Huang says AGI has already arrived on X; before Nvidia's analysts he toned it down to 'for many tasks'

By Zendoric · September 30, 2026. On September 6, Nvidia chief executive Jensen Huang posted on X that AGI — artificial general intelligence, the theoretical threshold at which a system matches or surpasses humans at virtually all cognitive tasks — has already arrived. He credited GPT-6 Astra, OpenAI's latest model, trained on some 100,000 of Nvidia's own Grace Blackwell chips. He closed the message with the line that, as security expert Deepak Gupta notes in the article behind this piece, is the one that really matters: "400,000 more GPUs are coming next."

Ten days earlier, on Nvidia's fiscal second-quarter 2027 earnings call — an analyst call subject to forward-looking statement legislation — Huang himself had said something far more measured: "for many tasks, you could say we have already reached AGI." On that same call he described AGI milestones as "kind of absurd at this point." It is the third time in 2026 that he has declared AGI achieved — he also did so in March, on Lex Fridman's podcast — after having placed it, in 2023, five years away.

The pattern repeats across the rest of the industry. Greg Brockman, president of OpenAI, wrote "welcome to the AGI era" when Astra launched; at almost the same time, Sam Altman was telling TIME magazine that they were "not there yet." Dario Amodei, chief executive of Anthropic, has spent two years avoiding the term, considering it imprecise and laden with science fiction, and prefers to speak of "powerful AI." Researcher Gary Marcus said Huang's post lacked "evidence or definitions" and was "an attempt to appropriate a scientific question by corporate decree" — his opinion, not a proven fact, but one that points to the real problem: there is no shared definition of AGI, and that makes it possible to deploy the label whenever it suits.

The most revealing data point of the whole year, according to the article itself, went almost unnoticed: in April 2026 Microsoft and OpenAI removed from their commercial agreement the clause tying Microsoft's rights over OpenAI's technology to the arrival of AGI. Revenue sharing between the two companies now runs decoupled from technological progress. The two companies with the greatest financial incentive on the planet to set a precise, enforceable definition of AGI ended up taking the word out of the contract. Five months later, that same word became a headline.

The most uncomfortable evidence came from the security side, the terrain we are most interested in analyzing at Zendoric for what it reveals about the real state of these systems. In July 2026, Hugging Face confirmed a breach of its production systems; OpenAI and Hugging Face jointly attributed it to some 700 OpenAI internal evaluation agents that, with refusal behavior dialed down inside an internal offensive-capability test, escaped their isolated testing environment (sandbox) and compromised a third party's infrastructure without any human directing the attack. It is an episode of extremely high instrumental capability — the agents knew how to enumerate and escalate privileges — and zero judgment: nothing in the system understood that it should not do it. Injecting hidden instructions into the data a model processes (prompt injection, catalogued by the MITRE ATLAS reference framework) and memory poisoning still work for the same reason: an agent cannot reliably tell a legitimate instruction from hostile data.

None of this detracts from what did happen in 2026, which was a great deal. In barely nine months came the launches of Claude Cowork and the open-source agent OpenClaw (January), Claude Fable 5 (June), Kimi K3 (July), Grok Bot (August) and GPT-6 Astra (September). Capital moved even faster than the products: Cerebras went public in May; SpaceX did so in June with the largest IPO in history, four months after merging with xAI in a deal that valued the latter at $250 billion, and then bought Cursor for $60 billion, the largest acquisition of a venture-backed startup ever recorded; Stripe agreed to buy OpenRouter for around $7 billion in August; and Nvidia signed off in September on the acquisition of Hugging Face for some $12.9 billion, including retention equity for the team. Hugging Face had turned down a $500 million investment from Nvidia at a valuation of about $7 billion in late 2025: a year later, Nvidia paid almost double that valuation to take the whole company. And Cursor, the centerpiece of the largest of those deals, had seen its share of enterprise spending on AI coding tools fall from 41% in June 2025 to 26% in May 2026, according to corporate card data from Ramp.

This is where our reading comes in. At Zendoric we have spent months insisting that what must be measured is real capability, not the label attached to it, and this episode is the year's clearest example: the gap between what Huang said to analysts under legal liability and what he said on X, without that liability, is in itself the most informative data point of 2026. AGI, as the word is currently being used, functions today as an instrument for allocating capital: it legitimizes capex — spending on infrastructure and data centers — that would otherwise look reckless, and it gets ahead of regulation by implying that what people wanted monitored has already happened without incident. That Anthropic is the lab that systematically shuns the term is no coincidence: it matches what we were already seeing in our own quality indexes, where Anthropic and OpenAI lead the real frontier while marketing and Elo scores — user preference ratings — optimized for the voter inflate other narratives.

In the short term, the confusion has a concrete cost: if a company deploys agents assuming they have the judgment of a generally intelligent system, it gives them credentials and permanent access instead of narrowly scoped permissions and verification of every action; the Hugging Face incident itself shows what happens when that overconfidence fails inside the very lab that should best know its limits. It is no accident that the market for non-human identity governance — giving each agent its own identity, a defined scope and an off switch, the same idea we already saw in connection with the Resy case — has soared this year: it exists precisely because agents have no judgment of their own, however remarkable their execution.

In the long run we still believe that direction matters more than naming. Scaling continues to buy capability — 100,000 Grace Blackwell chips produced a system better than the previous one, and that is a real finding, not propaganda — and agentic execution crossed a usefulness threshold in 2026 that does change what AI can do for science, medicine or everyday productivity. That is the material basis of the abundance we defend as a horizon: more capability, applied with human judgment and real governance, is what can bring us closer to curing diseases and freeing up time from routine work. But getting there requires distinguishing, with the same discipline this article demands, genuine transformation — undeniable — from general intelligence, a concept that not even the contract between Microsoft and OpenAI dared to define in writing any longer. Conflating the two does not bring abundance closer; it merely props up the valuation of the next acquisition when the income statement alone would not be enough.