NVIDIA GTC 2026, AMI Labs, the Anthropic-Pentagon Update, and Why Only 19% of People Want AI for Work

Good morning,

This week we cover NVIDIA's GTC 2026 and the dawn of the "agent economy," plus: Yann LeCun bets $1 billion (in France) that the entire AI industry has it wrong, the Anthropic-Pentagon standoff reaches the courts, Supermicro's co-founder is arrested for smuggling chips to China, and more top news. But my biggest read recommendation is the end section on Anthropic latest report: What 81,000 people want from AI.

NVIDIA's annual conference was the industry's most important event of the year so far. Jensen Huang used a nearly three-hour keynote to make one thing clear: the age of "agentic AI" is here, and it will need a lot of new hardware.

The headline numbers are staggering. Huang disclosed $1 trillion in cumulative orders for Blackwell and Vera Rubin chip systems through 2027. He unveiled the Vera Rubin NVLink 72 system (10× performance-per-watt gains, shipping H2 2026), a concept for orbital AI data centers (matching Elon Musk bet), and the DGX Station GB300, a desktop-sized AI supercomputer with enough compute to run a frontier model from your office.

But the real story wasn't hardware. It was Huang's framing of an "inflection point of inference", a shift from training models to deploying them as autonomous agents at massive scale. Every major partnership announced at GTC (Meta's $27B deal with Nebius, AMD's $100B agreement with Meta, NVIDIA's $2B investment in Synopsys) is about building infrastructure for a world where AI agents do things, not just answer questions.

NVIDIA also released Nemotron 3 Super (a 120B-parameter open model leading open-weight benchmarks), launched the Nemotron Coalition with Mistral, Perplexity, Cursor, and others, and debuted NemoClaw for their OpenClaw agent platform. Huang's pitch: "OpenClaw is the operating system for personal AI." Think Android, but for AI agents instead of apps.

Mistral chose GTC to launch Mistral Forge, an enterprise platform letting companies build frontier-grade AI models trained on proprietary data. CEO Arthur Mensch said Mistral is on track for $1B in annual recurring revenue.

Want to learn more about Nvidia and Jensen Huang? I highly recommend: The Thinking Machine by Stephen Witt.

In a period dominated by hardware, the most intellectually interesting story may be the quietest: Yann LeCun raised $1.03 billion in seed funding — Europe's largest seed round ever — to build AI that works nothing like ChatGPT.

LeCun is one of three researchers who won the 2018 Turing Award for inventing convolutional neural networks. He spent twelve years as Meta's Chief AI Scientist. In November 2025, he told Zuckerberg he was leaving. His reason: he believes the entire large language model paradigm is a statistical dead end for truly intelligent systems.

His new company, Advanced Machine Intelligence (AMI Labs), is headquartered in Paris. It's building "world models", AI that learns from reality, not from text. The core architecture, JEPA (Joint Embedding Predictive Architecture), learns abstract representations of how the physical world works rather than predicting the next word. Think of it as the difference between an AI that can write about how to assemble an engine and one that actually understands how engines work.

The investor list reads like a who's who: Bezos Expeditions, NVIDIA, Temasek, Samsung, Toyota Ventures, Bpifrance, Mark Cuban, Eric Schmidt, Xavier Niel, and Groupe Dassault. Valuation: $3.5 billion.

Why does this matter beyond the science?

First, it's a genuine bet against the consensus. Every major AI company is building bigger language models. LeCun is arguing, with a billion dollars behind him, that this approach has fundamental limits. If he's right, the implications are enormous.

Second, it matters for European sovereignty. AMI is a Paris-headquartered company with French leadership, backed by French state-affiliated investors (Bpifrance, Dassault). Combined with Mistral's rise to $1B ARR and Nscale's $2B raise in the UK, a genuine European AI ecosystem is taking shape.

Third, the commercial applications matter for insurers. AMI's first partner is a medical AI company. Target sectors are manufacturing, robotics, aerospace, and biomedical; areas where LLM hallucinations are dangerous and understanding physical reality is essential.

Quick follow-up to our lead story two weeks ago. The confrontation has moved to the courts.

On March 9, Anthropic sued the Pentagon, challenging its "supply chain risk" designation, alleging unconstitutional First Amendment retaliation. On March 18, the DOD filed a rebuttal claiming Anthropic's safety "red lines" make it an "unacceptable risk to national security." Then on March 20, Anthropic produced a bombshell: evidence that the Pentagon's own official had emailed CEO Amodei saying the two sides were "very close", the day after the designation was finalized.

Industry solidarity has deepened. Over 30 employees from OpenAI and Google DeepMind and 150 retired federal judges filed amicus briefs defending Anthropic. A preliminary injunction hearing was set for tomorrow, March 24. The outcome could set precedent for whether the government can punish AI companies for maintaining safety guardrails.

Beyond the actual ethical consideration, this is an interesting and fresh change of stance from an industry that - since his election last year - had shown strong tie to POTUS and his government.

I want to close with something different this week. Anthropic published what it calls the largest qualitative study ever conducted: 80,508 Claude users across 159 countries and 70 languages, interviewed in depth about what they want from AI, what it's already done for them, and what scares them. It is worth your time.

The headline finding cuts against how most companies frame AI adoption. When asked "If you could wave a magic wand, what would AI do for you?", only 19% said they wanted AI to help them do better work. The rest wanted their lives back: 13.5% wanted help managing the mental load of daily life, 11% wanted to reclaim time for family and relationships, 13.7% wanted personal transformation.

That's a striking mismatch with most enterprise AI strategies, which are built almost entirely around efficiency and productivity metrics.

The study's most useful concept is what Anthropic calls "light and shade": the things people love most about AI are the very same things they fear. People who value AI for emotional support are three times more likely to also fear becoming dependent on it. Entrepreneurs thrilled by economic empowerment watch their freelancer friends get squeezed out. A lawyer in Israel said: "I use AI to review contracts, save time… and at the same time I fear: am I losing my ability to read by myself?"

Hope and alarm don't divide people into camps. They coexist as tensions within each person. Any change management approach that treats AI adoption as a binary — resisters vs. champions — is missing this entirely.

A few other findings worth noting for us: job displacement was the top fear (22.3%), spread evenly across job categories, not concentrated among blue-collar workers. Unreliability was the primary barrier to adoption (26.7%), ahead of every other concern. People in Sub-Saharan Africa and Southeast Asia are significantly more optimistic about AI than those in Western Europe. And independent workers reported more than triple the rate of economic empowerment from AI compared to salaried employees — suggesting AI's benefits are accruing unevenly, with entrepreneurs capturing gains that traditional employees are not.

The full study includes a searchable "Quote Wall" where you can filter voices by region, concern, and vision. I'd recommend spending ten minutes with it. It's a useful corrective to the abstraction that dominates most AI strategy conversations and a reminder that the people we're building these tools for are more thoughtful, more conflicted, and more interesting than any executive summary gives them credit for.