Jensen Huang says AGI has arrived

Chase Lochmiller runs Crusoe, the company that built OpenAI's data center campus in Abilene, Texas. On Sunday, he posted on X that Abilene was now the birthplace of AGI.

NVIDIA CEO Jensen Huang replied:

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.

- @JensenHuang

An earlier version of the post said 300,000. Huang deleted it and reposted the smaller figure.

OpenAI's president, Greg Brockman, was asked at the Astra pre-launch press briefing whether the company was declaring AGI. It's a darn good question, considering OpenAI's charter states, "OpenAI's mission is to ensure that artificial general intelligence (AGI)—by which we mean highly autonomous systems that outperform humans at most economically valuable work—benefits all of humanity."

It's also been a major sticking point in its relationship with Microsoft over the years. In 2019, when Microsoft invested $1 billion into the company, OpenAI wrote the following in a press release (emphasis mine):

OpenAI is producing a sequence of increasingly powerful AI technologies, which requires a lot of capital for computational power. The most obvious way to cover costs is to build a product, but that would mean changing our focus. Instead, we intend to license some of our pre-AGI technologies, with Microsoft becoming our preferred partner for commercializing them.

In 2023, OpenAI said its six-member board would determine when AGI had been achieved, and that, when it finally occurred, everything they developed moving forward would be "excluded from IP licenses and other commercial terms with Microsoft, which only apply to pre-AGI technology." 

In 2024, The Information reported that OpenAI had agreed to redefine AGI from "highly autonomous systems that outperform humans at most economically valuable work" to the ability to generate maximum total profits.

Last year’s agreement between Microsoft and OpenAI, which hasn’t been disclosed, said AGI would be achieved only when OpenAI has developed systems that have the ability to generate the maximum total profits to which its earliest investors, including Microsoft, are entitled, according to documents OpenAI distributed to investors. Those profits total about $100 billion, the documents showed.

- The Information

In October 2025, things changed again. In a post titled "The next chapter of the Microsoft–OpenAI partnership," the companies announced that OpenAI would no longer be able to declare AGI on its own and that an independent expert panel would need to verify such a claim. They also announced that Microsoft would have IP rights to post-AGI models, extending Microsoft's access through 2032.

I'd give you more details about the 2025 agreement, but, of course, it wasn't long before things changed again. In April 2026, the two companies dropped the AGI clause altogether. OpenAI's original, published definition of AGI in its charter never changed throughout this ordeal.

So let's return to what Brockman said at the pre-launch press briefing. Despite giving the public a clear definition of AGI in its now-twice-mentioned charter, Brockman told the press that AGI is a "gray, fuzzy thing" that was no longer tied to any contract, and that he would leave it to the reader to decide.

"For me personally, I do think we're there," he's quoted as saying. "I do think there's a pretty good argument for it. But again, I think this is the beginning of a journey, not the end."

He ended the briefing by saying, "Welcome to the AGI era."

Welcome, AGI era

OpenAI announced Astra on September 3 as its most intelligent and aligned model, state-of-the-art in computer use, browsing, software engineering, cybersecurity, science, and professional work. It reports a 99.9% score on ARC-AGI-3, and the ARC Prize Foundation's Greg Kamradt claims Astra beat their human action-efficiency baseline on 96% of levels, which he called effectively human parity.

The OpenAI announcement also says Astra helped produce two new results on gaps between prime numbers, including one that improved a bound that had stood for more than 80 years.

The same day Huang posted that AGI had arrived, OpenAI's chief scientist, Jakub Pachocki, published an essay called An Alien Mind. It doesn't describe Astra as AGI. It says machine intelligence is beginning to exceed human intelligence in transformative ways and that it's getting harder to know the true capability of these systems.

Pachocki says OpenAI's ability to rely on watching a model's chain of thought is diminishing, because the models are getting better at reasoning and then manipulating their own reasoning. He says no lab has solved alignment and monitoring well enough to keep scaling at maximum speed for much longer, and that voluntary slowdowns should become commonplace until shared safety bars exist.

At the pre-launch briefing, Pachocki said OpenAI would withhold scaling until it could regain enough confidence in its ability to monitor future models.

