Good morning,

We cover OpenAI's remarkable week: killing its biggest consumer product and closing the largest private funding round in history on the same Monday; and the conclusion (for now) of the Anthropic–Pentagon saga. In Competitor Watch: Ping An's claims automation results are jaw-dropping and AIG doubles down in its annual letter. Plus the usual top news roundup.

On March 24, OpenAI shut down Sora, its AI video-generation app, barely six months after launch. The numbers were brutal: roughly $15 million per day in compute costs against $2.1 million in lifetime revenue. Monthly downloads had collapsed from 3.3 million at launch to 1.1 million by February. Disney's planned $1 billion partnership, signed just three months earlier, was effectively dead on arrival (which doesn’t help Disney’s new CEO). (above clip generated by Sora)

One week later, on March 31, OpenAI officially closed the largest private funding round in history: $122 billion at an $852 billion valuation. Amazon anchored at $50 billion, NVIDIA and SoftBank each contributed $30 billion. For the first time, OpenAI also raised ~$3 billion from individual investors via bank channels. To fund its $30 billion share, SoftBank took a $40 billion unsecured bridge loan from JPMorgan, Goldman Sachs, and four Japanese banks. Widely interpreted as a strong signal that an OpenAI IPO is coming within 12 months.

The juxtaposition is instructive. OpenAI is not a company that does everything well. It is a company that raises capital on the promise of the next thing. And the next thing is agents, not spectacle. GPT-5.5, codenamed "Spud," reportedly completed pretraining in late March. Altman told employees it could "really accelerate the economy." Polymarket gives it above 90% odds of shipping by June.

Two other strategic signals. First, the company acquired TBPN, a daily tech talk show, on April 2. Its first media acquisition, with a promise of editorial independence that drew immediate scepticism. Second, its advertising pilot hit $100 million in annualised revenue in under six weeks. OpenAI is becoming a distribution company that happens to build models.

Why it matters for strategy: The Sora failure is a cautionary tale about compute economics; a product can be technologically impressive and economically unviable at the same time. The $122 billion round, meanwhile, shows that capital markets are pricing in a future where a handful of AI companies become infrastructure-layer monopolies, not unlike the cloud hyperscalers a decade ago. If you work in any industry where distribution costs matter (insurance, for instance), the lesson is: AI's value accrues to whoever controls the customer relationship, not whoever has the fanciest model.

The ruling. If you have followed the last 2 edition, you know we have been tracking the Anthropic–Pentagon standoff. On March 26, U.S. District Judge Rita Lin issued a 43-page preliminary injunction blocking the Pentagon's "supply-chain risk" designation. She called it "classic illegal First Amendment retaliation" and wrote that "nothing in the governing statute supports the Orwellian notion that an American company may be branded a potential adversary and saboteur of the U.S. for expressing disagreement with the government." The Trump administration filed an appeal on April 2.

The leaks. In the same week as the ruling, Anthropic had two spectacular data spills. On March 26, a CMS misconfiguration exposed ~3,000 internal files including a polished, launch-ready blog post revealing "Claude Mythos" (codename "Capybara"), described as its most powerful model ever, positioned above Opus. The leaked documents also described a planned invite-only CEO summit in Europe. Days later, 512,000 lines of Claude Code source code were accidentally published to npm - exposing a Tamagotchi-style ‘pet’ and an always-on agent.

Anthropic called both incidents "human error." But the scepticism is widespread and, perhaps, reasonable. The timing, during peak news coverage of the Pentagon victory, just ahead of a European CEO summit, and with a Q4 IPO on the horizon, was remarkably convenient. The exposed Mythos documents were structured as a ready-to-publish web page, complete with headings and a publication date. Several analysts have noted that this looks less like a security failure and more like a controlled detonation of hype, designed to dominate the news cycle and position Anthropic as the company with the most powerful model in the world, at the exact moment when the Pentagon story was making it the industry's most sympathetic protagonist. Whether you call it a leak or a launch, it worked: Mythos is now in the head of most AI enthusiast.

For what it's worth: the model itself sounds genuinely significant. The leaked documents describe Mythos as "far ahead of any other AI model in cyber capabilities," and Axios reported that Anthropic privately warned senior U.S. officials that Mythos makes large-scale cyberattacks materially more likely in 2026. Cybersecurity stocks dropped sharply on the news.

Why it matters for strategy: The court ruling sets a judicial precedent: the government cannot weaponise procurement designations to punish companies for policy positions. That has implications well beyond AI. For insurers specifically, the Mythos cyber capabilities flagged by Anthropic (combined with the broader escalation of AI-powered attacks) should be on the radar of anyone working on cyber risk pricing.

Very little AI news from the last two years is not linked to Nvidia. The giant US chipmaker is driving the industry with monstrous investment and increasingly impressive chips. However, I have found that its business and moat are often misunderstood. For instance, because Nvidia mostly sells chips, it is often thought of as a hardware provider. But is that really true? Or is its moat instead its software, “CUDA,” a programming model that has become essential for many software engineers?

Given Nvidia’s position in the industry, it seems important to understand what Nvidia is, how it came to be, and how it sees itself. This week, I would like to recommend two ways to do so:

For the most passionate, I would recommend starting with the book, which will give you great insights into how Jensen thinks, before giving your time to the podcast.