The Sentence That Lost Two Words
An Nvidia VP said something true about his own team. Cut two words out of it, and it became the rumor stalling good companies across the country. The missing words are the whole story.
A single sentence is freezing good companies in place right now, and it is built on a lie of omission.
It shows up in board meetings, break rooms, and comment sections, usually like this: AI costs more than employees. Sometimes it arrives with a name attached, an Nvidia VP, which gives it the weight of a confession from inside the temple. If the people building the thing admit it costs more than people, the logic goes, then the dream is dead and the smart move is to wait.
I want to show you that sentence before it went through the telephone game. An Nvidia VP did say something close to it. Then it lost two words.
What Bryan Catanzaro, Nvidia's VP of applied deep learning, actually told Axios was that, for his own team, "the cost of compute is far beyond the cost of the employees."
Read that again and notice the part the rumor amputates: for his own team. His team trains frontier AI models. That is the most compute-hungry work being done anywhere on earth. Of course his compute bill dwarfs his payroll. It is like the head of a smelting operation saying his furnace costs more than his crew, and a reporter running the headline "Heat Is Too Expensive."
Cut two words, "my team," and a narrow truth about a research lab becomes a universal lie about your shop floor. That is not a quote anymore. That is a rumor wearing a quote's clothes.
Once you see the cut, you see it everywhere. And the bubble story is spreading fast.
You have heard that three-quarters of companies are abandoning AI. The real number, from a survey of twenty-five hundred senior leaders, is that three-quarters rolled back one kind of thing, a customer-facing agent, usually because nobody governed it. That same survey found three in five already running agents in production and nearly nine in ten planning to within a year. "Abandoning AI" and "killing one ungoverned bot while expanding everything else" are not the same sentence. One of them lost a lot of words.
You have heard that Starbucks proves AI does not work. Starbucks retired one inventory tool after nine months because it kept miscounting the milk. They kept the order-sequencing AI, the barista assistant, and the rest of the program. A Wharton operations professor's read on the whole episode was that there is "more hype than actual benefit" at the moment, because companies feel pressure to run AI before it is ready to deliver. That is not a verdict on AI. That is a verdict on deploying before you architect.
So look at who is actually bleeding, because that is where the real story is hiding.
Uber burned its entire annual AI budget in four months. One company ran up $500 million in a single month because nobody set a usage limit. Microsoft canceled most of its internal coding-AI licenses six months after rolling them out. These are not stories about AI being too expensive. They are stories about the same mistake made by everyone at once: they bought the tool and skipped the architecture. Their AI demand outran their architectural capacity overnight. Scaling Trapped, inside a single budget cycle.
The mechanism is always the same. You hand the whole team a top-shelf model, point it at everything, and set no rule for what runs where. So the routine work that a cheap, private model should handle for pennies runs on the most expensive model money can buy, all day long. You are paying premium freight to haul gravel. And you cannot see it, because there is no routing, no meter, and nobody governing the spend. It shows up as a number that climbs every month, and you keep paying it, because what are you going to do, turn AI off?
That number has a name. It is the invoice for deploying before you architect. Every operation that skipped the architecture is paying it right now, whether they have found it on the statement yet or not. You are stealing from yourself, and you handed yourself the card.
Here is the part the doom videos leave out, and it should stop you cold.
The same analysts writing the scary headlines wrote the fix in the very next paragraph. One prescription, repeated everywhere, lands embarrassingly close to a doctrine I have taught for years. Match the model to the task instead of aiming the most expensive engine at everything. Run the routine work cheap. Reserve the costly model for the moment it earns its keep. Govern it. Meter it. That is not a cost-cutting tip. That is architecture. The companies getting eaten alive are the ones who do not have one. The companies quietly compounding are the ones who do. The cost crisis is not evidence that AI failed. It is evidence that ungoverned AI fails, and that most companies are still ungoverned.
Which brings us to the word everyone is reaching for. Bubble.
Maybe it is one. I have watched four technology transitions across 42 years, and I lived the last great bubble up close. There was a stretch in the late nineties when you could not put "dot-com" on yourself without it sticking like something you had stepped in. Then the bottom fell out, and the smart money lined up to call the internet a fad.
The internet was not a fad. The bubble did not kill it. The bubble culled it. The pretenders went under, the capacity got cheaper, and the companies that had actually built something walked through the wreckage and owned what came next. That is what a cull does. It rewards the architected and buries everyone else.
So if AI is in a bubble, here is the only question that matters for your business: when it pops, which side of the cull are you on? The company that bought a pile of tools loses everything when those vendors disappear, and starts over from zero, which is the Catch-Up Penalty arriving all at once. The company that built an architecture barely flinches, because the architecture does not care which vendor survives. It points at whoever is left standing and keeps running. The tool is disposable. The architecture is what survives. That is the difference between an Unassailable Moat and a subscription.
There are two operations sitting exactly where you are right now. One has an architecture that decides what runs where, on what model, under whose watch. It pays a fraction of what its competitors pay, it compounds while they thrash, and when the cull comes it is still standing. The other bought tools, called it a transition, and is bleeding out on a meter it cannot read. From the outside, on a good quarter, they can look identical. They are not. And the most expensive part is this: right now, you do not know which one you are.
That is the gap. Not whether AI is too expensive. Whether you can see, in dollars, what your current setup is costing you, and whether what you have built survives the thing everyone is bracing for.
That is the first thing I find. The Strategic AI Intelligence Diagnostic™ puts a number on it: what you are spending, what it is actually buying, and what an architected operation in your position would pay instead. In dollars, on your operation. No tools pitched. Reach me directly at [email protected].
And if you want to look at the capital question before the cost question, my Forfeit Brief lays out the federal money already allocated to a business like yours and what you forfeit by leaving it unclaimed. It is free at carlpeterlin.com, access word Forfeit.
Everybody is selling AI tools. That's tactics. I work with you on the strategy and the architecture that runs it, so AI becomes your competitive advantage instead of a line item you are afraid to read.
Carl J. Peterlin Jr. is a Strategic AI Intelligence Architect™. Across 42 years he has watched four technology transitions and documented the single pattern that decides which companies survive them. He is the author of Your SMB AI Revenue Ratchet and Death To Excel!.
Edition 7 of The Fourth Intelligence, first published June 23, 2026. The argument stands as written. Offers, pricing, and product names referenced in the original have since advanced; current ones are at carlpeterlin.com.