Experience Stern | Faculty & Research

The four largest U.S. hyperscalers — Amazon, Alphabet, Microsoft, and Meta — each disclosed 2026 capex budgets above $100 billion. The group is on track to spend $635 to $690 billion this year, three-quarters of it on AI infrastructure: GPUs, high-bandwidth memory, networking silicon, transformers, cooling, and power. This is the largest concentrated capital cycle in corporate history. It is also, in my judgment, where the consensus AI trade has gone wrong. Cap-weighted indices have placed passive investors on the buyer side of this cycle just as the supply side began to capture its economics.

The conventional view

NVIDIA, Alphabet, Apple, Microsoft, and Amazon together account for approximately 27 percent of the S&P 500. NVIDIA itself, at roughly 7 percent, is the largest AI supplier in the world and the one mega-cap on the right side of the trade. The remaining four account for approximately 20 percent of the index, all on the demand side. Most U.S. equity portfolios are now an AI-buyer concentration trade with a single offsetting supplier position embedded within them. For two years, this was benign: the hyperscalers funded capex out of operating cash flow, margins held, and multiple expansion did the rest. That phase has ended.

What changed in the last two quarters?

Three shifts in recent disclosures have together inverted the trade. Firstly, capital intensity at the hyperscalers has reached utility-like levels. Microsoft's capex now consumes 45 percent of revenue; Oracle's recent quarter reached 57. Amazon's free cash flow collapsed to $1.2 billion last quarter as a $59 billion year-over-year increase in infrastructure spending consumed nearly all of its operating cash flow. The hyperscalers raised $108 billion in debt during 2025; cumulative issuance is projected past $1 trillion before the cycle ends.

Secondly, supply has become the binding constraint. SK Hynix has sold its entire 2026 HBM output, and Samsung's tripling of capacity by Q4 is unlikely to clear demand until late in the year. IDC projects DRAM revenues to nearly triple in 2026 to $418 billion, with the average price moving from $3.76 per gigabyte to a forecast $9.71. Microsoft has attributed $25 billion of its 2026 capex to component price inflation rather than capacity expansion: one company telling shareholders that supplier pricing power is reshaping its cost structure.

Thirdly, equity returns have followed the underlying economics. As of Tuesday's close, even after a sharp single-day selloff in the chip and memory complex, the year's leaders remain the suppliers: SanDisk (+490 percent YTD), Bloom Energy (+280), Intel (+220), Seagate (+185), Western Digital (+180), Lumentum (+165), Micron (+155), FuelCell Energy (+120), Vertiv (+105), and AMD (+90). The selloff is instructive. The memory and logic names dropped 7 to 10 percent on profit-taking; the behind-the-meter power names rose. Even after the pullback, the basket has returned several times what the hyperscalers have, which is zero to twenty percent over the same window. The capex is hitting the buyers' cost line; the cash is landing in supplier P&Ls.

Where the index leaves passive investors

At the start of 2026, the supplier names listed above held roughly 1.2 percent of the S&P 500's weight; two of them, Bloom Energy and Lumentum, remain outside the index. The four non-NVIDIA mega-caps held, and continue to hold, approximately 20 percent. NVIDIA contributes a further 7 percent, but as a single concentrated position rather than diversified supplier exposure. It is conspicuous from these figures that at the entry point passive holders were exposed to the buyer side at a ratio of approximately 17 to 1 over the diversified seller side, with NVIDIA carrying the entire diversified exposure on the index's behalf.

The basket has since appreciated to roughly 4 percent of the S&P 500's weight, narrowing the present-day ratio to approximately 5 to 1. This compression is itself part of the diagnosis. Cap-weighted indices mechanically rebalance toward outperformers; passive flows now follow the supplier basket at higher prices. The investor who held the index in January owned the cycle's economics at a 17-to-1 disadvantage; the investor who buys today owns them at 5 to 1, having forgone the appreciation. Concentrated capex cycles with binding supply constraints generate large transitory wealth transfers, and indices structurally underweight the seller side at entry, then catch up only after the gains have been realized.

The counterarguments

Three serious objections deserve treatment. The mean-reversion case writes itself: this year's leaders will revert. It conflates two phenomena. Momentum reversal is documented in price-driven runs without structural underpinning; the 2026 memory cycle is supply-constrained, with capacity additions visible in capex disclosures and forward agreements. A second objection is that hyperscalers will recapture margin once the buildout is complete. They almost certainly will. But the relevant horizon for the investor is the buildout itself, not the recovery beyond it: a 24- to 36-month window during which capital intensity, debt issuance, and component inflation are rising in parallel. The serious objection is memory's own cyclicality. Every supplier-side rally in NAND or DRAM has eventually unwound, often violently. That bear case may prove correct in time. The claim here is more specific: the next 12 to 24 months are a supply-constrained window during which the cycle's economics accrue disproportionately to suppliers. It is, by design, a tactical reallocation with a defined exit.

What follows from this

The implications differ by portfolio mandate. For unconstrained portfolios, a 10 to 15 percent reallocation from broad U.S. equity into a diversified supplier basket — memory (Micron, Western Digital, SanDisk, Seagate), AI logic (AMD, Intel, Teradyne), optical interconnects (Coherent, Lumentum), data center power and cooling (Vertiv), and behind-the-meter fuel cells (Bloom Energy, FuelCell Energy) — sizes the bet to match the diagnosis. Entry can be staggered into pullbacks of the kind seen today. The reallocation carries an 18- to 24-month review trigger: if hyperscaler capex guidance flattens, if HBM capacity clears the order book early, or if memory ASPs roll over for two consecutive quarters, the position is reversed.
For institutional portfolios constrained on active risk, a partial equal-weight S&P 500 overlay replacing 20 to 30 percent of the standard market-cap allocation captures the bulk of supplier dispersion without individual-name risk.

For passive holders who will not deviate from the cap-weighted index, the question is whether anything has fundamentally shifted. It has. The S&P 500 is no longer a balanced expression of the AI economy; it is a concentrated expression of the buyer side with a single supplier position (NVIDIA) embedded within it. Holding it is defensible for horizons that will outlast the current capex cycle. To believe one already owns the AI trade, diversified across its economic actors, is something else.

The dispersion may compress, hyperscaler capex may moderate, and memory may eventually behave as memory has always behaved. Each outcome is plausible. The current configuration — utility-scale capital intensity at the buyers, sold-out capacity at the sellers, and an index that began the year 17-to-1 against the contracting side and has only mechanically closed the gap as supplier prices rose — is, in my judgment, the largest tactical mismatch between an index and its underlying economic activity that the post-2010 market has produced. Over the next 18 to 24 months, owning the cap-weighted index alone is to be on the wrong side of the largest capital cycle in corporate history.
____ 
Haran Segram is an Adjunct Assistant Professor of Finance.