#494 — Jensen Huang: The 100,000 Agents Problem
Jensen Huang has a definition of AGI, and it fits neatly inside a sentence he made up on a podcast.
The exchange happened near the end of Lex Fridman Podcast #494, published March 23, 2026. Fridman asked Huang whether an AI could start, grow, and run a company worth more than a billion dollars. Huang responded: "I think it\u2019s now. I think we\u2019ve achieved AGI." His definition of the milestone: an AI that can launch something viral and hit $1B once. Not passing human-level tests. Not matching expert performance across domains. Not any of the definitions AI researchers have debated for years. Just: can it make something that takes off?
Huang immediately opened an escape hatch in his own logic. When Fridman followed up on whether an AI could run a company, Huang said: "You said a billion, and you didn\u2019t say forever." Then, in the same interview, he said this: "The odds of 100,000 of those agents building Nvidia is zero percent."
That is the whole interview, in two sentences.
NVIDIA is not worried about anything. The numbers tell you why. Huang said Nvidia has $1T in purchase orders through 2027, double the prior forecast. Revenue this quarter: $78B, up 77% year-over-year. Eleven straight quarters above 55% growth. The stock is down from its March peak, which puts Nvidia at roughly $4T \u2014 still the most valuable company in the world by a wide margin.
On hardware: Vera Rubin, named for astronomer Vera Rubin, is coming in the second half of 2026. It has 1.3M components per rack and delivers 10x the performance-per-watt of the current Blackwell generation. The Vera CPU delivers 2x energy efficiency versus x86 and 3x memory bandwidth per core. Rubin Ultra \u2014 the next architecture \u2014 is targeting 2027.
Nvidia also unveiled Groq 3 LPU at GTC 2026 (March 16\u201319), a chip that came via a roughly $20B technology licensing deal with Groq. Groq continues operating as an independent company; Simon Edwards stepped into the CEO role after the deal closed. Sunny Madra, Groq\u2019s former president, and Jonathan Ross, the founder, joined Nvidia along with the engineering team. The structure \u2014 a non-exclusive technology license plus talent acquisition \u2014 is one of several Big Tech deals structured this way in recent years to sidestep antitrust review.
Huang was candid about what he doesn\u2019t know. Asked about ASML, TSMC packaging, and SK Hynix as supply chain bottlenecks: "No," he said. "Because I told \u2018em what I needed. They understood what I need. They told me what they\u2019re gonna go do, and I believe them what they\u2019re going to do."
The AGI definition question is not new. At the 2023 NYT DealBook summit, the working definition was: software capable of passing tests requiring human-level intelligence, with a five-year timeline. Huang\u2019s 2026 definition \u2014 can it build a viral app \u2014 is different in kind, not degree. It is a definition any successful consumer software product has already met, going back decades.
Huang\u2019s definition conveniently coincides with what Nvidia builds. This is worth noticing, not because it discredits the hardware story \u2014 the numbers are real and they stand on their own \u2014 but because the AGI declaration and the hardware numbers are being covered as a single narrative. They are not.
What Huang demonstrated in the interview is more interesting than a milestone: he is a systems thinker who runs a company with 60 direct reports across memory, CPUs, optical engineering, GPUs, and algorithms, and has for 20 years been right about where computing is going. CUDA was added to GeForce in the mid-2000s; the stock dropped from $8B to $1.5B during the transition. "NVIDIA is the house that GeForce built," Huang said. He meant it as history. It reads as prophecy.
The 100,000 agents line is the one that should get attention \u2014 not as proof of anything, but as a window into how the most powerful voice in AI infrastructure thinks about what his products cannot do. Huang was explicit: he does not believe current AI agent systems could design and build Nvidia, even with 100,000 of them. The margin on that claim is worth sitting with.
Whether that view is right or wrong, conservative or too generous to human judgment \u2014 that is a different conversation. One that Huang\u2019s definition of AGI was never going to start.
Sources: Lex Fridman Podcast #494 (March 23, 2026), Nvidia press releases, CNBC, Reuters, Groq press release, Yahoo Finance.