Writing · Markets
The semiconductor sell-off is not the end of the AI trade. It is a shift from scarcity to utilization, and a search for the next constraint.
Everyone is asking whether they should buy the dip in Nvidia.
I think the more useful question is: if GPUs are no longer the only scarce resource, where does the AI bottleneck move next?
The recent semiconductor sell-off exposed a new burden of proof. Investors are no longer rewarding infrastructure spending simply because AI requires it. They are asking who is financing the buildout, who captures the pricing power, and whether actual utilization can support the returns.
In other words, the market has moved from a scarcity trade to a return-on-investment trade. Everyone joined the AI party. Now the market wants to know who is paying the bill.
First, it became crowded. The IEA estimates that five large technology companies spent more than $400 billion in 2025, with spending expected to grow another 75% in 2026. When that much capital chases one theme, expectations can become more fragile than the underlying demand.
Second, efficiency is improving. Open models, custom accelerators and more efficient inference can reduce the compute required for a given task. That does not mean total chip demand disappears, usage may grow even faster, but it changes which companies capture the economics.
Third, monetization is now the main event. Buying GPUs was the easy part. Making them earn their rent is the awkward second date. The market wants proof that AI workloads can produce revenue, productivity and durable free cash flow before it rewards another round of spending.
This is why strong fundamentals and falling share prices can coexist. The buildout can continue while the market becomes much more selective about where profits accrue.
The scarce asset is not electricity in the abstract. It is reliable power at the right site, with grid access, transformers, switchgear and permits. The IEA has identified grid connections, transformers and gas turbines as active constraints. That puts companies exposed to electrical equipment, grid construction and firm power on the research map, but only where contracts and cash flow confirm the story.
AI systems are increasingly constrained by moving and storing data, not only by arithmetic. Micron says DRAM and NAND supply should remain tight beyond 2027 and has signed multiyear supply agreements. If that holds, memory can remain a bottleneck even as accelerator supply broadens.
Thousands of accelerators are not very useful if they cannot communicate quickly. Nvidia's networking revenue grew 162% year over year in its fiscal third quarter of 2026, while Broadcom reported 106% growth in first-quarter AI revenue. As clusters scale, the interconnect can become as important as the chip.
Higher rack density creates a less glamorous problem: heat. Vertiv entered 2026 with $15 billion of backlog, up 109% year over year. Liquid cooling, switchgear and facility redesign may not look as exciting as a new foundation model, but the servers still need to avoid becoming very expensive space heaters.
This is the ultimate constraint. Infrastructure becomes durable only when companies create useful workloads, paying customers and workflow-level returns. Cloud platforms and software companies with distribution may capture the longer-term value, but only if AI revenue grows faster than depreciation and capital spending.
My base case is a barbell. Physical bottlenecks, memory, networking, power and cooling, can capture near-term scarcity rents. The owners of high-utilization workflows can capture the longer-term economic value.
For broad exposure, the ETF matters less than what it actually owns:
For individual-company research, I am watching memory names such as Micron; networking and connectivity names such as Broadcom, Arista Networks, Marvell and Amphenol; power and grid names such as Eaton, GE Vernova, Hubbell and Quanta Services; and thermal-management names such as Vertiv, Johnson Controls and Schneider Electric.
That is a research list, not a model portfolio. A shortage is not automatically a good investment. Capacity can arrive, technology can route around the constraint, customers can insource, and the market may already price in years of flawless execution.
I would look for the overlap of three conditions:
Then I would monitor physical utilization, HBM pricing and long-term contracts, transformer lead times, cooling backlog, cloud AI revenue, depreciation and free-cash-flow conversion.
The next bottleneck and the next winning investment are not automatically the same thing.
Scarcity creates the first winners. Utilization determines the durable ones.
Disclosure: This material is for education and discussion only. It is not personalized financial, investment, tax or legal advice. Securities and ETF holdings change, and past performance does not predict future results. The author may hold securities discussed.
Originally published on Substack.