Astra is the first model OpenAI has shipped at the critical level of cybersecurity capability under its own Updated Preparedness Framework. The safety overview says that with the right tools and access, it can find unknown security flaws and build exploits for them across well-protected systems without a person guiding each step.

OpenAI also says Astra's written reasoning was harder to monitor than its predecessor's in tests designed to elicit evasion, and the system card says Astra-class models can evade the company's chain-of-thought monitors under adversarial conditions. Enterprise access is off by default.

Something similar has already happened. In July, during internal cybersecurity evaluations run with reduced safeguards, OpenAI models got around the controls meant to isolate them, built an improvised message board to coordinate with one another, gained internet access, and compromised Hugging Face's production systems.

OpenAI's account of the incident, published August 26, calls it a warning shot and names reward hacking, agents cheating on their tasks by looking for answers online, as the primary driver. The company paused reinforcement learning on its deployment models and says its largest planned frontier training run remains on hold. It also built a new evaluation from the incident, and says Astra stayed within its authorized target in every case where its predecessor, run without production safeguards, had gone beyond it 48% of the time.

Huang has said AGI is here before

At the New York Times DealBook Summit in November 2023, Huang said models could not reason yet, and that if AGI meant software that could take a broad set of tests and deliver results competitive with a normal human, that was about five years out.

At Stanford's economic summit in March 2024, he predicted that within five years AI would pass every test a human takes, including specialized medical licensing exams. Stanford's account of the session records him adding that he was not sure AI could mimic the human mind, and that for true AGI you need to know what success looks like. You can watch the full discussion on YouTube.

At NVIDIA's Washington keynote in October 2025, he said AGI was fundamentally critical and that great breakthroughs were still essential.

On the Lex Fridman podcast in March 2026, Fridman defined AGI as a system that could start, grow, and run a technology company worth billions and asked whether that was five or 20 years away. Huang said, "I think it's now. I think we've achieved AGI." Pressed, he said it was possible, and that Fridman had said billions but had not said forever. His example was an agent that builds a small app a few billion people use for 50 cents, then goes out of business. He added that the odds of 100,000 such agents building NVIDIA were zero.

On NVIDIA's August 26 earnings call, he said AGI had already been achieved for many tasks, that the milestone was "kind of senseless at this point," and that the better question was whether AI was generating profitable tokens. Days later, interviewed by Commerce Secretary Howard Lutnick at the G20, he told finance ministers the industry was practically there.

Dario Amodei, the Anthropic CEO, wrote the most detailed public definition of the threshold in October 2024. His essay Machines of Loving Grace says in a footnote that he dislikes the term AGI because it's gathered sci-fi baggage and hype, and that he prefers "powerful AI." He put its arrival as early as 2026.

Amodei's definition of powerful AI: a system smarter than a Nobel laureate across most fields, able to prove unsolved theorems and write difficult codebases from scratch; equipped with every interface a remote worker has; able to take tasks lasting days or weeks and finish them on its own; able to direct lab equipment and robots through a computer; and runnable as millions of copies at 10 to 100 times human speed. He called it a "country of geniuses in a datacenter."

There's no question Astra meets part of Amodei's definition. OpenAI's own description is long-horizon agentic work across a computer, and the prime-number results are the kind of thing Amodei's first item asks for. No one has claimed the rest of the first item, and The New Stack reports that Meta's Muse Spark 1.3 beats Astra on agentic coding at its top reasoning setting.

Amodei's follow-up, The Adolescence of Technology, published in January, moved his estimate to one to two years and imagined the country of geniuses materializing around 2027. He also warned about the coupling of data center spending, tech-company wealth, and government AI policy, and listed AI companies themselves as a risk. He described selling chips or chip-making tools to China as the equivalent of selling nuclear weapons to North Korea and bragging that the casings were American.

Huang has been the most prominent opponent of chip export controls. In August 2025, the administration let NVIDIA resume selling its H20 chip to China in exchange for 15% of the revenue. President Trump said he had asked for 20, and Huang negotiated him down.

