The AI trade is shifting from chips to power, and the market is still underpricing that handoff
The AI trade is no longer just a semiconductor story; it is becoming a power, cooling, and data-center capacity story. That matters now because hyperscaler capex is still accelerating, while the real bottlenecks are moving downstream to the companies that can actually energize and house AI workloads.
The market is still talking about AI as if the scarce asset is the next GPU. We think that framing is already stale. The more important scarcity is moving into electricity, thermal management, switchgear, and financed data-center capacity, because hyperscalers can keep ordering accelerators only if someone can power and cool them at scale. Once the bottleneck moves, pricing power tends to move with it — and many infrastructure names are still being treated like accessories to
The cleanest evidence is where the money is going. Consensus discussion around AI used to center on chip lead times; now it is increasingly about whether enough power, land, and built capacity can be brought online. That shift is showing up in the capex stack itself: one major Wall Street forecast now sees the four largest hyperscalers spending a combined $5.3 trillion across fiscal 2025 through 2030, with private infrastructure and real estate capital expected to play a larger financing role. That is not a semiconductor-only buildout. It is a grid, equipment, and balance-sheet buildout.
The stocks most exposed to that handoff are not all priced as if they control a bottleneck. VRT is the clearest example. Its reported first-quarter 2026 net sales rose 30% year over year to $2.65 billion, while adjusted operating margin expanded 430 basis points to 20.8%. That is what scarcity looks like when demand stops being theoretical and starts hitting physical systems. ETN is seeing the same thing from a different angle: its Electrical Americas data-center orders were up roughly 240% in the quarter, and revenue in that end market was up about 100%. Those are not sidecar metrics riding on someone else’s product cycle. They suggest the electrical layer is becoming a primary beneficiary of AI capex, not a derivative one.
The market still tends to benchmark everything back to NVDA, and that is fair up to a point. Nvidia remains the profit center of the AI stack, with a 63.0% net margin, 65.5% revenue growth, and a $4.76 trillion market cap that reflects extraordinary execution. But that is exactly why the next leg of the trade looks different. When the incumbent chip winner is already operating at that scale, the incremental debate naturally shifts from who has the best accelerator to who can remove the next constraint on deployment. If the gating factor for new AI capacity is no longer silicon availability but megawatts and cooling density, then the market should stop valuing infrastructure enablers as mere attachments to the chip cycle.
A few comparative numbers make the handoff visible:
That spread matters. VRT has already rerated, so this is not an argument that every power-and-cooling name is undiscovered. Some of the move is obvious and deserved. But the broader market still seems more comfortable paying up for visible AI compute than for the assets that make compute usable. ETN at 5.68x sales and CEG at 2.93x sales do not screen like classic scarcity monopolies, even though the debate around AI buildout increasingly revolves around power access. The same goes for VST, which sits at 3.21x sales and is down 6.8% year to date despite being tied to the exact question investors keep asking: where does the electricity come from?
That is also why ORCL is such a useful bridge case. Oracle’s revenue growth is real — 17.4% on the comparative data, with reported fiscal fourth-quarter revenue up 21% — yet the narrative around the stock has shifted toward AI infrastructure spending and backlog conversion. In other words, demand is not the issue. Conversion capacity is. The market is starting to understand that booking AI demand is one thing; turning it into delivered compute requires power, buildings, and financing. That is the handoff in plain sight.
Yes, the bulls on NVDA have a strong counter: every AI data center still starts with accelerators, and Nvidia still captures the fattest economics in the stack. That is true. And yes, skeptics can point out that VRT at nearly 70x earnings is hardly neglected. But both objections miss the same point. This is not a call that chips stop mattering; it is a call that the marginal bottleneck has moved. Once hyperscaler capex starts pressing against power availability and build timelines, the companies that can solve those constraints gain strategic leverage whether or not they ever approach Nvidia-like margins.
The financing angle strengthens the case. AI infrastructure is becoming capital-structure heavy in a way the market still underappreciates. A recent 15-year, $5.2 billion hyperscaler lease in the data-center space and the broader expectation that hyperscaler capex could approach the full run rate of operating cash flow in 2026 both point to the same conclusion: this buildout now depends on long-duration funding, not just product demand. Even SMCI, which looks optically cheap at 11.87x earnings and 0.52x sales, underscores the point from the server layer. Cheap multiples do not mean low risk when the business model is becoming more capital intensive just to secure components and fulfill orders.
The market’s old AI map put semiconductors at the center and everything else in orbit around them. The new map looks more like a chain of physical constraints, and the next one is power. That does not make NVDA irrelevant; it makes the surrounding infrastructure more important than the market has been willing to admit.
What we would watch from here is simple: data-center power awards, utility contracting, electrical equipment order growth, and whether cooling and capacity vendors keep expanding margins as demand scales. If hyperscaler capex cools sharply, or if power availability stops being the gating issue, this thesis weakens. Until then, we think the AI trade is being misframed: the handoff from chips to infrastructure is already underway, and the market is still pricing too many of the bottleneck owners like supporting actors.
Our take, not advice. This is opinion commentary — informational only, not personalized investment recommendations. Markets carry risk. Do your own research and consider your own situation before any trade.
▌The Daily Briefing · Free
A new stock idea, every evening.
One stock worth watching each weekday, plus the analysis behind it. Free, in your inbox.
