Seven checks. Seven zeros.
I run every piece of financial copy through a fixed diagnostic before I read it for content. Last week it was a Crypto Briefing item: Nvidia CEO Jensen Huang says AGI has arrived, and analysts name the stocks to watch.
Does it name the stocks? No. Are the analysts named? No โ "analysts," plural, unaffiliated, untraceable. A direct quotation? No. A date, a venue, a model version, a benchmark score, a single integer? No, no, no, no, no.
The piece is a title, a dek, and a byline. The body restates the title in longer words.
That is not a reporting failure. It is a production method. Take a keyword cluster with heat โ AGI, Nvidia, stocks โ assemble a headline that matches search intent, fill the remainder with an abstract that commits to nothing, and drop it in front of an audience already trained to treat a concept as a tradable instrument. A crypto news desk is precisely that audience.
AGI is not a technical term anymore. It is a container.
There is an operational definition โ a system that clears human baselines across a broad battery of standardized tests. Under that reading the claim has partial footing; frontier models have saturated several verifiable domains. There is the classical definition โ continual learning, sample efficiency, cross-domain transfer, autonomous goal-setting. Under that one the claim is simply false.
Huang has said both things. Within the same period he has described AGI as already achieved in a narrow sense and as roughly five years out in the general sense. Two incompatible statements, one speaker, no cost. That is the diagnostic: when a term can be both true and false without penalty, it has stopped describing a system. It is pricing one.
The outlet matters too. A crypto desk covering AI equities is not an editorial accident. It is audience migration โ when marginal returns on crypto narratives compressed, the same traffic moved toward AI equity exposure. A structural signal about attention. Nothing to do with technology.
Start with the definition, because the definition is the entire argument. The article never states which benchmark, which threshold, which evaluation protocol. Without that, "AGI has arrived" is unfalsifiable โ and an unfalsifiable claim in a market is not information. It is a lever.
Now the measurements. The capability profile of frontier models is jagged, not general. They exceed expert baselines in verifiable domains โ competition mathematics, code generation โ and sit well below human level in physical common sense, long-horizon planning, persistence of memory. ARC-AGI-2, released March 2025, kept the best systems in the single-digit percentage range against a human baseline near 60. That gap is not a tuning problem. It is the operational definition of "not general."
Sample efficiency is worse. A human acquires a new skill from single-digit to low-hundreds of examples. Current systems need 10ยณ to 10โต labeled or synthetic samples for an equivalent step. And continual learning โ online weight updates inside deployed commercial systems โ is still absent. Not delayed. Absent. Entropy finds its way through the gap, and that gap is where every AGI claim quietly relocates.
I audited the BAYC contract line by line in 2021 and found the fault somewhere else entirely: 15% of tokens carried corrupted metadata, and the corruption lived in off-chain indexing, not in the Solidity. The code was clean. The index was not. This article is the same architecture. The on-chain layer โ the headline โ is verifiable and true in its way: Huang said the words. The off-chain layer โ the body โ is where the tokens are wrong. There is no body.
Solidity does not lie, it only omits. Here the omission is the product.
Then there is the commercial function, which the article never touches and which is the only part that moves money. Nvidia sells compute. It does not sell AGI. Data center revenue is roughly 88% of the total and the customer base is a handful of hyperscalers. In a concentrated B2B structure, a supplier CEO setting public expectations is the cheapest instrument available for extending procurement commitments and investor patience.
"AGI has arrived" is a capex exemption mechanism. When a CTO asks a board for a billion-dollar compute budget, that sentence reclassifies the request from an efficiency project โ which requires ROI โ into a survival project, which does not. No benchmark score converts budget like that.
I mapped a version of this in 2025 for the spot Ethereum ETF custody filings: 90% of staked ETH under three entities, multi-sig key management with single points of failure, presented as decentralization. The structure had not changed. The label had. Institutional decentralization denial is not a political position. It is a reading of the key list. Nvidia's version is milder and identical in kind: rename the exposure, keep the concentration.
The analyst line โ "analysts name the stocks to watch" โ is the emptiest ring in the chain. The analyst is a converter: narrative in, ticker out. A converter with no name and no model has an output of zero. The likely names are inferable โ TSMC, HBM suppliers, transformer and switchgear makers, liquid cooling, neoclouds โ and the article supplies none of them. That is the proof of method. Research produces a list. A template produces a promise of a list.
Silence in the logs speaks louder than noise.
Here is what the bulls got right, and it survives every objection above.
The physical constraint migration is real and independent of any definition of AGI. The binding limit moved from GPU supply โ CoWoS packaging, HBM output โ to interconnection: transformer lead times of 18 to 24 months, grid capacity, site permitting, community approval. Value is relocating toward power equipment, cooling, and modular data centers whether or not the word "AGI" means anything.
The Jevons point stands as well. Efficiency gains do not reduce total compute demand in the near term; they expand the set of runnable tasks. Every argument that a smarter algorithm will shrink the compute bill has been wrong for three cycles.
And the bulls' strongest position is the one the article stumbled into by accident. The claim does not need to be true to be load-bearing. A narrative that cannot be falsified can still coordinate a thousand budgets. That is not a defense of the article. It is an acknowledgment that the emptiness is functional rather than stupid โ and that the function is the thing worth pricing.
The next verifiable sample is easy to specify. A named analyst, a named ticker, a target range, a valuation assumption. The first time those four appear together, the narrative becomes testable and the drift stops.
Until then, watch two clocks. One: continual learning landing in a production system โ the hard gate, for which no announcement substitutes. Two: the quarter when the word stops moving the price. We trace the fault line, not the earthquake. The earthquake is already priced in.