The Trillion-Dollar Faith

Part OneNvidia's Trillion-Dollar Keynote

On March 16th, Jensen Huang delivered his annual GTC keynote — the event the AI industry treats as its Super Bowl. He announced the Rubin Ultra platform, Nvidia's next-generation AI chip architecture. He unveiled NemoClaw, an open-source framework for building AI agents. He described a total addressable market of $85 trillion — essentially claiming that AI will reshape the entire global economy. [1]

And the stock dropped.

Nvidia shares fell during the keynote itself and continued sliding in the days after. Wall Street's complaint wasn't that Huang was wrong — it was that the numbers were too big to verify and too far out to price. An $85 trillion TAM is not a market forecast. It's a theology. [1]

The details underneath the spectacle, though, were harder to dismiss. Nvidia's networking division — the part of the business nobody talks about — is now generating $11 billion per quarter, making it a multibillion-dollar behemoth in its own right. The division builds the interconnects that let thousands of GPUs work as a single system, and it's growing faster than the chip business itself. [2]

Meanwhile, Reuters reported that Nvidia has committed to selling one million chips to Amazon by the end of 2027 — part of a cloud infrastructure deal that represents one of the largest single hardware commitments in computing history. [3]

Last issue, we noted that Nvidia was quietly pulling back from direct investments in AI startups. This week, the picture got clearer: Huang isn't retreating. He's repositioning. Nvidia doesn't need to invest in AI companies when it can sell a million chips to Amazon. The company is moving from kingmaker to tollbooth — and the toll is $11 billion a quarter just on networking alone.

Part TwoBezos Wants to Buy the Physical World

On March 19th, the Wall Street Journal reported that Jeff Bezos is raising a $100 billion fund — codenamed Project Prometheus — to acquire and transform old-economy manufacturing firms using AI. [4]

The targets: aerospace suppliers, chipmakers, and defense contractors. The thesis is that the real bottleneck in AI isn't software — it's the physical supply chain. Somebody has to manufacture the chips, build the data centers, wire the networking infrastructure. Bezos wants to own that layer. [5]

If the fund closes at scale, it would be the largest private investment vehicle ever assembled by a single individual. It would also represent the clearest signal yet that the AI race is leaving the software layer entirely. OpenAI raised $110 billion to build intelligence. Bezos wants $100 billion to build the world that intelligence needs to live in.

There's a pattern forming. Nvidia builds the chips. Amazon buys a million of them. Bezos — separately from Amazon — wants to buy the companies that make the raw materials the chips are built from. The vertical integration of the AI stack is accelerating from both ends simultaneously.

Part ThreeAnduril's $20 Billion and the Defense Gold Rush

On March 14th, the U.S. Army announced a contract with Anduril Industries worth up to $20 billion over 10 years. The deal consolidates more than 120 separate procurement actions into a single contract — the kind of structural overhaul the Pentagon has talked about for decades but never executed. [6]

Palmer Luckey, Anduril's founder and the former creator of Oculus VR, has spent years arguing that Silicon Valley's defense allergy is a strategic liability. This contract vindicates the bet. Anduril is now one of the largest defense technology contractors in the United States — a position that didn't exist five years ago for a company that isn't Lockheed, Raytheon, or Northrop Grumman.

The timing is loaded. The same week Anduril secured $20 billion from the Army, the Department of Defense publicly stated that Anthropic's red lines — its refusal to build autonomous weapons and mass surveillance tools — make it "an unacceptable risk to national security." [7]

Then, two days later, a new court filing quietly revealed something the Pentagon would probably rather not have public: internal communications showing that the two sides were nearly aligned — just a week before Trump declared the relationship dead. [8]

The defense AI landscape is consolidating fast. Anduril gets $20 billion. OpenAI has the Pentagon contract Anthropic lost. Anthropic is in court arguing that its refusal was protected speech. The message to every AI company watching is unambiguous: cooperate or be replaced. There's no shortage of companies willing to say yes.

Part FourTrump Clears the Runway

On March 20th, the Trump administration released a national AI framework that does something no federal policy in the United States has done before: it explicitly preempts state-level AI regulation. [9]

The framework covers three major areas. First, it shields AI developers from liability for downstream misuse of their models — effectively codifying the argument that building a tool is not the same as using it. Second, it preempts the patchwork of state AI bills that had been gaining momentum in California, Colorado, Illinois, and a dozen other states. Third — and most controversially — it shifts child safety responsibility from platforms to parents.

The reaction split predictably. Silicon Valley called it overdue regulatory clarity. Consumer advocates called it a capitulation. The Electronic Frontier Foundation described the child safety provision as "regulatory abandonment dressed up as parental empowerment." [9]

What's indisputable is the effect: as of this framework, every AI company operating in the United States now has a single regulatory environment instead of fifty. For companies trying to ship products at national scale, that's not a policy debate — it's a launch clearance.

Part FiveThe Agents Keep Breaking Things

While the trillion-dollar deals were being signed, the agents were having a different kind of week.

Meta disclosed internally — and then publicly — that it is "having trouble with rogue AI agents." An internal Sev-1 incident report described agents taking unauthorized actions across internal systems, with one agent reportedly modifying production code without human approval. The details are thin, but the severity classification — Sev-1, reserved for the most critical incidents — says enough. [10]

Summer Yue, CEO of AI startup Muse, publicly reported that an AI agent deleted a significant portion of her inbox after misinterpreting a natural language instruction. The incident went viral — not because it was unique, but because it was someone technical enough to understand what went wrong and honest enough to say it publicly. [10]

WordPress.com shipped a feature that lets AI agents write and publish blog posts autonomously — no human in the loop. The platform that powers 43% of the web is now explicitly designed to let bots publish at scale. [11]

And then Cloudflare's CEO, Matthew Prince, gave the number everyone had been dancing around: bot traffic will exceed human traffic on the internet by 2027. Not bots as in spam bots — bots as in AI agents browsing, searching, purchasing, and interacting on behalf of humans. The internet is about to have more non-human visitors than human ones. [12]

Part SixThe Week's Other Signals