The market is no longer trading AI as one undifferentiated bubble. The bigger risk now is investors paying infrastructure-winner multiples for second-tier beneficiaries that still lack the margins, pricing power, or earnings conversion of the leaders.
That distinction matters because the market’s concentration risk is real, but it is being misunderstood. Public market data now shows technology at more than 39% of the S&P 500’s market cap, above the 2000 bubble peak. The easy conclusion is that all of AI is one giant excess. The harder, and more useful, conclusion is that capital is clustering around a narrow set of businesses with genuine scarcity economics while the rest of the ecosystem is being priced as if adjacency alone guarantees the same outcome.
Start with NVDA. If this were simply a replay of late-1990s speculation, the valuation would be doing all the work. But NVDA is trading at 33.15x earnings while still posting 65.5% revenue growth and a 63.0% net margin. That is not a normal momentum profile; it is what dominant pricing power looks like when demand is still outrunning supply. Nvidia’s latest capacity comments only reinforce that the buildout remains alive. So the debate should shift from whether Nvidia is crowded to whether other AI names deserve to trade as if they have Nvidia-like economics when they plainly do not.
The gap is easiest to see in the names investors keep treating as interchangeable AI beneficiaries.
That is the real fault line. AMD has had a huge run, up 139.8% YTD, and bulls have a credible case: data center momentum is real, and earnings growth has accelerated sharply. But 179.34x earnings with a 13.4% net margin is not the same setup as NVDA at 33.15x with 63.0% margins. The market is charging investors a premium for participation, not just proven monetization. That does not mean AMD cannot keep executing; it means the valuation risk has migrated away from the obvious leader and toward the names still trying to prove they can convert AI enthusiasm into sustained profitability.
The same logic applies to networking and custom silicon. ANET is a very good company, and its recent results back that up: strong growth, high operating discipline, and raised guidance. But this is exactly where investors can get sloppy. Arista still sits closer to enterprise and cloud capex timing than to the core compute bottleneck itself, yet it trades at 59.40x earnings and 22.29x sales. That may hold if spending stays broad and fast. If customers become even slightly more selective, the market will care a lot more about where the bottlenecks are structural and where they are merely cyclical.
If anything, MU is the better example of where the economics are improving for real. Memory has historically been the kind of business investors should hesitate to crown as a durable AI winner. But constrained HBM supply and rising revenue estimates are changing that equation, and the numbers now look more like scarcity economics than commodity drift. MU at 50.28x earnings with 48.9% revenue growth and a 41.5% net margin is expensive, yes, but at least the market can point to a tightening part of the stack where pricing power is visibly improving. That is a more defensible premium than simply buying anything with AI exposure in the slide deck.
AVGO complicates the bubble narrative in a useful way. At 94.42x earnings and 33.64x sales, it is hardly cheap, but the business is not being valued on hope alone. Consensus around its quarter centers on roughly $22 billion in revenue and about $10.7 billion in AI semiconductor revenue, with margin structure that remains far stronger than most of the field. That is why the old “next Cisco” framing misses the point. Cisco became the cautionary tale because investors extrapolated demand into businesses that lacked today’s margin durability and earnings conversion. Broadcom, like Nvidia, is expensive because it is already capturing the spend.
Yes, the bulls can still argue the whole capex cycle is early, and they are not wrong about demand. Nvidia says supply can support robust growth, Broadcom is still seeing AI acceleration, AMD’s data center business is expanding, and Micron is benefiting from HBM tightness. But that is precisely why the risk has moved. When the buildout is real, investors stop getting paid for owning the theme broadly and start getting paid for owning the businesses with the strongest claim on the profit pool. Everyone else has to justify a premium the old-fashioned way.
Our verdict: the AI bubble debate is now aimed at the wrong target. The market is already telling us that the issue is not whether AI spending vanishes, but whether every company attached to that spending deserves a leader’s multiple. It does not.
What we would watch from here is simple: margin durability, evidence of pricing power, and whether earnings growth keeps converting from AI demand into cash-generating scale. If second-tier beneficiaries start showing leader-level economics, this view weakens. If not, the next real air pocket in AI is more likely to come from the expensive followers than from the companies already proving they own the bottlenecks.
Palo Alto Networks, Inc. (PANW) drops after a sharp post-earnings reversal, even as the company raised full-year guidance and analysts lifted price targets. The move looks driven more by valuation and profit-taking than by weak fundamentals, with investors reacting to a rich multiple after a strong report.
IBM fell sharply after a strong run, even as analyst upgrades and upbeat quantum and AI news stayed in place. The move looks more like profit-taking and valuation pressure than a sign of operational trouble, with the company still showing solid earnings momentum and a credible long-term strategy.
Sprinklr is a slower-growth software name, but stronger profitability, robust free cash flow, and a cash-rich balance sheet make it a selective Buy. The key debate is whether improving renewals and AI traction can offset soft customer trends and margin pressure.