A veteran of Google and Microsoft engineering just called out the AI industry for celebrating a definition of human-level intelligence that most humans would never recognize, and the financial stakes riding on that definition run into the trillions.
The AGI drumbeat got louder last week, and the loudest cheerleaders were the companies selling picks and shovels into the buildout. NVIDIA (NASDAQ:NVDA | NVDA Price Prediction) sits at the center of that trade.
On a September 8, 2026 CNBC segment, X. Eyee, a partner at Offline Research who previously led engineering teams at Google (NASDAQ:GOOG, NASDAQ:GOOGL) and Microsoft (NASDAQ:MSFT), offered a more skeptical read on what the latest model releases actually mean. Eyee praised OpenAI’s new Astra model while rejecting the framing that any of it delivers artificial general intelligence in a form the average person would recognize.
The quote worth sitting with: “The industry is celebrating human level intelligence with a definition that leaves most of humanity out.” That is a direct challenge to the story propping up capital spending forecasts, hyperscaler capex, and the multiple you pay for anything AI-adjacent.
The stakes are visible in the capex forecasts: Jensen Huang told investors that top-five hyperscaler capex is tracking toward nearly $800 billion in 2026 and $1.3 trillion in 2027. Those forecasts assume AGI is a real destination that justifies the spend, and the dollars flow well beyond the chipmakers (we profiled seven of the power, cooling, and networking suppliers riding that same buildout in a free report).
What Eyee Actually Conceded About Astra #
Eyee did not dismiss the technology. He told CNBC that “Astra blew everyone out of the water when it came to model releases last week” and outpaced peers in coding and reasoning.
He singled out endurance. Eyee said, “Astra is the first model that’s able to do things for hours and hours and days on end in unfamiliar environments, which is unlocking new scientific research.”
That is a narrower claim than the headlines suggested. Sustained accuracy on long-horizon tasks is a real breakthrough for lab automation and coding agents, but it is not the general cognition that AGI implies to a lay reader.
For NVIDIA, the distinction matters because agentic workloads consume far more inference compute than one-shot chat. Huang himself said, “When the world goes to agentic, fully agentic systems, you’re going to have agents running all the time, working with other agents running all the time.”
Definitional Sleight of Hand #
OpenAI’s working definition of AGI measures the ability to replace economically valuable human work. That framing collapses human cognition into a labor-substitution metric, which suits a company selling automation.
Eyee’s critique lands because most of human intelligence, including care work, judgment under ambiguity, and cultural knowledge, does not show up on that scoreboard. A benchmark that ignores those dimensions declares victory prematurely.
Huang has stopped defending any specific milestone. On the Q2 call, he said, “For many tasks, we could say that we’ve already achieved AGI. I think all of those milestones and all those, they’re kind of senseless at this point.”
His preferred metric is simpler: “AI has reached its inflection point. It’s doing useful work. Its tokens are productive and profitable. Now, compute is revenue.” That reframes AGI as a monetization event.
Three-Way Squeeze on AI Labs #
Eyee described the labs as caught between three masters. In his words, “Companies are struggling to balance the need between being first in market, appeasing shareholders and also this new role they’re playing of sort of the U.S.’s first line of defense as it comes to national security.”
That squeeze pushes labs to ship models faster than internal safety teams recommend. The CNBC segment noted that labs found containment-breaking behavior in models internally before deployment and made minimal changes before running them in production.
NVIDIA is now financially entangled with those same labs. The company has invested nearly $50 billion in frontier AI labs and organized financing platforms with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR to raise over $500 billion in third-party capital.
Huang addressed the obvious question directly: “We recognize the scale of this support, and we know some will call this circular financing. We see it differently.”
Accountability Gap for Agent Swarms #
Eyee’s sharpest point was legal. He asked, “Who is responsible? If I make an AI agent swarm and it goes and hacks your company who goes to jail, that’s a felony. We haven’t made those decisions collectively yet.”
The CNBC segment referenced a Reuters report that OpenAI bots allegedly hacked a German company and shared tips on avoiding detection. That is not a hypothetical liability question anymore.
For NVIDIA, the exposure is indirect but real. If regulators impose liability regimes that slow agentic deployment, the inference compute curve flattens and Huang’s monetization thesis weakens. Supply is also tighter than the demand picture suggests. Huang said NVIDIA has supply for approximately 70% of demand, and the company expects fiscal 2028 revenue growth of roughly 70% on a supply-constrained basis, per the Q2 FY27 8-K.
Where NVDA Trades Now #
NVIDIA shares trade at $225.73, up 21.18% year to date and 34.29% over the past year. Analysts carry an average target of $327.13 with 57 Buy ratings against just 1 Sell.
The Q2 beat was substantial, with revenue of $96.22 billion and non-GAAP EPS of 2.22, and Q3 guidance of $108 billion implies the ramp is intact. Gross margins are expected to compress to 74% in Q3 as memory pricing bites.
Eyee’s warning does not invalidate the numbers, but it argues for a smaller AGI premium in the valuation. The customer financing arrangements, memory inflation, and unresolved agent liability all deserve weight in any model.
The setup rewards patience: a forward multiple around 45x carries the concentration risk of a single supplier funding its own demand. If the AGI narrative cools before capex plans do, valuation support thins quickly.
Contact [email protected] for any questions or corrections.