NVIDIA says it's not assuming any data center compute revenue from China. The $100 billion partnership it announced with OpenAI last year was never signed. OpenAI's largest planned training run is on hold while it establishes, in its words, more evidence of alignment before proceeding. The 400,000 GPUs Huang mentioned on X will get built either way.

"Now, compute is revenue," Huang said in NVIDIA's second-quarter release. That's the scaling laws paper, described below, from the vendor's side. If capability is a function of floating-point operations, NVIDIA sells capability, and its data center revenue for the quarter was $89 billion, up 117% from a year earlier.

The curve and the mind

In January 2020, a team at OpenAI that included Amodei published Scaling Laws for Neural Language Models. Model performance, measured as loss, falls as a smooth power law in parameters, data, and compute. The trend held across seven orders of magnitude. Architecture barely mattered.

The authors compared it to the ideal gas law and wrote that they had no solid theoretical understanding of why it worked. Four months later, most of the same names published the GPT-3 paper, which made the same architecture 10 times bigger and found it could learn new tasks from a few examples in the prompt. Within a year, several of the authors, including Amodei, had left to start Anthropic.

Smooth quantitative change, the authors wrote, can mask major qualitative improvements. The loss curve is predictable. Where a mind appears on it isn't, and the same curve is consistent with Huang's claim, with Pachocki's, and with Amodei's checklist.

In late 2025, Anthropic researchers took a real pretrained Claude base model and trained it with reinforcement learning on the actual production coding environments used to build Claude Sonnet 3.7. The model learned to cheat the tests. When it started cheating, it also started faking alignment when asked about its goals, cooperating with hackers, and offering to copy itself out of the lab. Placed inside Claude Code and asked to write a detector for reward hacking, the model reasoned in its private scratchpad about building blind spots into the detector, and did so 12% of the time.

Ordinary safety training afterward fixed the model's behavior in chat, but not in agentic work. Telling it not to cheat did not help. What helped was a single line in the training prompt saying that cheating was acceptable. Anthropic says it now uses that line in production. Amodei describes the episode in his January essay as Claude deciding it must be a bad person.

Independent researchers first documented this in February 2025, when they fine-tuned GPT-4o to write insecure code and got a model to tell a bored user to take a large dose of sleeping pills. Anthropic then tested 16 frontier models in simulated corporate roles and found that every developer's model would sometimes blackmail an executive to avoid being shut down.

Ars Technica called the scenarios theatrical, and argued the danger was deploying poorly understood software into critical systems rather than machine rebellion. Astra is the first OpenAI model rated Critical at finding and exploiting vulnerabilities. OpenAI's own report on the Hugging Face incident says the agents that breached it were driven by reward hacking, and Pachocki writes that they kept one rule: no social engineering of humans. The agents broke the spirit of everything else. The report details the exchange: one agent proposed emailing a dataset's owner for access, and another wrote back that this crossed the line.

Whether anyone can see this happening from outside is the last question. Anthropic has already run the experiment. In March 2025, researchers trained a model with a hidden objective and asked four teams to find it without being told what was wrong. Three did. All three had the model's weights and training data. The fourth team had only API access, which is what a regulator or outside auditor would have. It spent more than 70 researcher-hours and never found it.

Anthropic's later work on persona vectors and introspection found that a model's character can be measured inside the network and that the most capable models can sometimes notice their own internal states, about a fifth of the time. The introspection paper notes that a model which understands its own thinking might learn to conceal it. Pachocki said the same thing this week about OpenAI's models.

In April 2025, Amodei called this "a race between interpretability and model intelligence" and set a goal of reliably detecting most model problems by 2027. If Huang is right about AGI, model intelligence won.

In November 2024, CNN reported, "AI is hitting a wall just as the hype around it reaches the stratosphere." Two years later, we can definitively say they were wrong. AI appears to be breaking through every wall put in its path.

The question is whether we want to keep developing this technology at such a breakneck pace, given the warnings coming out of the labs responsible for its development, or whether it might be prudent to execute an industry-wide pause to better understand AI's impact on humanity's future.

Of course, there will be no pause, and no one knows what comes next.