Tech Digest – April 27, 2026
AI Governance & Risk
Major Insurers Win Approval to Exclude AI-Related Damages from Corporate Policies
Major insurers including Berkshire Hathaway, Chubb, and Travelers have secured regulatory approval to strip AI-related damages from standard corporate liability policies. Since January 2026, ISO exclusion endorsements have allowed carriers to remove generative AI exposure from commercial general liability coverage, and at least six major carriers have filed their own AI exclusions with US state regulators.
Chubb’s approach specifically excludes “widespread” incidents where a single model failure affects many clients simultaneously — the systemic risk scenario insurers fear most. W.R. Berkley has gone further, proposing to bar claims involving any actual or alleged use of AI, including products that merely incorporate the tools.
Note: The risk hasn’t disappeared — it’s been transferred. Every organisation deploying AI tools now carries liability that used to sit with its insurer. Procurement teams reviewing vendor contracts should check whether their coverage still means what they think it means.
Sources: The Information, Tom’s Hardware
Bipartisan House AI Bill Targets Deepfakes and Whistleblower Protections
A bipartisan AI bill backed by Rep. Ted Lieu and Rep. Jay Obernolte would crack down on deepfake distribution and non-consensual AI-generated images while creating federal protections for whistleblowers reporting AI safety risks. The legislation draws on recommendations from the bipartisan House AI Task Force and would require US participation in international organisations developing technical standards for AI. Obernolte is working on a separate Republican AI package expected later this year.
Note: The whistleblower provision is the sleeper clause. As AI systems handle more institutional functions, the people most likely to spot failures are the employees operating them. Whether those reports surface early or late depends on legal protection — and that calculus just changed.
Sources: CNBC
Compute Demand & Hardware
Hyperscale Data Centre in Utah Will Consume More Electricity Than the Entire State
Kevin O’Leary is planning a hyperscale data centre in Utah’s Box Elder County designed to generate its own power, treat its own water for return to the Great Salt Lake, and consume more electricity than the entire state currently uses. The project is self-contained by necessity — no existing grid connection could supply it.
The scale reflects a demand curve that shows no sign of flattening. AWS CEO Matt Garman confirmed that the company has “never retired an A100 server” — six-year-old GPUs remain fully utilised because demand continues to outpace every generation of hardware. As OpenAI’s Noam Brown frames it: model weights now matter relatively less than securing inference compute capacity. The strategic prize is no longer the recipe — it’s the kitchen.
Note: When a single facility needs more power than an entire US state, data centre siting becomes energy policy. European regions competing for digital infrastructure investment should note what’s required: not a business park, but a power plant.
Sources: Salt Lake Tribune, Data Center Dynamics
OpenAI Enters Hardware: AI Smartphone with Custom Processors Targeted for 2028
Supply chain analyst Ming-Chi Kuo reports that OpenAI is developing an AI-native smartphone with Qualcomm and MediaTek as chip partners and Luxshare as manufacturer, targeting mass production in 2028. The device — reportedly codenamed “Gumdrop” — aims to replace the app model entirely with task-driven AI agents. Qualcomm shares surged 13% on the reports, though none of the three companies have officially confirmed the partnership.
Note: If the app-based smartphone gives way to agent-based devices, the enterprise mobility stack — MDM policies, app stores, procurement frameworks — was built for a model that no longer applies.
Energy Beyond the Grid
Nuclear and Geothermal Startups Hit Public Markets — Meta Signs for Space Solar
The AI energy race reached two milestones in one week. Amazon-backed nuclear startup X-energy raised $1.02 billion in its IPO — the first traditional IPO for an advanced reactor company — with shares surging 27% on the first trading day to a $9.1 billion valuation. X-energy already has an order pipeline exceeding 11 GW from Amazon, Dow, and Centrica. Geothermal startup Fervo has filed at roughly a $3 billion valuation.
Meanwhile, Meta signed a first-of-its-kind agreement with Overview Energy for up to 1 GW of space-based solar power, beamed from geosynchronous orbit to ground-based solar installations via near-infrared light. Commercial delivery is expected by 2030. Meta has now committed to 30 GW of renewable capacity across nuclear (Oklo, TerraPower), geothermal, and orbital solar. SpaceX is thinking even bigger — a new share incentive plan ties 60 million additional shares for Elon Musk to delivering “100 terawatts of compute per year” from space-based data centres.
