Most retail investors hold one expression of the AI investment thesis: a US semiconductor exposure, often through an index fund. That position has been right. The Philadelphia Semiconductor Index — a basket of US-listed semiconductor companies — finished a 16-day winning streak on April 22, up more than 30% for the month (UBS CIO, “Transformational Innovation,” April 23, 2026).

For two years that one trade was the AI trade. The pattern is now changing. UBS’s sector view, published on the same day, downgraded the broad US Information Technology sector to Neutral, citing “capex moderation risk, AI tools commoditizing software, smartphone cycle cooling” (UBS CIO, “US Equity Sectors,” April 23, 2026). UBS’s sector ratings are cited here as third-party context. EXDS does not endorse or recommend allocation changes based on outside firms’ ratings.

The point is not the rating. The point is that the firm with the loudest AI-bull thesis in 2025 is signalling the priced part of the trade is mostly behind us, while the unpriced part is mostly ahead.

What the chip index actually owns

The graphics-processing units that do the heavy mathematical work of training and running AI models. This is the layer the equity gains have come from. It is the layer the financial press writes about. It is the layer almost every retail investor with AI exposure already owns.

Goldman Sachs’s sales note on April 24 made the next part of the conversation visible in trading-floor language: “The story is evolving beyond GPUs… memory and broader AI ‘scaffolding’ ie CPU + optics are emerging as the real bottlenecks” (Goldman Sachs, “Basics — Mag 6 Capex / AI Jobs,” April 24, 2026; sales note, not Goldman Sachs Investment Research). The largest AI sales desks at the largest banks are no longer asking clients about chip exposure. They are asking about everything the chips need around them in order to function.

What the chip index doesn’t own

Five exposures sit outside most diversified US equity funds and outside the semiconductor index entirely. They are the AI build that has not yet been priced.

Memory and CPUs. Memory chips store the data the AI model is actively using. Central processing units coordinate the rest of the system around the GPUs. Both have been treated as commodity components for two decades. Both are now constraining how fast new AI capacity can come online — the bottleneck Goldman named directly.

Optics. Once memory and CPUs are in place, the data has to move between server racks fast enough that the GPUs are not waiting. That movement happens through high-speed optical-fibre connections. There is no mass-market equity index for optical networking. There are specialist suppliers, and they are not in most diversified US equity funds.

Data centers. AI servers run hot — much hotter than traditional cloud servers. Cooling them, delivering reliable power to them, and finding sites with appropriate water and electricity access has become a constraint in its own right. None of the companies that make this build possible appear in the headline AI tickers.

The grid. This is where the most concrete data point in the UBS note lives. Citing Bloomberg New Energy Finance, UBS estimates that data centers could account for up to 9% of total US electricity consumption by 2035, against approximately 4% currently. A doubling of share over ten years, on a base — total US electricity demand — that is itself growing. The US grid was not built for that load. UBS’s sector ratings catch up to this directly: Utilities is rated Attractive, with the explicit rationale that “significant growth in AI data centers” will drive a strong increase in electric power demand. Roughly a quarter of utility companies in the sector index have material business exposure to AI-driven power demand (UBS CIO, “US Equity Sectors,” April 23, 2026).

Power generation and critical minerals. Behind the grid, the actual electricity has to come from somewhere. UBS’s thematic note broadens the value chain to “grid resilience, renewables, nuclear, industrial automation, and critical minerals.” The nuclear inclusion is recent in sell-side framing — nuclear had been treated as a mature, low-growth utility sub-sector. AI demand is rewriting that view. So is the read on copper, transformers, and the physical components of new transmission. None of this is in most diversified US equity funds either.

The honest risk

The migration view has a specific failure mode that the sell-side itself names openly.

Capital expenditure by the largest US technology companies — what the industry calls “hyperscalers” because they build computing capacity at very large scale — is rising fast at the same time those companies are cutting jobs. Goldman’s note observes that two of the largest hyperscalers, Meta and Microsoft, announced “mid-single digit %” headcount reductions while their data center build budgets continue to expand. The implicit question Goldman flags directly: “Are you about to get some serious question on cost pressure/ROI from the hyperscalers?”

If revenue from AI products catches up to spending, the migration trade works in both directions — chips and the unpriced infrastructure. If revenue does not catch up, the spending gets cut. Chips fall hardest, because chips are where the trade already is.

What to read in the next two quarters

The largest US technology companies report Q2 results from late July through August. The clean test for the migration thesis is in the published 2026–2027 capex guidance numbers in those reports — rising guidance keeps the unpriced exposures propagating; trimmed guidance freezes them and applies first pressure to chips.

The second test is relative-equity performance. Utilities and grid-related companies should outperform broader equities and even the semiconductor index over the next two quarters if the migration is being priced. If they continue to lag while semiconductors run further, the broadening view is a sell-side narrative that the market has not yet acted on — and the simple chip trade still has the dominant flow.

Conclusion

For two years, “AI investment” and “buy the chip index” were the same trade. They are not the same trade anymore. The chip-index buyers got one thing right: AI is real and compute would be the first beneficiary. The trade now splits into one priced exposure and five unpriced ones. Most retail capital owns the priced exposure and not the others.

Stay invested. Stay informed.