Jensen Huang has declared that artificial general intelligence has arrived. The Nvidia chief executive made the claim on Lex Fridman's podcast, in an episode framed around his company's $4 trillion valuation, and the exhibit is GPT-6 Astra. What makes the statement worth parsing is not the model but the measuring stick: Huang is not applying the research definition of AGI, and the definition he appears to be using was built for buyers rather than for reviewers.
Jensen Huang: Nvidia — a $4 trillion company and the AI revolution | Lex Fridman Podcast #494
Source: forbes.com
In the strict research sense, AGI is a system that reliably learns, reasons, adapts and applies knowledge across a wide range of unfamiliar tasks, at human level or above. Against that bar, Astra can be a large step forward and still fall short of proof. A step is not an arrival, and nothing about the system on its own settles the question.
The working definition behind Huang's claim is economic. On that reading, AI becomes general when it can do intellectual work valuable enough to create, run or substantially support a billion-dollar company. That is a threshold about revenue, not about cognition. It measures the buyer as much as the model — a system clears the bar when an organisation is willing to build a business around it.
There is a coherent case for talking this way, and it is a case about deployment rather than science. When a chief executive says AGI has been reached, enterprise adoption accelerates and large infrastructure commitments become easier to justify. It also sits close to Nvidia's own position as one of the principal suppliers of the technology used to train frontier models. The definition that declares victory is, conveniently, the definition that keeps the training clusters selling.
It also matches how Silicon Valley has long reasoned about capability: if AI is already producing visible economic value, it must possess something AGI-like. And it matches how companies actually deploy. Most enterprise AI is a portfolio — frontier models, open models and narrow specialised ones, each assigned to a task — rather than a single omniscient system. In that context, AGI can simply mean AI powerful enough to reorganise whole industries, which is a claim about industries, not about intelligence.
The critics decline the label. Researchers including Gary Marcus argue that current models, Astra among them, have not demonstrated the ability to reason, to plan, to grasp cause and effect, or to perform consistently in environments they were not prepared for. Marcus called Huang's statement a declaration of victory without evidence and without clear definitions: corporate rebranding of narrow AI rather than the achievement of general intelligence.
The absence of an agreed definition is what makes claims like this possible, and it is also what makes them unfalsifiable. Researchers, companies and policymakers all need to assess what these systems can actually do and what they can actually break, and that requires shared and more precise language than either side of this argument currently supplies.
My own reading is that the honest label for this moment is proto-AGI: an intermediate state in which AI has become markedly more advanced without reaching the destination. The evidence for the first half is real. Today's models write software, draft strategic plans, synthesise research findings and coordinate complex workflows. The evidence for the second half is just as real. They remain brittle and inconsistent, and they lean heavily on human direction. They are exceptionally good at recognising and generating patterns, and they lack full understanding, long-horizon planning, and the ability to operate autonomously in an unpredictable environment.
That is not a downgrade, it is a specification. It tells a buyer what to plan around. The useful enterprise question is not whether a particular vendor's model has crossed a threshold that nobody can define, but the four things that show up in an operating review: what the productivity gain is, what the risks are, what it costs, and who governs it.
The part of the claim that goes unexamined is its own test. If AI counts as general when it can create, run or substantially support a billion-dollar company, the obvious follow-up is which company — the definition has a testable shape and arrives without a worked example. It is also a certification issued by the supplier rather than by anyone who builds or operates the systems being certified. Huang sells the picks; the assay comes from the same counter.
Huang is right about the commercial reality, and that is the part worth taking seriously. Systems like GPT-6 Astra can generate substantial economic value and change how intellectual work is organised. What they have not done is settle the technical question, and the available evidence shows AI growing more capable, more general and more economically consequential without showing that AGI has been reached.
The gap that matters is not the one between Huang's definition and Gary Marcus's. It is the gap between what companies will buy on the strength of the word AGI and what the systems they install can actually be left alone to do. The contracts get signed against the first number; the operations run against the second.