Note: A nuclear IPO oversubscribed at $9.1 billion and a contract for orbital solar in the same week. The energy procurement playbook for AI infrastructure has moved well past “negotiate a grid connection.” The options are multiplying faster than most institutional energy strategies can absorb them.
Sources: TechCrunch, CNBC, Bloomberg, The Information
Deployment Surface, Attack Surface
China Deploys Thousands of Robots Across Its Power Grid as Stolen Drones Trigger FBI Alert
China’s State Grid Corporation is investing approximately ¥6.8 billion ($1 billion) in 2026 on 8,500 robotic systems for grid operations — including 5,000 robot dogs for substation inspection and humanoid robots for high-voltage maintenance on the ultra-high-voltage network. China Southern Power Grid has already deployed humanoid robots in live construction at a 220 kV substation in Yunnan. Industry estimates put total Chinese utility investment in embodied intelligence above ¥10 billion this year.
On the other side of the dual-use ledger, 15 Ceres Air C31 agricultural drones were stolen from a New Jersey logistics facility on March 24 by a thief posing as a delivery driver with forged documents. Each drone is roughly the size of an ATV, carries a 31-gallon tank expandable to 40 gallons, and can lift nearly 180 kg. The FBI is investigating what it calls a potential “nightmare scenario,” citing long-standing US Army assessments that agricultural drones could serve as chemical or biological delivery platforms.
Note: China is replacing high-voltage line workers with robots to eliminate human risk. In New Jersey, drones designed for crop spraying are now a biosecurity concern. The technology is the same; the governance gap is the story. Any institution deploying autonomous systems at scale will need to answer for the attack surface those systems create.
Sources: South China Morning Post, New York Post, DroneXL
Science at Machine Speed
Amateur Cracks 60-Year Erdős Problem with a Single AI Prompt
Liam Price, a 23-year-old with no advanced mathematics training, solved a 60-year-old Erdős problem using a single prompt to GPT-5.4 Pro. The AI found a proof using Markov chains combined with von Mangoldt weights — a technique available for 90 years that no mathematician had thought to apply to this class of problems. Terry Tao suggested that human researchers had hit a “mental block” from making “a slight wrong turn at move one.” Oxford mathematician Jared Lichtman called it the first AI result at the level of Erdős’s Book.
Note: The AI didn’t outthink the mathematicians — it took a path they’d never considered. When the constraint on solving hard problems is cognitive path-dependency rather than raw intelligence, an AI prompt becomes a legitimate research instrument. The raw proof still needed expert interpretation, but the direction came from a machine and a non-specialist.
Sources: Scientific American
First In Vivo CRISPR Gene Editing Succeeds in Phase 3 Trial
Intellia Therapeutics reported positive Phase 3 results for lonvo-z, the first in vivo CRISPR treatment to reach this stage. In the 80-patient HAELO trial for hereditary angioedema, patients receiving the one-time infusion saw attack rates drop 87% versus placebo, with over 60% becoming entirely attack-free. The treatment permanently inactivates the KLKB1 gene that causes the condition — not a recurring therapy, but a single edit to the disease-causing DNA. Intellia has begun a rolling FDA submission and targets a 2027 launch.
Note: This is the dividing line between “gene editing works in theory” and “gene editing is a treatment your health system will need to fund.” One-time edits that replace lifelong medication reshape cost-effectiveness calculations from the ground up.
Sources: CNBC, Intellia Therapeutics, STAT News
Today’s digest traces a single thread through very different domains: the gap between what’s technically possible and what institutions are prepared for. Insurers are excluding AI risk before most organisations have assessed theirs. A single data centre needs more power than an entire state. A 23-year-old with a chatbot solves a problem that stumped mathematicians for six decades. Gene editing just entered the regulatory pipeline as a one-time cure. The technology is arriving across every front — the question is whether the institutional frameworks that need to absorb it can move at anything close to the same speed.
