Jensen Huang says AGI arrived. Nobody agreed on the finish line

On September 6, Nvidia CEO @JensenHuang posted that OpenAI's newest model had crossed the line. His words: "From ChatGPT to o1 to Astra in 4 years. AGI has arrived." He credited the hardware in the same breath, saying GPT-6 Astra trained on more than 100,000 Nvidia Grace Blackwell NVLink72 systems, with another 400,000 GPUs coming online next.

So the man selling the compute is the one announcing that the compute produced general intelligence. That is not a scandal, but it is worth holding onto while you read the rest of this.

The claim, and the missing definition

Ninety minutes later, @GaryMarcus answered. He wrote that it was "Sad to see Jensen claim that AGI has arrived, with no evidence and no definitions," and pointed Huang at agidefinition.AI, a framework co-authored by Dan Hendrycks, Yoshua Bengio and others including Marcus himself. His objection was procedural rather than dismissive. By conventional definitions Astra still falls short, and declaring a win without saying what winning is muddies the water for everyone downstream.

Marcus is a professional skeptic and usually gets discounted for it. Here the discount does not apply, because he is not arguing about whether Astra is impressive. He is arguing that AGI has no agreed threshold, which means the sentence carries almost no information. Two people can read Huang's post and one of them hears a scientific result while the other hears a sales figure, and neither one is misreading it.

The number nobody stopped on

Marcus posted again on September 7, and that one landed harder. He asked whether Huang meant "100K+ GPUs, or 100k+ NVLink72 server racks," noting that each of those racks holds 72 Blackwell GPUs. Huang's phrasing points at racks. If it is racks, Marcus figures the hardware behind Astra would sell for something on the order of a quarter trillion dollars.

That is a 72x spread in the most quoted technical detail of the week, and it ran through dozens of outlets without anyone pausing on it. I do not know which reading is correct, and nobody outside Nvidia and OpenAI does either. The number everybody repeated turns out to be ambiguous inside its own sentence, which tells you something about how carefully any of this gets read.

My honest read

Huang is not lying. He is doing his job. Nvidia's business case rests on every frontier release being understood as proof that more GPUs buy more capability, and "AGI has arrived" is the tightest possible compression of that case. The hardware figure is real. The jump from a hardware figure to a civilizational milestone is marketing, and you can say that without calling anyone a fraud.

What bothers me is the second-order effect. When the loudest definition of AGI comes from the supplier, the word stops being useful for deciding anything, and it ends up describing a mood instead of a capability.

What this means if you run a business

Some version of "AGI is here" will show up in a sales call you take this month, probably attached to a price that assumes you believe it.

It does not matter. Nothing about how you quote a job changes based on whether Astra clears an academic bar. The questions that decide whether an AI tool earns its keep are boring and specific:

• Does it handle this exact task on your real inputs, not a demo?
• How often is it wrong, and what breaks when it is wrong?
• What does it cost per month against the hours it gives back?

You can answer all three this week with a two-hour test, and none of them requires a position on AGI. The businesses pulling real value out of AI right now are mostly ignoring frontier benchmarks. They wired up one annoying handoff, watched it for a month, then wired the next one.

There is one place the definition fight does matter, and it is regulatory. Massachusetts is arguing over rules that would apply to frontier labs above a revenue and spending threshold, with Anthropic backing them while OpenAI and Google push back. Notice that those thresholds are written in dollars rather than in capability claims. My read is that dollars are the only part anyone can actually measure.

For the rest of us the job has not changed since last year: find the task, test the tool, watch the error rate, then keep it or kill it. If you want a second opinion on which task to start with, New Face Design does a free process audit. We map what happens between a lead coming in and the invoice getting paid, then tell you where a tool would genuinely save time and where it would just add a subscription.