The AI selloff is really a pricing-power test for the whole buildout
This week’s AI wobble looks less like the end of the trade than a new test of who can still charge through the cycle. As hyperscalers scrutinize returns, the stack’s next winners may be the companies with bottleneck power in memory, networking, and tools — not just the highest-beta GPU names.
The market is finally asking the right AI question, and it is not whether demand exists. It is who gets paid as the buildout gets more expensive and buyers get tougher. That is why the June 23-26 selloff reads less like a verdict on AI and more like a pricing-power test across the stack: if hyperscalers are going to keep spending at this pace, suppliers have to prove they can defend margins, not just ship volume. The names that matter from here are the ones sitting on bottlenecks.
Start with the simplest tell: the market is no longer rewarding “AI exposure” as a category. Broadcom put up the kind of numbers that should have ended the bubble talk on the spot — revenue up 48% year over year to $22.2 billion, Q3 guided to $29.4 billion, up 84%, with non-GAAP operating margin steady at 67% — and the stock still sold off because investors had already moved on to the next question. If even a supplier with that margin profile cannot simply coast on AI enthusiasm, then the debate has clearly shifted from demand to monetization.
That shift matters because the buildout is broadening beyond the obvious GPU winners. NVIDIA still looks dominant on the raw numbers, with 65.5% revenue growth and a 63.0% net margin, and its valuation at 32.99x earnings is hardly irrational against that backdrop. But the market is starting to treat GPU leadership as the entry ticket, not the whole thesis. The harder call now is which adjacent layers can force customers to absorb higher costs without blowing up returns on capital.
Memory is the cleanest example. Micron is trading at 23.36x earnings with 48.9% revenue growth and a 55.9% net margin, and that combination only makes sense if investors believe memory has become a choke point rather than a commodity sidecar. The recent rally after its forecast was not just momentum chasing; it was the market recognizing that constrained supply and rising content per AI system give memory vendors unusual leverage. When estimates suggest memory could rise to roughly 30% of hyperscaler AI capex in 2026 from about 8% in 2023-24, that is not a side story. That is a warning that one of the most important margin debates in AI may sit upstream of the GPU.
Networking deserves the same treatment. Arista at 49.90x earnings is not cheap, but 28.6% revenue growth and a 38.3% net margin tell you why investors are willing to pay for it. In a world where AI clusters have to move more data, faster, with less tolerance for downtime, networking is not just connective tissue. It is part of the bottleneck. If hyperscalers start pressing for better returns, they may squeeze plenty of vendors, but they are less likely to skimp on the layer that keeps expensive compute fully utilized.
The equipment names reinforce the same point from another angle. If the AI capex cycle were actually cracking, wafer fab equipment would be one of the first places to show it. Instead, the spending outlook still points higher, with bull-case wafer fab equipment demand projected at $145 billion in 2026, $200 billion in 2027, and $250 billion in 2028. That helps explain why Lam Research has rallied 107.2% year to date and Applied Materials 141.5%, even though their current growth rates are far less explosive than the market’s favorite AI stories. Investors are underwriting the idea that the buildout continues — they are just becoming more selective about who captures the economics.
NVDA: 32.99x P/E, 65.5% revenue growth, 63.0% net margin
MU: 23.36x P/E, 48.9% revenue growth, 55.9% net margin
ANET: 49.90x P/E, 28.6% revenue growth, 38.3% net margin
Yes, the bears have a real point: some of this is still a crowded-trade unwind. Broadcom’s post-earnings drop showed how quickly the market can punish even strong results when expectations are stretched, and the whole AI complex still tends to trade as one factor on bad days. But that explanation is incomplete. If this were only a multiple reset, we would not be seeing such a sharp distinction between companies exposed to financing pressure and those with clearer bottleneck leverage.
That is where Oracle becomes the useful contrast. Oracle trades at just 19.46x earnings with 17.4% revenue growth, but the stock is down 25.3% year to date because the market is focused on the cost of funding the buildout, not the slogan attached to it. Public reporting around its June selloff centered on capex running to about $55.66 billion in 2026, above a $50 billion target, while free cash flow remained negative. That is the other side of the AI trade: somebody has to pay for all this infrastructure, and if the buyers cannot show credible returns, they will push harder on every supplier that lacks real pricing power.
So the right frame for this selloff is not “AI is over.” It is “prove you can hold the line.” The companies with the strongest hand are the ones selling scarce inputs, mission-critical throughput, or tools required to keep the capacity race going. The rest of the stack may still grow, but growth alone is no longer enough once hyperscalers start asking what the returns look like after the invoices arrive.
That is why we would treat this week’s wobble as a sorting mechanism, not a collapse. The market is moving from first-order excitement about AI demand to second-order scrutiny of who can defend margins as spending scales. In that regime, NVIDIA still matters enormously, but so do Micron, Arista, and the equipment suppliers that sit closer to the bottlenecks than to the hype.
What would change our mind? A real break in infrastructure spending would do it, especially if tool orders softened and memory pricing lost momentum at the same time. Short of that, the better read is that AI is entering a tougher, healthier phase — one where pricing power, not just narrative beta, decides who wins. On that basis, the TickerSpark Score matters most when it confirms durable margin leverage rather than just headline AI exposure.
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.
