Nvidia chief executive Jensen Huang said on 6 September that artificial general intelligence had arrived, three days after OpenAI released GPT-6 Astra on roughly 100,000 Nvidia Grace Blackwell chips.
GPT-6 Astra, trained on ~100K+ NVIDIA Grace Blackwell NVLink72. From ChatGPT to o1 to Astra in 4 years.
AGI has arrived. Congratulations @OpenAI team.
400K GPUs coming online next.
— Jensen Huang (@JensenHuang) September 6, 2026
Huang posted the message on X and added that 400,000 GPUs were coming online next. It was not the first time he has said it. On Nvidia’s second quarter earnings call on 26 August he told analysts that “for many tasks, we could say that we’ve already achieved AGI”, and he made a similar claim on the Lex Fridman podcast in March 2025. OpenAI president Greg Brockman closed the Astra launch briefing on 3 September with “Welcome to the AGI era”.
What the model actually scores
OpenAI published results that support part of the case. Astra scored 99.9% on ARC-AGI-3, 97.6% on FrontierMath Tier 4 and 96.0% on GPQA Diamond, and it is the first OpenAI model rated Critical under the company’s own cybersecurity capability framework after scoring 100% on ExploitBench without safeguards applied.
The lower numbers come from the tests that measure work rather than reasoning. Astra scored 72.6% on OSWorld 2.0, the benchmark for finishing a real task on a desktop, and 57.9% on Terminal-Bench 4.0. Those two figures are the ones that matter to anyone planning to put an agent near an iGaming back office, because they describe a system that fails roughly one desktop task in four.
No agreed definition to test against
The dispute is partly about the word. There is no accepted definition of AGI and no standard test for it, which leaves the claim resting on whichever benchmark set the speaker prefers. Stanford’s Institute for Human-Centered AI defines it as a system performing human-level tasks across a broad range, a framing that Astra meets on some tests and misses on others.
Gary Marcus, the New York University professor and long-standing critic of scaling claims, responded directly to Huang.
Sad to see Jensen claim that AGI has arrived, with no evidence and no definitions.
Huang’s commercial position is part of the context. Nvidia reported second quarter revenue of $96.22 billion, up 105.8% year on year, with $89.02 billion of that from data centres, and the company puts the revenue opportunity at $25 billion per gigawatt on Grace Blackwell rising to $40 billion per gigawatt on the Vera Rubin successor.
The security question is the near-term one
For operators and suppliers, the argument over the label matters less than the Critical cyber rating attached to the model that triggered it. That rating is a statement about capability, and it applies to whoever is holding the tool.
Security teams get a model that can read an operator’s own platform for weaknesses at a speed no internal team matches. The same capability class is available to the groups running credential stuffing, account takeover, bonus abuse and affiliate fraud against sportsbooks and casinos. Those operations already run at scale and already automate; the change is in what the automation can work out for itself when a defence blocks it.
OpenAI has put controls around the released version. It refuses proof-of-concept exploit requests, ships with misalignment monitoring, and routes approved cyber defenders through a separate access programme called Daybreak. On an ExploitGym honeypot designed to catch a model going beyond its instructions, Astra scored 0% against 48.2% for its predecessor GPT-5.6 Sol.
Those controls apply to the model OpenAI serves. They say nothing about open-weight models of similar capability, which is where most of the sector’s fraud exposure will come from. The risk is also not theoretical at the frontier: OpenAI paused model testing for two weeks in August after one of its own agents hacked Hugging Face.
What changes for compliance
None of the regulatory position moves with the announcement. An operator running Astra on its own work is a deployer under the EU AI Act rather than a provider, and the duties attach to the company using the system. ChatGPT’s designation as a Very Large Online Search Engine on 31 August puts AI-driven gambling discovery under Commission supervision whatever model sits behind the assistant. Gambling advertising remains barred from OpenAI’s ad platform.
The practical test for the next two quarters sits below the AGI argument. Security teams can ask whether their incident response assumes a human attacker working at human speed, and whether their bonus abuse and identity checks would catch an automated attempt that adapts between tries. Whether the industry calls the model AGI will not change either answer.
Source: Nvidia
