The AI trade isn’t breaking — it’s punishing the wrong kind of exposure
This week’s selloff looks less like the end of the AI trade than the end of lazy AI positioning. The market is getting more selective, favoring the parts of the stack with hard bottlenecks, pricing power, and visible infrastructure demand while punishing anything that relied on AI halo alone.
The cleanest read on this week’s AI selloff is not that the thesis broke. It’s that the market finally started separating core AI economics from borrowed AI excitement. Broadcom’s post-earnings drop turned a routine reset into a bubble debate, but the underlying numbers still point to a capex cycle that is alive, funded, and increasingly concentrated around the infrastructure layer. When a company can post $22.19 billion in quarterly revenue, up 48% year over year, reiterate a $100 billion AI chip sales target for 2027, and still get sold, that is not demand collapse — it is the market refusing to pay up for anything short of unmistakable operating leverage.
That distinction matters because the AI trade was always too broad. For a year, investors could buy almost any stock with data-center, networking, custom silicon, cooling, or software exposure and call it an AI position. This pullback is narrowing that field. The market is moving away from companies whose AI upside is real but diffuse, and toward the names sitting on actual choke points: accelerated compute, power, thermal management, and the networking fabric that keeps the whole buildout running.
NVIDIA remains the clearest proof that the core thesis is intact. Its latest quarter delivered $44.1 billion in revenue, up 69% year over year, with data center revenue of $39.1 billion, up 73%. Those are not the numbers of a market that has suddenly discovered AI demand was fictional. They are the numbers of a platform vendor still benefiting from scarcity, scale, and pricing power. The valuation snapshot reinforces the point: NVDA trades at 31.47x earnings with 65.5% revenue growth and a 63.0% net margin. That is expensive in absolute terms, but not obviously irrational relative to the growth and profitability it is producing right now.
Broadcom is where the debate gets noisy, but the selloff actually supports the sharper version of the bull case. AVGO trades at 63.11x earnings and 23.98x sales, so the bar was already high. Its quarter was strong; the stock still fell because investors wanted cleaner evidence that custom silicon and AI networking can keep accelerating from here. Bears will say that is the first crack in AI monetization. We think that overstates it. A stock dropping after a 48% revenue-growth quarter says more about expectations than about end demand, especially when hyperscaler spending plans and financing around AI infrastructure still point to a buildout phase rather than a digestion phase.
The better way to frame the move is by asking which businesses have the most direct pull-through from that capex cycle.
NVDA: 31.47x P/E, 65.5% revenue growth, 63.0% net margin
AVGO: 63.11x P/E, 23.9% revenue growth, 38.8% net margin
That list shows why the old “AI bubble” label is too blunt. Vertiv and Eaton are not being rewarded because they can say “AI” on a conference call; they are being rewarded because data centers need power distribution, cooling, and electrical gear whether the application layer monetizes next quarter or two years from now. Vertiv just beat expectations with $1.17 in quarterly EPS versus $1.00 expected and raised full-year guidance to $6.30-$6.40. Eaton posted 21% growth in Electrical Global and explicitly tied demand to data centers, while its Boyd Thermal acquisition is expected to add at least $1.7 billion of 2026 revenue. That is what durable AI exposure looks like: not concept, but order flow.
This is also why the June semiconductor reversal should not be read as a verdict on the whole stack. Yes, the Nasdaq fell 4.3% over two sessions and the semiconductor ETF dropped 12.3%, which is exactly the kind of air pocket that feeds bubble talk. And yes, Ray Dalio’s comparison to 1929 and 2000 gives the bears a serious macro frame: if capex outruns monetization for too long, multiples can compress hard. But that comparison cuts both ways. In every major buildout cycle, the market eventually stops rewarding generic exposure and starts rewarding the owners of the bottlenecks. That is not the death of the theme. It is the maturation of it.
The valuation spread across the group makes that point more clearly than the headlines do. Marvell sits at 95.61x earnings and 27.77x sales after a 211.1% YTD run, which leaves very little room for execution wobble even with 42.1% revenue growth. Arista at 56.32x earnings and 21.23x sales still has a strong case because networking remains a real AI constraint, but it no longer gets the benefit of indiscriminate multiple expansion. Meanwhile Microsoft, despite a -18.8% YTD move, still matters because the capex engine is not theoretical: its reported 2026 capex plan is $190 billion, and its AI run-rate has been cited at $37 billion. The market may be less patient with software-adjacent narratives, but it is not walking away from the spending backbone that supports the whole ecosystem.
That is the key reframing. The AI trade is no longer “buy anything exposed to AI demand.” It is “own the layers where demand is hardest to substitute and easiest to monetize.” Chips with dominant economics, networking that removes throughput constraints, and electrical and thermal infrastructure with visible backlog all fit that test. Names that depend on looser, longer-dated AI monetization stories do not. Broadcom’s selloff did not invalidate the theme; it reminded investors that in a more mature phase of the trade, good businesses can still be bad stocks if the market has already priced perfection.
What we would watch from here is not whether AI spending slows in a headline sense, but whether the bottlenecks start to ease. If hyperscaler capex plans hold, if infrastructure suppliers keep raising guidance, and if the leading chip vendors maintain margins that signal pricing power, then this pullback will look like a rotation inside the AI trade rather than the end of it.
What would change our mind is straightforward: a real break in infrastructure demand, not just a valuation reset. If data-center power, cooling, networking, and accelerated compute all start missing at the same time, then the bear case gets stronger fast. Until then, the market is not rejecting AI. It is punishing the wrong kind of exposure and paying up for the parts of the stack that still have the clearest claim on the dollars.
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.
