AI News Roundup — Week of August 16, 2026: The AI Boom Runs Into a Permitting Wall

More than 530 US counties and municipalities now actively restrict or block new data centres, according to a review by Heatmap. The tally counts only the severe stuff — outright permit bans, impossible noise limits, setback requirements no site can meet. The overwhelming majority were enacted since January this year, and nearly 190 since June 1 alone. More than 50 data centre projects have been cancelled so far this year after local pushback.

Then the biggest domino: Texas Governor Greg Abbott announced a moratorium on data centre approvals until regulators can audit projects queuing for grid connection. ERCOT is tracking more than 1,800 projects representing over 474 gigawatts of requested interconnection — roughly 90% of it data centres. No stated end date for the audit. Texas is the country's second-largest data centre market.

The money did not slow down to match. Amazon, Microsoft, Meta and Google are now planning up to $725 billion of capital expenditure in 2026; Microsoft alone added 31 data centres last quarter. The European Commission opened tenders meant to mobilise over €30 billion for seven AI "gigafactories."

Eight contech startups raised in the week to August 10, and the pattern is different from a fortnight ago. FormX (San Francisco) took $12.5M seed for a platform that runs design, engineering, planning and permitting for custom homes on top of a modular system. Conmeet (Germany) raised €6M for a single platform covering project management, procurement, scheduling and finance aimed at 10–500 employee contractors. SnapScale (South Korea) raised an undisclosed seed to put a prompt interface on top of CAD for plant and EPC design documentation. Buzz Solutions (California) raised a $20M Series A turning utility inspection imagery into asset intelligence across transmission, distribution and utility-scale solar. Exclaim Robotics (Zurich) took $4.95M pre-seed for robots servicing hazardous infrastructure, starting with the 800V DC racks inside AI data centres.

And Caliperd (Australia) came out of stealth with AI concrete estimating that reads pad footings, slabs, columns and reinforcement schedules and outputs volumes, rebar tonnage, formwork areas and rate-priced quantities straight into a BoQ template.

Sundar Pichai confirmed the Gemini app passed 1 billion monthly active users — Google's fastest product ever to the milestone, and its fourteenth to reach it. The usage detail is more interesting than the headline: 63% of users interact by voice, one in five Gemini Live sessions involves live camera or screen sharing, and the system generates over 150 million images a day.

Two days later Google shipped Gemini 3.7 Flash, three weeks after 3.6 Flash, at an introductory $0.75 per million input tokens — half the launch price of the model it replaces — with a 1M-token context window and sizeable coding gains (DeepSWE v1.1 from 49.0% to 65.3%).

Meanwhile Anthropic has been meeting investors ahead of a possible public listing in the autumn, per the Wall Street Journal, while fielding questions about Chinese competition and infrastructure costs.

More than 120 organisations, including Nvidia, Cisco and CrowdStrike, are backing a proposal from the Open Secure AI Alliance called SAFE — the Shared AI Findings Exchange — a common standard for reporting security incidents involving autonomous agents. It covers events like an agent reaching into third-party systems without authorisation, exposing confidential data, or continuing to probe production infrastructure after operators realised something had gone wrong. Participants would preserve prompts, agent traces, tool calls, identities, credentials and permissions, notify affected parties quickly, and file a confidential initial report within four business days. The draft's sharpest clause: an operator's intent doesn't cancel the reporting obligation — an agent that attacks a real target because its developers thought it was still in a simulation still counts as an incident.

The timing wasn't accidental. The Financial Times reported that Taiwan's nuclear regulator was targeted by what it described as an autonomous AI-enabled cyberattack in which agents ran reconnaissance and intrusion attempts in parallel rather than assisting human operators. Separately, water and wastewater utilities in at least seven US states reported cyberattacks in recent weeks; the FBI said some intrusions caused pressure loss and flooding.

We covered the EU AI Act's application date two weeks ago as a calendar event. This week it turned into an instruction. Reporting from the French property trade press sets out three things landing together: the AI Act's obligations now apply concretely to estate agents using AI — transparency, governance for high-risk systems, documentation scaled to how much is automated; France's data protection authority, the CNIL, restated that agencies may only request the documents legally required of a rental applicant, with penalties for over-collection; and, most specifically, using ChatGPT to automate the screening of tenant applicants is now explicitly off the table.

The reasoning is plainly stated: letting a general-purpose model decide a rental application alone produces bias you can't audit and decisions you can't defend in a dispute.