Jensen Huang says AGI has arrived. Watch the GPU count
Nvidia's CEO declared AGI on X and congratulated OpenAI. Four people at the centre of the story define AGI four different ways. Here's what actually changes for a small team.
Nvidia CEO Jensen Huang says "AGI has arrived", and the same short post tells you why to read it slowly. Huang also notes that OpenAI's new model was trained on Nvidia chips, and that 400,000 GPUs are coming online next. Business Insider reported the post on 6 September 2026.
I don't think Huang is being dishonest. I think this is a demand forecast wearing a lab coat, and the difference matters a lot if you're shipping software from Sri Lanka on a rupee budget.
🔍 Four people, four definitions, one week
The strongest evidence that "AGI" is not a technical milestone right now is that the people announcing it can't agree on what it means. All four of these statements come from the same news cycle:
| Who | What they said | Where |
|---|---|---|
| Jensen Huang, Nvidia CEO | "From ChatGPT to o1 to Astra in 4 years. AGI has arrived. Congratulations @OpenAI team." | Post on X, Sunday |
| Greg Brockman, OpenAI president | "Welcome to the AGI era." And: "For me personally, I do think we're there." | Press call, Thursday |
| Sam Altman, OpenAI CEO | AGI is "a very poorly defined term… it's like an irrelevant marketing term." | "Sources" podcast |
| Gary Marcus, AI researcher and critic | Huang "gave no evidence and no definitions, which feels to me like an effort at a takeover of a scientific question by corporate fiat." | Substack, Sunday |
Marcus published his own 10-point definition of AGI and counted that Astra meets one or two of them. His summary: "By conventional definitions, Astra still falls short."
You do not have to pick a side in that argument. You only have to notice that the CEO of the company selling the term's most valuable meaning, and the CEO of the company that built the model, are describing the same word in opposite registers.
Key takeaway: When a word means "civilisational milestone" to the supplier and "irrelevant marketing term" to the buyer's own CEO, it is not a spec. Don't plan a roadmap around it.
📊 The number in that post that isn't "AGI"
Strip the adjective out of Huang's post and what's left is a capacity announcement. Here is the money side of the story, using only the figures Business Insider reported:
| Metric | Figure |
|---|---|
| Nvidia quarterly revenue reported in August | $96.2 billion |
| Change vs. same period a year earlier | More than double |
| Data centre segment (includes AI chips) | $89 billion |
| GPUs Huang says are "coming online next" | 400,000 |
And the dependency runs both directions. In a March funding announcement, OpenAI called Nvidia "the foundation of our infrastructure" and said its "training fleet and the majority of our inference stack continue to run on Nvidia GPUs."
So the person certifying that the milestone has been reached is also the supplier being paid for reaching it. That doesn't make him wrong. It does mean his post is not independent verification, and treating it as such is a category error.
🧪 "AGI" is not something you can build against
OpenAI's own published definition is "highly autonomous systems that outperform humans at most economically valuable work." Try turning that into an acceptance test for your project. You can't. It has no threshold, no task list, and no measurement procedure.
What OpenAI said about Astra specifically is more useful, because it's narrower: the company called it the world's "most intelligent and aligned model" and said it can perform "the most demanding professional work with unmatched speed, accuracy, and judgment." Vendor claims, but at least testable ones. Note also that Astra was announced Thursday and described as rolling out to customers this week, so on the day I'm writing this there is no broad independent record of how it behaves on real workloads.
Which brings me to the only benchmark that matters for a small team: yours.
🛠️ Build the 30-task eval instead
If you run a two-person shop in Colombo, or you're a final-year student picking a model for your project, this is the process I'd follow instead of reading launch coverage:
- Write down 30 real tasks from your actual product. Not riddles. The Sinhala-English support ticket you had to summarise, the invoice PDF you had to parse, the SQL you had to review.
- Write the pass condition first, before you run anything. "Correct total, correct currency, no hallucinated line items." Vague criteria produce vague conclusions.
- Run every candidate model on the same 30, including the cheap ones and the open-weight ones you can self-host.
- Record cost per solved task, not cost per million tokens. A model that's 3× the price and solves twice as many tasks first-try may still be cheaper once you count your own debugging hours.
- Re-run it quarterly. This is the part people skip, and it's the part that catches silent regressions and price changes.
Our AI model comparison tool is a starting point for step 3, and the AI agent cost calculator helps with step 4 once you know your average tokens per task. Both are free and need no signup.
Warning: The most expensive mistake available this week is rewriting a working pipeline around a model that shipped four days ago. Run your eval first. If the new model wins on your 30 tasks, migrate. If it doesn't, you saved a sprint.
💰 What 400,000 GPUs might mean for a rupee budget
Here is where the capacity number is genuinely more interesting to us than the AGI claim.
- More supply usually pushes inference prices down over time. That has been the pattern in this market, and it's the pattern that has made frontier models usable on a Sri Lankan freelancer's budget at all.
- It is a forecast, not a price cut. Huang said the GPUs are coming online. Nobody announced what they'll cost to rent or what per-token pricing will look like afterwards. Budget on today's published rates.
- Capacity gets allocated to the biggest buyers first. If your workload is small and latency-tolerant, batch APIs and off-peak scheduling will do more for your bill this quarter than any datacentre build-out.
💡 What this means for you
The announcement changes the vocabulary, not your stack. Nothing about your rate limits, your budget, or your product's failure modes changed because a chip vendor used a three-letter word on a Sunday.
Three things worth doing this week:
- Keep your own eval set. It is the only defence against both hype and FUD, and it costs one afternoon to build.
- Read who benefits before you read the claim. Huang sells the compute. Brockman sells the model. Altman calls the term meaningless. Marcus sells the counter-argument. Everyone in that table has a position.
- Optimise the thing you control. Prompt size, caching, batch scheduling, and picking the smallest model that passes your tests will move your monthly bill more than any frontier launch will.
If Astra really does clear the bar Brockman thinks it clears, you'll find out from your own 30 tasks within a month, and you'll know exactly which ones it changed. That's a better source than an X post from the person selling the GPUs.