The AI bubble debate is missing the power bottleneck
AI demand can be genuine without making every AI-linked stock a durable winner. As the market demands earnings proof, the better trade may be the constrained infrastructure needed to power, connect, and cool AI.
The AI bubble debate is aimed at the wrong bottleneck. With 32% of fund managers calling an AI bubble the market’s top tail risk and NVDA reporting on Aug. 26, investors are asking whether hyperscaler spending can keep supporting semiconductor valuations. But the more useful question is what happens when demand survives and the physical system cannot keep up. AI can be a real growth cycle while the next winners shift toward networking, electrical equipment, and reliable power rather than every company benefiting from the same capex wave.
The constraint is moving from imagination to infrastructure. Electricity demand in the United States is expected to rise at a 2.6% annual average rate over the next decade, with data centers a major driver, while market research increasingly identifies power as a key bottleneck for the AI ecosystem. The most revealing figure is not a chip shipment or a model benchmark: only 5 of 16 gigawatts of announced 2026 data-center electricity capacity is currently under construction. That gap says AI demand may be arriving faster than the grid, transmission connections, and generation projects needed to serve it.
This changes the investment debate. When capacity is scarce, exposure to rising capital expenditure is not enough; the scarce input has to translate into pricing power, visible orders, or a strategic role in getting a facility online. A chip designer can benefit from every additional accelerator deployed, but a data center still needs the electrical distribution, cooling, grid connection, and high-speed networking that make those accelerators productive. The market should therefore distinguish between companies selling into the AI narrative and companies addressing the physical reasons the narrative could remain durable.
That does not make the leading chip names hollow. AVGO is the clearest evidence that AI spending is already converting into earnings: it reported $22.2 billion of second-quarter fiscal 2026 revenue, up 48% year over year, while AI semiconductor revenue reached $10.8 billion, up 143%. Its $16.0 billion third-quarter AI semiconductor revenue guide implies more than 200% year-over-year growth. Those figures are not merely a proxy for investor enthusiasm. They show a supplier monetizing a genuine buildout, which is why the right conclusion is not to short AI indiscriminately but to become more selective about where the economics are strongest.
The same distinction matters for NVDA ahead of its Aug. 26 report. The company’s listed metrics show 65.5% revenue growth, 66.0% earnings growth, and a 63.0% net margin, with a market capitalization of $5.07 trillion. That is an unusually powerful earnings engine, not a business surviving on thematic association. Yet its scale also raises the standard for fresh evidence: investors need continued conversion of AI demand into revenue and profit, not just more announcements about future spending. At 35.84 times earnings on the listed metrics, NVDA is not priced like an ordinary semiconductor company, even if its 0.29 PEG ratio reflects the pace of expected growth. The stock can remain a core monetizer while still being vulnerable to any sign that spending is being delayed by power or productivity constraints.
Yes, AI bulls can point to server and model productivity gains, hyperscaler commitments, and the possibility that better economics will unlock still more demand. They can also argue that a power bottleneck delays deployment rather than cancels it. But that argument strengthens the case for the infrastructure layer: if demand is real and deployment is delayed, the suppliers that solve the delay become more strategically important. ANET is a useful example. Arista posted its first $3 billion quarter in the second quarter of 2026 and described networking as the central nervous system of AI and data-center environments. Its stock trades at roughly 40.6 times forward earnings, a premium that demands execution, but its role is tied to the bandwidth required to connect expanding compute clusters rather than simply to the hope that AI spending continues.
The power trade is broader and more complicated, which is why the comparison should not be reduced to buying every utility or electrical-equipment name. ETN offers a direct way to express the electrical infrastructure bottleneck: its second-quarter 2026 results showed 14% organic sales growth and strength in data-center end markets, while the stock trades at about 27.9 times forward earnings. Its listed revenue growth is only 10.3%, so the market is already paying for a durable infrastructure opportunity rather than a near-term hypergrowth profile. That can be a more defensible setup if AI deployment keeps expanding but the scarce value shifts toward power management and distribution.
CEG shows both the opportunity and the risk. It trades at about 22.2 times forward earnings and signed an additional 920 megawatts of long-term power purchase agreements, while full-year operating EPS guidance was raised to $11.50–$12.50. But its listed earnings growth is negative 37.8% and its YTD return is negative 25.7%, a reminder that a power bottleneck does not automatically produce smooth earnings or immediate stock performance. CEG is exposed to the right constraint, but investors still have to assess generation economics, contract quality, execution, and valuation. The infrastructure thesis is therefore about selectivity, not a blanket rotation from technology into utilities.
The valuation comparisons reinforce that point. On the listed metrics, AMD trades at 79.11 times earnings after a 103.9% YTD gain, despite 34.3% revenue growth, while NVDA trades at 35.84 times earnings with much higher revenue and net margins. AMD’s 164.4% earnings growth is meaningful, but the multiple and the stock’s run leave less room for disappointment. By contrast, AVGO combines a 38.8% net margin with 286.6% earnings growth on the listed figures, while ANET’s 38.4% margin supports a networking role that is increasingly difficult to bypass. These are not identical businesses, and valuation alone does not settle the trade. The point is that capital should favor demonstrated monetization and indispensable constraints over the weakest form of AI beta.
The next phase of the AI trade should be judged by what the system cannot do without. NVDA and AVGO still have the strongest cases among the core monetizers because their growth is visible in revenue and earnings, but the bottleneck gives ANET, ETN, and selectively CEG a more durable strategic angle than companies whose main asset is exposure to rising capex. We would rather own the infrastructure that turns demand into deployable capacity than pay indiscriminately for the promise of more demand.
What would change our view is evidence that power availability is improving fast enough to remove the constraint, while chip and networking suppliers lose pricing power or fail to convert orders into earnings. Until then, the market should stop treating the AI debate as a binary question of bubble or no bubble. The better question is which businesses control the scarce inputs that allow the boom to continue.
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.
