Two men run companies that, together with their largest peers, are reportedly burning through more than a trillion dollars in AI capital expenditure over two years. One of them stood at Davos in January and named the exact condition under which all of that spending curdles into a bubble. The other, six months later, quietly told analysts he was doing the opposite of what would prevent it.
Satya Nadella's line was clean enough to survive translation into every finance newsletter on earth. "For this not to be a bubble by definition," he said in January 2026, "it requires that the benefits of this are much more evenly spread."
Diffusion or bust. That is the whole thesis, delivered by the man best positioned to know.
The people warning loudest about concentration are the ones structurally guaranteeing it. That contradiction is not hypocrisy. It is a signal.
Here is the uncomfortable part. The company Nadella runs is spending its way toward the opposite outcome, and so is every peer he shares the trillion-dollar capex bill with. The people warning loudest about concentration are the ones structurally guaranteeing it.
That contradiction is not hypocrisy. It is a signal. And once you see how the smartest capital allocators are actually behaving—as opposed to what they say at conferences—the bubble question stops being about valuations and becomes about physics.
Start with what "diffusion" has to mean for the numbers to work. Nadella is not being poetic. The market has priced AI on the assumption that its value will spread the way software always has: a model ships, an API opens, and within months millions of users in every industry are transformed by a network call.
That is the diffusion machine that built the cloud economy. Text moves at the speed of a network call, so value moved that fast too.
The problem is that software is less than 0.5% of global GDP. SaaS, all of it, comes in under $500 billion against a $100 trillion-plus world economy. AI has conquered that 0.5%. The valuations require it to conquer the other 99.5%—manufacturing, energy, logistics, construction, agriculture—on roughly the same timeline.
The 99.5% does not care about your API.
We already have the first data point on what happens when the diffusion assumption meets reality, and it comes from inside the 0.5% itself. Per the MTS Blog's account of a Goldman Sachs report, the IGV index of top US software stocks fell 17% in the first half of 2026 and 26% from its October 2025 peak. The top ten software companies reportedly shed nearly $800 billion in market cap in six months.
Not revenue. Expectations. The market repriced what software is worth in a world where AI agents do the work, and pricing power migrates from per-seat to per-outcome. Goldman's Brian Kane calls software "the canary in the coal mine for AI economics"—dying first to warn everyone else what the gas smells like.
Hold that number against the one that should terrify anyone who bought the diffusion story. Goldman's own banking report—not its research analysts, its bankers, the people who write for CEOs signing checks—cites US AI investment at roughly 1.2% of today's annual GDP.
That is well below the railroad buildouts of the 1800s, which ran 3% to 4.5% of GDP. So, on the article's own reading, the physical transition that valuations demand appears, per the Goldman report, to be financed at less than half the intensity of a technology whose diffusion took decades and required laying literal steel across a continent—though the same report notes that AI investment will rise further if current projections are realized.
Eight hundred billion dollars already evaporated in software, the easy 0.5%, on the mere suspicion that the old pricing model is dead. The hard 99.5% is being funded at a fraction of historical buildout intensity. Something in that math does not close.
The Goldman banking team spelled out why in a thirty-page report called Harnessing AI for the Real Economy, and their framework is the load-bearing evidence here. They map five gates between AI's software success and its deployment into physical industries. None of them is a technology problem.
Gate one is physics. AI optimizes beautifully in simulation, then meets materials with real tolerances, heat that behaves differently than modeled, dust and vibration and humidity. A perfect design fails the moment it touches a factory floor.
Gate two is private data. ChatGPT can write your emails but cannot optimize your production line, because the data that would train it lives inside factories and quality-control logs, not in Common Crawl.
Gate three is edge deployment. A production line cannot pause because a cloud region hiccups, so models have to run locally on hardware never designed for inference.
Gate four is certifiability, and it is the wall. In software a hallucination wastes tokens. In aviation or energy or automotive, a hallucination kills people. Traditional certification demands you prove why a system made a decision under all conditions—and a black box, by definition, cannot supply that proof. We do not yet have frameworks for certifying black-box systems in safety-critical environments.
Then there is the constraint no amount of capital dissolves: people. Per the Goldman report cited by the MTS Blog, the US power sector alone needs 500,000 new workers by 2030. Training an electrician takes three to four years. Per that same report, in 2024 the entire US energy industry graduated 45,000 apprentices against a need of 65,000 per year—before counting retirements.
Capital accelerates. Chips iterate. Training a human being who can wire a substation has no shortcut. The irony Goldman's bankers land on is sharp: Wall Street's real labor anxiety is not white-collar displacement, it is a shortage of electricians, welders, and pipefitters.
These five gates are not friction that clears with a better model next quarter. They are the reason the diffusion valuations require is physically blocked, not merely delayed. You cannot patch physics, you cannot download proprietary factory data, and you cannot certify a black box into an FAA-grade process by writing a bigger check.
Which brings us back to the contradiction we opened with, and to the tell that matters more than anything said at Davos.
Watch where the smartest allocator of AI compute on the planet is actually pointing its chips. On Alphabet's second-quarter earnings call, a Goldman analyst asked Sundar Pichai how Google balances selling TPUs to customers against keeping them.
Pichai did not hedge. "In terms of allocating our TPUs," he said, "our first priority is making sure we are allocating what we need to compete at the frontier in terms of AGI development." Pressed again, he repeated it: "the baseline with which we start is what it takes to continue AGI development at the frontier."
Read that against the backdrop. Google Cloud grew 82% year over year to $24.75 billion in the quarter, with a $514 billion backlog of customers who have signed up for services they haven't yet consumed. Demand for the diffusion machine is nearly infinite, and Google is choosing to feed itself first.
This is the signal hiding in plain sight. Google owns the rental engine that broad diffusion would run on. If the company genuinely believed the near-term future was AI spreading into the 99.5% at software speed, renting TPUs into that build-out would be the most lucrative deployment of compute imaginable—a railroad charging every train on the continent.
Instead, per Pichai's own account of allocation priorities, Google appears to be reserving a significant portion of those chips to chase AGI internally. That is not the behavior of a company betting on imminent broad diffusion. It is the behavior of a company that believes the real returns for the next several years live at the frontier, inside its own walls, not distributed across the physical economy that its customers keep begging to serve.
The same institution that produced the five-gates report—Goldman—is the one whose analyst extracted that admission from Pichai. The bank's bankers mapped why diffusion is blocked, and the bank's analysts, arguably, documented the most sophisticated player in the industry acting as if it knows.
So the picture resolves into something colder than the usual bubble debate. Nadella is right that without broad, even diffusion this is a bubble by his own definition. Goldman's bankers are right that five physical gates make that diffusion, at software speed, impossible.
And Pichai is quietly confirming both by reserving his most valuable asset for the one place where returns are real—the frontier—rather than renting it to the economy that would need it for diffusion to occur.
The investors who priced AI for broad, near-term diffusion should study that allocation decision more carefully than any earnings beat. The people with the best information and the most compute are not betting on the outcome the valuations require.
They are building for a narrower, more concentrated future—the exact future Nadella warned would define a bubble. The tell is not what they say. It is where the chips go.