Jensen's 7.5 Million Agent Bet

Edition 168 | March 23, 2026

Coinbase CEO showing agentic payments happening in real time. Still early, but you can start to see where this is going.

Welcome back to Building AI Agents, your biweekly guide to everything new in the field of agentic AI!

In today’s issue…

• Jensen Huang lays out NVIDIA's bet on an agent-powered workforce

• Google turns AI Studio into a full-stack app builder

• A rogue AI agent inside Meta exposes internal data for two hours

• Stripe's coding agents now ship 1,300+ pull requests per week

• An AI agent designs a working CPU from a single prompt

…and more

🔍 SPOTLIGHT

At every GTC, Jensen Huang sells a vision of the future and the hardware to power it. This year was different. His two hour keynote spent surprisingly little time on GPUs. Instead, he laid out something much larger: a vertically integrated stack for the agent economy.

Nvidia, in his telling, is not just in the business of selling chips anymore. It is in the business of selling the infrastructure agents run on, an “AI Factory”. And the number he used to anchor that vision, 75,000 employees alongside 7.5 million AI agents inside Nvidia within a decade, was not a flashy line. It was a statement of how quickly he thinks this shift is coming.

That helps explain the phrase he kept returning to during GTC: the inference inflection point. Put simply, inference is when a trained model is actually being used to generate an answer, make a decision, write code, take an action, etc. For the last few years, the AI boom has mostly been about training larger and larger models. But if Jensen is right about agents, the next phase will look very different. The next wave of demand will come from agents running constantly, not just training models.

That is the logic behind the 7.5 million figure. Jensen still believes AI tools will get better but he is betting that agents will become a normal part of how work gets done, a picture of what he thinks companies will look like in the near future.

With that context, some of Jensen’s other comments start to make more sense. On the All-In podcast, he said he would be “deeply alarmed” if a $500,000 engineer was not using at least $250,000 worth of tokens per year. The exact number he used matters less than the meaning: token consumption, in his view, will become a standard input to knowledge work. The future he is describing is not one where people occasionally open ChatGPT. It is one where every high value worker is surrounded by agents doing meaningful work all the time.

That is also why Nvidia is trying to position itself as more than a chip company. At GTC, it did not just talk about faster hardware. It showed pieces of a broader stack for an agent-heavy future: compute systems, inference infrastructure, OpenClaw for running agents, NemoClaw for enterprise governance, and simulation tools for physical AI, including the Frozen Olaf demo. Jensen’s goal is to help power the entire system that keeps agents running.

That strategy only makes sense if Jensen is right that agents are the next major unit of software work. As Jensen described, most agents are still brittle, supervised, and far from reliable enough to become true digital workers at scale. But that is the point. He is not describing the present. He is placing a very large bet on what comes next.

The question is what happens if Jensen is right about the speed. He is not just making a long-term bet on agents. He is signaling that he believes the transition will happen fast, fast enough that Nvidia has to build the infrastructure for it.

As always, keep learning and building!

—AP