Tech Digest – May 11, 2026
Infrastructure Meets Reality
AI Data Centers Hit Three Walls at Once — Ratepayer Revolt, Sovereign Limits, and Warzone Routing
Maryland’s Office of People’s Counsel filed a complaint with the Federal Energy Regulatory Commission over $2 billion in grid upgrades allocated to the state for data centers operating primarily in Virginia. The complaint argues Maryland ratepayers would shoulder an extra $1.6 billion over the next decade — roughly $345 per residential customer — to subsidise infrastructure they never asked for and won’t directly benefit from. The state is invoking the “ratepayer protection pledge” that tech companies made to the White House, arguing the companies themselves should be billed for the upgrades they require.
In Kenya, Microsoft and G42’s $1 billion geothermal-powered data centre has stalled after the government couldn’t meet payment guarantee demands. At full build-out, the facility would have consumed more electricity than half the country uses at peak demand — President Ruto said openly that powering it would mean switching off half the nation. Meanwhile, U.S. hyperscalers are routing Gulf data centre traffic through fibre-optic cables that an Iraqi telecom has strung alongside crude-oil pipelines, running AI infrastructure literally through active conflict zones.
Note: Three stories, one pattern: the physical world is pushing back. Ratepayers don’t want to subsidise someone else’s compute. Sovereign nations can’t guarantee what hyperscalers demand. And the shortest path between a data centre and its users now runs through a warzone. Infrastructure planning that assumes smooth expansion is planning for a world that doesn’t exist.
Sources: Tom’s Hardware, Bloomberg, Rest of World
Model Governance at a Turning Point
OpenAI Winds Down Fine-Tuning as Foundation Models Go General-Purpose — Cisco Ships a Model Provenance Kit
OpenAI has begun winding down its self-serve fine-tuning API, blocking new users and giving existing customers until January 2027 to create new training jobs. Inference on already fine-tuned models will continue until those base models are deprecated, but the customisation pathway is closing. The move reflects a strategic bet that general-purpose models are now capable enough that weight-level adjustment is unnecessary for most use cases — with OpenAI pushing prompt engineering, retrieval-augmented generation, and custom GPTs as alternatives.
Separately, Cisco released its open-source Model Provenance Kit, which examines model metadata and weights to trace shared origins and detect tampering — a supply-chain transparency tool for the AI models organisations are deploying. The timing is instructive: as customisation options narrow, knowing exactly what’s inside the model you’re now committed to matters more.
Note: The fine-tuning shutdown isn’t just a product decision — it’s an architectural signal. Organisations that built differentiation through custom-tuned models are being herded back onto the general-purpose layer, where switching costs drop but vendor leverage increases. Anyone whose AI procurement assumed customisation as a differentiator should revisit that assumption.
Sources: OpenAI Developer Community, Cisco Blog
The Stack Gets Competitive
AMD’s ROCm Stack Jumps 75x in Two Weeks — Nvidia’s Software Moat Looks Shallower Than Assumed
SemiAnalysis reports that AMD’s ROCm software stack has improved inference performance by over 75x in the 14 days since DeepSeek V4’s launch, through fused operations and reduced CPU overhead that improve HBM memory utilisation. The remaining gap to Nvidia’s B200 has narrowed to roughly 7.5x. AMD’s Instinct MI350 and MI400 accelerators, paired with the vLLM-ATOM plugin, are now competitive for several major open-source models.
Note: For years, Nvidia’s software ecosystem (CUDA) mattered more than its hardware lead. A 75x improvement in two weeks suggests that moat is shallower than procurement teams assumed — and that single-vendor GPU commitments may be locking in cost that competition is about to erode.
Sources: SemiAnalysis, AMD
Alphabet Briefly Overtakes Nvidia in Market Cap — The Market Reprices Who Owns the AI Stack
Alphabet briefly surpassed Nvidia in market capitalisation in after-hours trading on May 10, capping a 160% rally over the past year. The catalyst: Q1 2026 revenue of $109.9 billion (up 21.8% year-over-year), with Google Cloud growing 63% to $20 billion and carrying a $460 billion backlog. The company once dismissed as an AI laggard now owns search, cloud infrastructure, custom training silicon (TPUs), the world’s dominant mobile OS, and the browser most people use.
