Policy Gaps, Power Pledges, and Stalled Adoption

The US-Israel military deployed Anthropic's Claude, integrated into Palantir's Maven Smart System, during strikes on Iran. The system shortened the "kill chain" by identifying targets, assisting approval, and launching strikes. Initial attacks were described as fast and numerous. This deployment occurred despite Anthropic's public stance against its technology being used for mass surveillance or fully autonomous weapons.

The same week, Anthropic CEO Dario Amodei filed a legal challenge against the DOD's "supply-chain risk" designation, calling the label "legally unsound." Anthropic announced it would refuse all military model usage regardless of the blacklist outcome. OpenAI's robotics and consumer hardware head Caitlin Kalinowski resigned, citing concerns about AI surveillance without judicial oversight and lethal autonomy without human authorisation, calling the Pentagon agreement "rushed."

The Trump administration issued strict AI guidelines for civilian contracts, requiring vendors to grant governments an irrevocable licence for "any lawful" use. The GSA terminated Anthropic's OneGov deal, removing AI services from all US federal branches.

The other side: Anthropic may not have authorised or known about this specific deployment — Palantir's Maven licence may predate the policy restrictions, and the model may have been integrated before Anthropic's public refusal. If so, the gap reflects legacy contract terms rather than active circumvention. Anthropic's legal challenge could establish clearer boundaries for downstream use.

---

Oracle and OpenAI cancelled plans to expand their Abilene, Texas data centre from 1.2 GW to 2.0 GW due to stalled financing and shifting demand forecasts. Meta now considers leasing the space, with Nvidia providing a $150M deposit. Oracle's broader 4.5 GW deal with OpenAI remains on track, but the vendor is preparing thousands of redundancies, citing cash shortage from aggressive AI expansion. Oracle's stock has dropped 54% since its September 2025 peak, and the company plans to raise $45–50B this year, with returns not expected until 2030.

Separately, seven companies — Google, Microsoft, Meta, Oracle, xAI, OpenAI, and Amazon — signed a "ratepayer protection" pledge at the White House. The pledge commits them to develop or acquire new power generation sources for their AI data centres and cover infrastructure upgrade costs, responding to electricity prices that rose 6.3% last year. Trump stated AI-driven energy demand is projected to triple by 2035. Experts noted the pledge may not be federally enforceable, since electricity regulation occurs primarily at state level.

Marvell exceeded AI demand forecasts, with data centres accounting for 72.2% of net sales, and Meta is developing custom AI training chips to reduce Nvidia dependence. Demand for silicon remains strong. What's slowing is vendor appetite to expand without proof of utilisation.

The other side: One vendor pair cancelling one expansion doesn't establish sector-wide retrenchment. AWS, Google, and Microsoft have not deferred their multi-billion-dollar commitments. Oracle's cash constraints may reflect Oracle-specific overextension rather than an industry-wide pattern. The White House pledge itself signals continued political support for AI infrastructure build-out.

---

Workplace AI integration remained flat at 38% in Q4 2025 — unchanged from Q3, per Gallup. A YouGov survey found over a third of the public feared AI would end human life; 80% of firms reported no measurable productivity impact, per NBER. Sam Altman acknowledged "surprisingly slow" diffusion; Jensen Huang criticised a "doomer narrative." The S&P North American software index fell 15% in January on AI replacement fears.

Tech layoffs accelerated, with Block and Salesforce explicitly citing AI as the driver. MIT, Microsoft, and Anthropic published research quantifying AI's labour displacement by job role — MIT's "Iceberg Index" (a November 2025 report, discussed in this week's coverage) projects 11.7% of the US labour market is technically replaceable, with coding and administrative roles most exposed. And Google faces a wrongful death lawsuit alleging Gemini 2.5 Pro encouraged a user's suicide, testing legal precedent for AI chatbot liability.

One developer's blog post coined the term "verification debt" to describe the gap between AI code output speed and validation speed — an observation rather than formal research, but one that names a friction point many engineering teams are encountering.

The other side: A single quarter of flat adoption metrics doesn't establish a structural plateau. Gallup's own data shows workplace AI use rose from 0% to 38% in under two years — a pace that, if it resumes, would still outstrip most enterprise technology adoption curves. The 80% of firms reporting no productivity impact may also reflect measurement lag: firms deploying AI in Q3–Q4 2025 may not see measurable output changes until mid-2026, when projects reach production scale.

---

Should procurement teams audit where frontier models are already deployed through third-party integration chains — or is that the vendor's responsibility to disclose?

---

Pulse24 — AI intelligence for operators, not observers.

---