The AI trade is no longer about owning semis — it is about owning the bottlenecks
The AI trade is getting more selective, and Broadcom’s post-earnings drop made that impossible to ignore. Demand is still real, but the market is starting to reward the parts of the stack with the most durable pricing power: networking, memory bandwidth, and system integration.
The easy AI trade is over. Owning anything with semiconductor exposure was enough when the market was paying for narrative, but this week’s reset showed investors are now paying for scarcity inside the stack. Broadcom delivered big AI growth and still got hit, which tells us the debate is no longer boom versus bust. It is about which companies still control a bottleneck when expectations are punitive and capital intensity starts to matter.
Broadcom’s selloff was the tell. The company reported fiscal Q2 revenue of $22.19 billion, up 48% year over year, and AI semiconductor revenue of $10.8 billion, up 143%, yet the stock still dropped sharply after earnings. That is not a market rejecting AI demand; it is a market rejecting the idea that "AI exposure" alone deserves a premium. When a company can post that kind of growth and still disappoint, the message is clear: investors are no longer rewarding the whole complex equally.
That harsher sorting process matters because the old frame was too lazy. NVDA is still the center of the AI profit pool, and the numbers remain extraordinary: 65.5% revenue growth, 66.0% EPS growth, and a 63.0% net margin, with a TickerSpark Score that still reflects elite execution. But scale cuts both ways. At 31.52x earnings and 19.69x sales, the stock is not being priced like a speculative flyer; it is being priced like a dominant platform that now has to keep clearing an enormous bar. Bulls will say that is exactly why you still own the GPU leader. Fair enough. But once expectations get this high, the next leg of outperformance usually shifts toward the chokepoints that every AI deployment still has to pass through.
That is where the comparison gets more interesting. ANET is not the loudest AI name, but networking is where clusters either scale cleanly or break down. Arista is growing revenue 28.6% with a 38.3% net margin, and its underlying business lines up with the market’s new obsession: moving data fast enough between increasingly dense compute systems. MRVL sits even closer to the bottleneck argument. Its 42.1% revenue growth and 27.00x sales multiple look aggressive, but that premium is the market paying for exposure to custom silicon and interconnect where scarcity still has teeth. If the AI buildout keeps running into bandwidth and architecture constraints, those are not side bets on the trade. They are the trade.
The system layer matters for the same reason. DELL is not going to win the glamour contest against GPU names, but the company’s latest results showed AI demand is becoming a deployment story as much as a chip story. Dell booked $24.4 billion in AI orders and recognized $16.1 billion in AI server revenue, reinforcing that the bottleneck is moving into racks, integration, and getting hardware installed at scale. That helps explain why the market can punish a chip name for forward commentary while still keeping the broader AI infrastructure thesis alive. If enterprises and cloud customers are still spending heavily to assemble systems, then the value is broadening beyond the silicon die itself.
This is also why Broadcom’s reset should be read carefully rather than dramatically. Yes, some of the drop was positioning after a huge run, and yes, the broader semiconductor pullback over June 4-5 looked like a multiple reset as much as a fundamental one. But that is exactly the point. In a forgiving tape, everything AI rallies together. In a selective tape, investors start asking which layer has pricing power when customers push back, which layer can defend margins, and which layer is not one guidance line away from compression. AVGO still has real strengths, but at 62.50x earnings and 23.67x sales, the stock was priced for more than strong demand; it was priced for near-flawless messaging.
The weaker links are the names that look like AI beneficiaries but do not own enough scarcity. SMCI is the clearest example of why investors are becoming choosier. Its 46.6% revenue growth is strong and its valuation looks cheap on the surface at 18.38x earnings and just 0.63x sales, but a 3.7% net margin tells the real story. System assembly without durable pricing power can become a volume business fast, especially when customers are spending aggressively but still demanding efficiency. That does not mean the company cannot work tactically. It means the market is less willing to pay up for exposure that sits close to the spending wave but does not clearly control the choke point.
The better historical analogy is not "AI is a bubble" but that infrastructure booms mature. Early on, the market buys the obvious winner and then buys everything adjacent to it. Later, it figures out where the true scarcity sits. In AI, that scarcity is increasingly about moving data, feeding memory, and integrating systems at scale. The companies that solve those constraints can still command premium multiples even in a tougher tape. The ones riding the theme without owning the bottleneck are the ones most exposed when expectations turn hostile.
The right question now is not whether AI demand is real. Public filings and company commentary already answer that. The right question is which parts of the stack can still defend margins and valuation when investors stop giving out free passes for "AI exposure." On that test, networking, custom interconnect, memory-adjacent plumbing, and system integration look more durable than the old blanket semiconductor trade.
What would change our mind? If the next round of results shows GPU leaders reaccelerating without any sign of networking or system constraints binding, then the market may go back to rewarding compute concentration above all else. But after Broadcom’s reset and the broader June pullback, our view is that the smarter AI trade is no longer about owning the headline chip. It is about owning the bottleneck.
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.