Note: When the company that owns most of the AI stack — from training chips to the distribution layer — briefly becomes the world’s most valuable, it tells you where the market thinks value settles. Not in any single component, but in the full vertical.
Workforce Under Pressure
Women Hold 83% of the Most AI-Vulnerable Jobs — and the Least Capacity to Adapt
A new analysis finds that women hold 83% of positions in the 15 occupations most vulnerable to AI automation, despite making up 47% of the total workforce. The most exposed roles — office clerks (2.5 million workers), secretaries and administrative assistants (1.7 million), and receptionists (965,000) — are overwhelmingly female. Of the roughly 6 million workers with the least capacity to adapt to AI-related displacement and find new positions, 86% are women. Women of colour make up over 30% of the most AI-vulnerable workforce.
Note: Workforce transition programmes that treat AI displacement as gender-neutral are designing for a population that doesn’t match the one actually being displaced.
Sources: Inc., National Partnership for Women & Families
McClatchy Journalists Withhold Bylines from AI-Generated Articles
Journalists at McClatchy, one of the largest newspaper chains in the United States, are withholding their bylines from articles that have been rewritten or generated by AI tools. The action is part of a broader dispute over how newsrooms deploy AI in editorial workflows — specifically, whether human journalists’ work can be fed into systems that produce output under the publication’s brand without the original author’s consent.
Note: Bylines are the last unit of individual accountability in institutional publishing. When journalists refuse to attach their names to AI output, they’re drawing a line that other knowledge workers — analysts, auditors, consultants — will eventually face too.
Sources: The Spokesman-Review
The AI Attack Surface Widens
40% of Breaches Experian Handled Last Year Were AI-Powered
Experian reports that 40% of the approximately 5,000 data breaches it serviced in 2025 were AI-powered, a sharp increase from prior years. The company’s 2026 forecast predicts agentic AI will overtake human error as the leading cause of data breaches. Global fraud costs have surpassed $534 billion.
Note: Every AI procurement business case should now include a line item for defending against AI. The threat isn’t theoretical — it’s already responsible for two in five major breaches.
Sources: Bloomberg via Techmeme, Experian 2026 Forecast
Razorback: An Autonomous Combat Vehicle That Powers Drones, Fires Lasers, and Needs No Crew
Utah-based Hypercraft unveiled Razorback, a software-defined autonomous ground vehicle with a 300 hp hybrid-electric drivetrain, 1,090 kg payload capacity, 450 km range, and 38 kW of exportable power — enough to charge drone swarms, power directed-energy weapons, and sustain forward command posts. The platform’s open architecture supports over-the-air mission reconfiguration: a logistics vehicle one day, a counter-drone platform the next. No human onboard.
Note: The 38 kW of exportable power is the specification that matters most. This isn’t a transport robot — it’s a mobile power grid for autonomous warfare. The shift from crewed to uncrewed ground systems now has a production-ready reference architecture.
Sources: Defence Blog, Defense Advancement
Research Frontiers
MIT Ships FINGERS-7B — The First Foundation Model Built to Prevent Alzheimer’s
An MIT-led team has released FINGERS-7B, the first AI foundation model designed specifically for Alzheimer’s prevention. Unlike traditional approaches that analyse one data source at a time, FINGERS-7B integrates lifestyle, clinical, genomic, and proteomic signals simultaneously to identify preclinical biomarkers — disease signals that can precede memory symptoms by a decade or more. The model achieves 4x more accurate preclinical diagnosis than existing tools. It is open-source, deployed in MIT’s AD Workbench, and was presented at ICLR in April.
Note: Open-source, multi-omic, and already deployed. Health institutions evaluating AI for preventive care now have a benchmark — and under the AI Act’s high-risk classification for medical AI, procurement decisions here will require documented evidence of accuracy and bias testing. FINGERS-7B ships with both.
Sources: MIT Picower Institute, Neuroscience News
Today’s threads converge on a single tension: AI’s ambitions keep scaling — into healthcare, into combat, into the full market-cap leaderboard — but the physical, financial, and social infrastructure underneath is straining visibly. Maryland ratepayers don’t want to subsidise someone else’s compute. Kenya can’t power it. Journalists won’t put their names on it. And 83% of the workers most exposed to it are women who were never part of the planning conversation. The institutions that act on these friction signals now — in procurement, workforce transition, and security posture — will be better positioned than those waiting for the friction to resolve itself.
