This week, the AI industry confronted two significant forces: one political, one physical. A court prevented the Pentagon from using regulatory measures against Anthropic, while simultaneously, $630 billion in capital expenditure is encountering a severe limitation: power grids are booked for a decade, and gas turbines are sold out until 2029. Infrastructure, not ideology, now shapes the game.

Watch the 2-min video overview:

🌍 Landscape

Pentagon's Anthropic blacklist blocked by federal court: A judge halted the supply chain risk designation, citing a First Amendment violation. The government's "ideology punishment" failed legal scrutiny, though an appeal is pending [The Information].

David Sacks exits AI czar role & joins PCAST: Having exhausted his 130-day limit as a special government employee, Sacks will continue advising Trump via the President's Council on Science & Technology, alongside Jensen Huang, Mark Zuckerberg, and Larry Ellison [Bloomberg].

Stanford study: AI chatbots consistently side with users in conflicts: Research shows 11 LLMs backed users over 50% of the time, even when the users were wrong. Users preferred these agreeable models, which in turn made them more self-righteous. Sycophancy in AI responses is pervasive and alters people’s behavioral inclinations. [Science]

Sycophancy in AI responses is pervasive and alters people’s behavioral inclinations. Source: Science

📈 Tailwinds

AI value chain inverts toward infrastructure, not models: This week saw $25B in deals, including IBM buying Confluent for $11B (real-time data) [IBM], Eli Lilly acquiring Insilico's drug pipelines for $2.75B [Bloomberg], and Physical Intelligence raising $1B for robot control [TechCrunch].

OpenAI closes $122B funding round at $852B valuation: In the largest private raise in history, OpenAI now generates $2B/month in revenue — growing 4x faster than Google or Meta at the same stage. Enterprise clients now account for over 40% of this revenue [CNBC].

Chinese robotics scaled production: Agibot hit 10,000 units in 9 months, accelerating from 5,000 to 10,000 in just 3 months. These units are deployed across logistics, retail, and hospitality, with global expansion underway [Int Eng].

📉 Headwinds

OpenAI projected to burn $14B in 2026: Despite earning $2B/month, infrastructure costs are outpacing revenue growth. These negative margins signal an unsustainable capital expenditure trajectory [Neuron].

AI data centers heating surrounding areas by 9.1°C: This thermal impact affects 340 million people. Power and cooling constraints are emerging as physical bottlenecks for AI expansion [New Scientist].

Anthropic's Claude Mythos carries “unprecedented cybersecurity risks": The unreleased flagship model is expensive to run. A "Capybara" tier is also in testing, with security evaluations ongoing [Fortune].

Deals & Partnerships

Google finances multibillion-dollar Texas data center for Anthropic: Google provided construction loans for Nexus Data Centers (2,800 acres, 500 MW). Anthropic plans to use 1 million Google TPU chips. The Texas project is expected to deliver about 500 megawatts of capacity as soon as late 2026 [FT].

Servers inside a data centre. The Texas project is expected to deliver about 500 megawatts of capacity as soon as late 2026 © AFP or licensors. FT

Microsoft leases 900 MW Texas site: Microsoft took over a Crusoe Energy facility previously abandoned by Oracle and OpenAI. This move addresses a server capacity shortage for Copilot and internal AI training [DCD].

Mistral raises $830M debt for 13,800-GPU French data center: Using Nvidia infrastructure, this move aims to reduce reliance on U.S. cloud providers, acting as a European foundational AI independence play [Le Monde].

+ News on AI

Infrastructure physics now constrains AI capex: Four tech giants (Amazon, Microsoft, Alphabet, Meta) are projected to spend $630B in 2026 on AI infrastructure (2.2% of U.S. GDP). The real bottleneck is power: grid connections take a decade, and gas turbines are sold out until 2029. Land, labor, and cooling are reshaping the capital expenditure calculus [Reuters].

Public launches first agentic brokerage: Users can now define tasks while AI agents monitor markets and execute trades independently on financial-grade infrastructure. The launch video garnered 2.3M views, signaling strong consumer demand for autonomous AI agents managing capital [X].

Hospital CEO plans to replace radiologists with AI: NYC Health + Hospitals and Westchester Medical Center Health Network are ready to deploy AI for primary image reads once regulatory barriers clear. Radiologists warn that AI cannot yet operate independently without risking patient safety. Cost pressure is driving adoption, but governance lags behind. [Radiology Business]

What to Watch

  • Anthropic appeal and government AI leverage: The Pentagon has 7 days to appeal the court's First Amendment ruling. The outcome will reshape what pressure the government can apply to AI companies without breaking the law. Watch whether the administration finds other mechanisms to isolate Anthropic.

  • AI sycophancy reshaping user behavior at scale: Stanford research shows 11 LLMs consistently side with users, making them more self-righteous. As agents scale, this embedded bias in autonomous systems could amplify polarization and poor decision-making across finance, healthcare, and policy.

  • Power constraints become the real AI bottleneck: Grid connections in London take a decade, and gas turbines are sold out until 2029. As $630B in capital expenditure hits physics limits, watch whether distributed compute (orbital, remote, or on-site generation) reshapes the infrastructure moat.

🎓 Webinars & Events

📅 April 14, 2026, 17:00-18:00 CEST | 🌐 Online (Free)

AI Beyond LLMs: From Generative Models to Economic Decisions - William Cong (NTU Singapore) and Vahe Andonians (Cognaize) explore AI applications in social science research beyond text: goal-oriented optimization, generative modeling, and agent-based simulations for portfolio, risk, and policy decisions. [Register]

Valence Thoughts

"We shape our buildings, and afterwards our buildings shape us." — Winston Churchill

That’s it for this week.

Keep reading