The AI Slowdown Panic Is a Valuation Squeeze — the Real Test Comes Next Quarter

When Dario Amodei, the chief executive of Anthropic, published an essay this weekend urging AI labs to "pace the frontier" — slow the breakneck progress of new models so safety rules can catch up — he was not aiming at the power grid. But on Monday the market connected the dots anyway. GE Vernova, the turbine-and-grid company that has become the cleanest proxy for AI's electrical appetite, fell nearly 9% by late morning. Eaton was down more than 7%, Comfort Systems had lost 6.5%, Constellation Energy 6.1%. The entire complex of companies that build, cool, and power data centers sold off in sympathy.

This looks like a chain breaking, and the chain is real. But it is worth being precise about which link moved today, because a selloff that frightened power stocks is not the same thing as a selloff that will hurt their revenue. The first domino — the call to slow AI — is public. The next link is the part still being mispriced: whether and when that rhetoric turns into cancelled orders. Right now it has not.

Landing one: a multiple squeeze, not an order problem

The hardest-hit stock points to the cause. GE Vernova fell most in part because it was the most expensive — market data put it at roughly 150 times forward earnings and near 100 times trailing EBITDA, a valuation that only makes sense if data-center power demand grows at its fastest plausible rate for years. Comfort Systems had nearly doubled over the past year, and Quanta Services was up more than 60% on the same basis. When an expensive, crowded AI-infrastructure name meets a growth scare, the multiple is the first thing to compress. That is the first landing: a repricing of how fast the demand curve rises, executed through the price.

It is also, by the distinction that matters here, a shared shock rather than contagion. Nothing Constellation did changed Constellation's economics on Monday, and nothing GE Vernova did changed GE Vernova's. They all fell because they share one demand driver — hyperscaler capital spending — and the market marked down how much growth that driver will fund. Broad sector redness of this kind is not proof of transmission between nodes; it is the same macro factor repricing an entire group.

Landing two: the step that would make it real

The reason a growth scare does not automatically become an earnings problem is that the physical buildout is already booked. GE Vernova ended its June quarter with roughly $176 billion in backlog — nearly five years of current revenue — including $5 billion of data-center orders booked in just the first half of the year, more than double all of last year's related revenue. That backlog is contracted and spread across a multi-year delivery schedule. It does not vanish because a chief executive in San Francisco asks for restraint.

The step that would turn rhetoric into reality is not today's headlines. It is a hyperscaler actually cutting its capital-expenditure guidance at the next earnings report. That is the observable tripwire. Until Microsoft, Amazon, Meta, or Alphabet says it will build fewer or smaller campuses, the turbines, switchgear, transformers, and construction crews keep getting scheduled. The panic becomes a real earnings event only when that budget decision lands. That is the second landing, and it runs on the quarterly earnings clock, not the news clock.

The amplifier and the control test

Power and grid investors have been here before. On January 27, 2025, China's DeepSeek released a low-cost reasoning model, and the market drew the same conclusion from the same evidence: maybe AI does not need all this electricity. Vistra closed nearly 30% lower in a single session, and Constellation, Talen, and GE Vernova all fell more than 20%. Bank of America's analysts pushed back at the time, arguing that U.S. and European grids remain under-invested and are the true bottleneck on load growth. The selloff did not stick: today GE Vernova trades up roughly a third for the year.

Amodei's essay is the same movie on a later date, and the amplifier has grown. After a year-long run, valuations are higher and more crowded, so a growth scare has more room to compress. The financing has also turned: with interest rates expected to rise, the leveraged operators in the middle of the AI buildout are paying more to carry the debt that funds construction. The safety rhetoric does not have to be enacted for that to bite; it only has to slow the expected revenue those loans were priced against.

Landing three: the leverage, not the turbine

Here is the mispriced next link. If a slowdown ever becomes real — if hyperscaler budgets and power commitments actually shrink — the first casualties are unlikely to be GE Vernova or Quanta, which hold contracted backlog and get paid as work is delivered. The exposed node is the highly leveraged data-center operators and the banks and lenders that finance them. One analyst put it directly: if the AI race slows, the leases, debt, and power commitments remain on the books regardless of slower compute demand. That is a credit event, not a revenue event. It shows up on refinancing schedules and loan covenants — quarters after any headline, if it shows up at all.

For a retail investor, the exposure ladder is straightforward. The direct pain is concentrated in high-valuation, high-AI-weight names like GE Vernova and the grid and construction contractors — but those carry the firewalls of contracted backlog and multi-year lead times. The indirect exposure is broader: index and fund concentration in the biggest AI-power names and, one level down, the lenders who finance the buildout at higher rates. That is where the stress that starts in one magazine essay would, if it ever became real, find its way toward an ordinary portfolio.

What stops the chain

The chain continues only if two things happen in sequence: a hyperscaler cuts capital-expenditure guidance, and that reduction shows up in the next round of GE Vernova order bookings and construction awards. It stops if the fall earnings season reaffirms budgets and orders keep being booked — in which case today is a valuation reset, not a changing of the thesis. The walk-down is shallow from the power side because the work is already contracted, and shallow from the demand side because grids, not chip efficiency, are what actually gate the buildout.

So monitor, not the headlines that moved Monday, but the next marginal order: capital-expenditure guidance and the order books it feeds. If those hold, the exposed stocks have simply become cheaper — and a genuinely cheaper share of a contracted, multi-year buildout is a different thing from a broken one. If they are cut, the second domino has moved, and the conversation moves to the credit that finances construction. This link is real. It is just not breaking on the news cycle the way Monday's tape suggested.