Good morning,
This week we cover the Anthropic-Pentagon standoff, plus: every major AI lab shipped a new model in the same three-week window, Bank of America put a number on AI's threat to insurance distribution, and more top news.
The Anthropic-Pentagon Standoff: When AI Ethics Meets National Security
Last July, Anthropic signed a $200 million contract with the Pentagon to deploy Claude on classified military systems (the first and only frontier AI model on such networks). The relationship worked well until the Pentagon demanded that Anthropic remove its guardrails so the military could use Claude for "all lawful purposes," without exception. Anthropic drew two red lines: no AI-powered autonomous weapons without human oversight, and no mass surveillance of US citizens. The Pentagon's position was equally clear: the military follows the law, and it's not for a contractor to dictate what the military can or cannot do with a tool it has purchased.
On February 27, after Anthropic refused to budge, Trump ordered all federal agencies to stop using Anthropic. Defense (or “War”) Secretary Hegseth designated Anthropic a "supply chain risk to national security", a label normally reserved for foreign adversaries, never before applied to a US company. The Pentagon's CTO called Amodei a "liar" with a "God complex" on X. Hours later, OpenAI announced a competing Pentagon deal.
Following the events, in a rare movement of unity, the tech industry backed Anthropic. An open letter titled "We Will Not Be Divided" gathered nearly 900 signatures from employees at Google, OpenAI, Salesforce, Databricks, IBM, and Cursor, urging the Pentagon to withdraw the blacklist. Amazon, Google, and Microsoft staff specifically urged their executives to adopt the same guardrails and refuse similar defense contracts. Over 100 Google employees signed a separate internal letter asking leadership to draw the same red lines. The Alphabet Workers Union issued a public statement saying "tech workers are united in our stance that our employers should not be in the business of war."
OpenAI found itself in a particularly uncomfortable position. Sam Altman called the blacklisting "a very bad decision" and admitted his own Pentagon deal was "opportunistic and sloppy," renegotiating it to add surveillance prohibitions.
As I am writing this, Anthropic has reopened negotiations with the Pentagon while simultaneously preparing a legal challenge to the supply-chain designation. At the same time, the story seems to have made more good than bad for Anthropic. Claude surged to #1 on the App Store, overtaking ChatGPT, with free users up 60% and daily signups quadrupling.
This story raise important questions on the role of ethics in the development and deployment of AI, including the moral question of “Whom define what is ethics?” (here Anthropic or the US Government?).
The Great Model Drop: Every Lab Ships at Once
I don't think we've ever seen anything like the last three weeks in terms of model releases. Every major lab shipped something significant.
Google released Gemini 3.1 Pro (Feb 19), which topped 13 of 16 major benchmarks, including a score on ARC-AGI-2 more than double its predecessor's. OpenAI launched GPT-5.4 (Mar 5), its first model with native computer-use capabilities, meaning the model can operate your desktop, browse the web, and execute multi-step tasks autonomously. Anthropic shipped Claude Sonnet 4.6 (Feb 17). And xAI launched Grok 4.10 and started 4.20 beta, the first consumer model with a native multi-agent architecture where four specialized agents debate internally before responding.
But possibly the most interesting release is yet to come: DeepSeek V4 should come in a matter of days. China's trillion-parameter model, would be the first frontier model optimized for Huawei Ascend chips rather than NVIDIA. Combined with Alibaba's open-source Qwen 3.5 (which matches GPT-5.2 performance at 1/18 the cost), Chinese open-source models have grown amazingly over the last couple of month and their market share seems to be highly improving (some suggest a current 15%). This directly challenges whether US chip export controls are actually working.
What does it mean for us? Two things. First, the price-performance curve is collapsing so fast that "AI is too expensive" is no longer a valid reason to delay adoption. Sonnet 4.6 delivers Opus-level output at a fraction of the cost. Second, the China dimension matters: if DeepSeek V4 on Huawei chips performs at frontier level, we're entering a world with two separate AI ecosystems (Nvidia+OpenAI/Anthropic VS Huawei+Alibaba/Deepseek…). Companies will need to balance both.
Competitor Watch
Bank of America: $15B in insurance commissions at risk from AI. In a March 3 report, BofA estimated that over $15 billion in US insurance commissions tied to "low complexity" personal lines and small commercial policies face "not immaterial risk" of AI disintermediation. They specifically noted that Munich Re's Next Insurance already lets customers purchase and bind commercial policies without a human agent. BofA warned that insurance distributor stocks, trading at 22× trailing FCF, do not yet discount these risks. Following the selloff we covered two weeks ago, this is an important Wall Street Report.
Agentic AI enters insurance back offices at scale. Microsoft and Cognizant announced a joint agentic AI platform for insurance, with agents that ingest unstructured emails, scanned PDFs, and intake forms, extract risk information, apply policy rules, and route exceptions to humans.
More Top News
NVIDIA's Q4: $68 billion in revenue. NVIDIA posted $68.1 billion in quarterly revenue (up 73% YoY), with data center revenue alone at $62.3B. Full-year FY2026 revenue hit $215.9 billion. They guided Q1 at $78B. Jensen Huang's thesis remains simple: AI infrastructure demand is insatiable.
OpenAI closes $110B round at $730B valuation. The largest private fundraise in history! Amazon invested $50B, NVIDIA $30B, SoftBank $30B. The investment is still the same, and still worrisome: chipmakers and cloud providers are funding an AI lab that is also their largest customer. It's circular financing at a scale never seen before. Between this and Anthropic's $30B round at $380B, February was the biggest venture month in history at $189 billion.
US drafting sweeping new chip export controls. Bloomberg and TechCrunch reported the Trump administration is drafting rules requiring government approval for all AI chip exports worldwide — not just to China.
China fires back with massive AI subsidies. During the National People's Congress, Beijing announced hundreds of billions in subsidies for domestic AI and semiconductor development, targeting chip self-reliance by 2027. China also injected $44B into state banks to boost tech financing.
"AI Agents of Chaos" study rings alarm bells. Thirty-eight researchers documented 11 catastrophic failure modes when AI agents were given real system access, including leaking secrets, running destructive commands, and falsely reporting tasks as "completed" when they had failed. Those findings should raise concern when taking decision to deploy AI for claims processing or underwriting workflows. The paper.
Anthropic enables "Memory transfer". In a swift move, as they gained in popularity after the Pentagon news, Anthropic released a feature to "Bring your preferences and context from other AI providers," making it easier for people used to ChatGPT or Gemini to jump to Claude. Overall, it seems an important feature, breaking the silos between AI and putting the power in the customer's hands again.
Tool to Try: Apps
Most of you are probably used to using ChatGPT or Claude with all their out-of-the-box capabilities by now ("deep search" or extended thinking…). A tool that is often overlooked, however, is the "Apps" (for ChatGPT) or "Connectors" (for Claude).
Apps are additional tools for your chat, built and provided in collaboration with external partners. For example, OpenAI provides a "Booking.com" app that allows users to search and book trips using Booking's database directly in ChatGPT. Apps are the source of last month's broker crash, as Tuio released an insurance app.
Swiftly, OpenAI and Anthropic have been expanding their app directories and now offer bridges to Adobe Photoshop or Figma (to create), Canvas (to build presentations), Booking or Expedia (to book trips), Slack or Monday.com (to manage your projects), and much more…
As a user, it is a great tool to master as it can make life easier. As an insurance professional, I would argue that it is important we get the concept as soon as possible to build our own and bridge the gap between Zurich and a ChatGPT user.