What would actually change on your screen tomorrow if AGI showed up today? Not in a keynote. Not in a post. On your screen, in the tool you already pay for, on the task you already gave up on automating. Hold that answer, because Jensen Huang just declared that the moment has arrived, and I’d like to know what any of us are supposed to do with that.
The Nvidia CEO posted congratulations to the OpenAI team over GPT-6 Astra, noting it was trained on roughly 100,000-plus Grace Blackwell NVLink72 systems, tracing a line from ChatGPT to o1 to Astra in four years. “AGI has arrived,” he wrote. He also mentioned 400,000 GPUs coming. In a March 2026 interview, Lex Fridman asked him whether an AI that could start, build, and run a billion-dollar company was achievable within the next 20 years. Huang’s answer: “I think it’s now.”
Then he downplayed it. Which is the part everyone skipped.
The claim and the invoice arrive in the same sentence
I review AI tools for a living. That means I spend most of my week watching agents fail at things a competent intern would finish before lunch, so I’ve developed an allergy to declarations that arrive bundled with a hardware order. The AGI announcement and the 400,000-GPU note sit in the same breath. I’m not accusing anyone of lying. I’m pointing out that the person calling the finish line also sells the track, the shoes, and the stadium.
This matters because AGI has no agreed definition. It’s generally described as software that can handle any intellectual task a human can, or at least most thinking work. Nobody agrees on when that threshold gets crossed, or who gets to certify it. Huang picked a definition he liked, the billion-dollar-company test, and answered it in the affirmative. He’s entitled to. So is everyone who thinks that’s absurd. The term is a Rorschach test with a marketing budget.
What makes it slippery is that “AGI arrived” is unfalsifiable in either direction. There’s no benchmark suite, no audit, no regulator stamping a certificate. So the phrase does what vague phrases always do in this business: it moves money, shifts expectations, and gives procurement teams a reason to sign something large before they’ve tested anything small.
What I test instead
Since nobody can define AGI, I’ve stopped trying. Here’s what I actually measure when a new model or agent lands, and none of it requires philosophy:
- Can it complete a multi-step task without a human unsticking it halfway through?
- Does it fail loudly or quietly? Quiet failure is worse, and it’s still the norm.
- Does it hold state across a long job, or does it forget the constraint you gave it twenty minutes ago?
- When it’s wrong, does it say so, or does it produce something confident and useless?
- Does the second run match the first? Reliability beats peak capability in every real workflow I’ve seen.
Notice what’s missing. There’s no line item for “displays general intelligence.” Because if a system passes those five checks on your actual work, you don’t care what it’s called. And if it fails them, no announcement makes your Tuesday better.
The honest read on Astra
I haven’t put GPT-6 Astra through my own testing, so I’m not going to pretend I know how it performs. What I can say is that the pattern from ChatGPT to o1 to Astra in four years is real, fast, and genuinely notable, and that fast progress and “AGI has arrived” are different statements. One is a trend line. The other is a claim about a destination that nobody has mapped.
I’d also gently note that Huang has made versions of this call before. That’s not a gotcha, it’s context. Repeated declarations from an interested party should be graded on a curve.
What to do Monday morning
If you run a team that buys these tools, the AGI headline should change exactly nothing about your process. Keep your evaluation use. Keep your regression tests. Keep a human in the loop on anything that touches money, customers, or code that ships. The models are improving quickly enough that you should re-test quarterly. They are not improving in a way that makes verification optional.
If you’re an individual user, the practical question stays the same one it’s always been: does this thing save you time on the specific work you do, measured over a week, not a demo? That’s a question you can answer yourself. It’s more useful than anything a chip executive can tell you.
AGI may well have arrived. I’ll believe it when the agent finishes the job and I don’t have to check its work. Until then, I’m still checking.
🕒 Published: