT-4 in 2023. He now frames the slow pace as “a positive.”
Note: Altman sees inertia; a16z measures a 108x surge. Both are true — economy-wide, transformation is slow; at the desk level, it’s already happened. The question for any institution isn’t whether AI adoption is coming, but whether your teams are in the 108x cohort or the inertia cohort.
Sources: a16z, Gavin Baker, Yahoo Finance
China Runs 70% of Its AI Tokens on Video While America Ships Text Models
Alibaba launched Wan3.0, a video generation model that turns spreadsheets, slides, and documents into 30-second videos at up to 1080p resolution. The release came one day after a $10.2 billion share sale and a week after AI capital expenditure cut Alibaba’s earnings by 75%. Cristóbal Valenzuela, CEO of Runway, reports that video generation now accounts for roughly 70% of AI token consumption in China — driven by short-form content production and robotics training data. “America is LLM-pilled,” he wrote. “China is world-model-pilled.”
Note: This is a procurement map, not a geopolitical talking point. EU institutions buying video and simulation capabilities will increasingly draw from Chinese-origin models. Text-based reasoning tools will come from American labs. The AI supply chain now splits by modality.
Sources: Reuters, Cristóbal Valenzuela
Anthropic Hits $65 Billion Revenue — Model Pricing Enters iPhone Phase
Ramp spending data published by the Financial Times shows Anthropic’s annualised revenue has reached $65 billion, up from $9 billion at the end of 2025, ahead of a potential $2 trillion IPO. Within that growth, model economics are splitting: Fable 5 accounts for just 11% of Anthropic’s enterprise spend, overtaken by Opus 5. The drivers aren’t capability — they’re price and a 30-day data retention mandate that enterprise buyers reject. OpenAI has separately passed $40 billion in annualised revenue.
Gavin Baker projects that closed frontier models will follow iPhone economics: 15 to 25% of total AI token volume, but 60 to 90% of economic value. If the pattern holds, institutions will pay premium rates for a small share of their AI usage — and the compliance terms attached to that premium tier will matter more than the benchmarks.
Note: The most revealing signal isn’t the revenue — it’s what held Fable 5 back. A data retention mandate and higher pricing outweighed superior capability. For any institution evaluating AI vendors, this confirms what procurement officers have long suspected: compliance terms drive adoption more than benchmark scores.
Sources: Financial Times, TechCrunch, Steve Hou / Gavin Baker
The Human Baseline
Goldman Partner Warns of “Cognitive Atrophy” as One in Ten Web Pages Turns AI-Written
Chris Churchman, a Goldman Sachs partner who leads the bank’s institutional digital platform Marquee and co-chairs its AI working group, warned that outsourcing reasoning to AI models risks “cognitive atrophy.” Speaking on the firm’s Exchanges podcast, he compared it to how GPS eroded navigational skills and search engines hollowed out memorisation. “You still need to reason about problems and structure it into an argument,” he said. “Now we’re delegating reasoning.”
Meanwhile, a Pew Research Center study of 490,000 English-language web pages finds that 10% now show significant signs of AI authorship — rising to over a third among pages published since ChatGPT’s launch. Usage of em dashes, a hallmark of AI-generated prose, has doubled. Government and education domains show roughly 1% AI authorship; commercial domains are at 10% and climbing.
Note: The feedback loop is worth tracing: AI writes the web that trains the next AI, and the humans in the middle gradually lose the reasoning skills to tell the difference.
Sources: CNBC, Pew Research Center
Fertility Declining in 219 of 236 Countries — Population Peaks Near Nine Billion by 2056
Economists Jesús Fernández-Villaverde (University of Pennsylvania) and Patrick Norrick (Northwestern University) model fertility trends across 236 countries and territories from 1950 through 2023. They find declining fertility in 219 of them, with no sign the trend is levelling off — and project global population peaking near nine billion around 2056, roughly 30 years earlier than conventional UN estimates. Some of the fastest fertility declines are in middle-income countries, not the wealthiest, challenging the standard development narrative.
At the micro scale, the trend is already visible. A 350-person village in Andalusia is refilling — not with births, but with remote workers drawn by affordable housing and fibre broadband. Synthetic minds scale just as biological ones plateau.
Note: Every pension model, municipal service plan, and workforce projection in the EU is built on a population trajectory. If the peak arrives 30 years ahead of schedule, those models need revising — and the automation reshaping the workforce is partly a response to the demographic ceiling arriving faster than expected.
Sources: Fernández-Villaverde & Norrick (UPenn), Condé Nast Traveler
Alzheimer’s Blood Test Cleared at Over 90% Accuracy — Screening Without Spinal Taps
The US Food and Drug Administration cleared PrecivityAD2, a blood test by C2N Diagnostics for Alzheimer’s disease evaluation in adults aged 40 and older with cognitive symptoms. The test measures plasma amyloid and tau peptide ratios using mass spectrometry, detecting amyloid pathology at over 90% accuracy — with a 97.6% positive predictive value validated against amyloid PET scans. It is the third FDA-cleared blood biomarker test for Alzheimer’s, and the first available for patients as young as 40.
Note: The shift isn’t just diagnostic — it’s structural. When screening requires a PET scan, only symptomatic patients in specialist settings get tested. When it requires a blood draw, primary care can screen at population scale. That moves Alzheimer’s care from late-stage management to early-stage intervention — a different capacity plan entirely.
Sources: Washington University Medicine
Today’s items converge on one signal: the constraints on AI are no longer technical. Voters block data centres, grids can’t absorb the batteries meant to power them, and a drone on a €200 consumer chip crosses a line no treaty covers. Adoption explodes 108x in legal departments while the CEO of the leading AI company admits the broader economy barely noticed. The institutions that navigate the next decade won’t be those with the best technology strategy — they’ll be those that planned for the world the technology has to run in.
