The AI Supercycle and the Question of Sustainability

A demand-led boom or a supply-driven mirage?

The current AI buildout is the largest concentrated capital cycle in modern technology history. Whether it joins the railroad and fiber buildouts as transformational — or the telecom bust as a cautionary tale — depends on a single question: can end-user revenue catch the spending curve before the financing math breaks?

Major Holders has reviewed the latest cycle data from MUFG, RBC Wealth Management, Goldman Sachs, Morgan Stanley, IDC, TrendForce and the principal hyperscaler 10-K filings. Our read is that the buildout is real, the contracted demand is real, and the productivity case in narrow verticals is real. But the gap between infrastructure spending and end-user value capture is widening, not closing, and the suppliers with the most information — memory manufacturers — are voting with their capacity decisions to not believe the projections.

This note synthesizes the data into four questions: the scale of the spend, the memory-chip chokepoint, the enterprise payoff gap, and the structural bottlenecks that will dictate the pace of the next phase.

$775B in one year — and climbing.

Microsoft, Alphabet, Amazon, Meta and Oracle have collectively guided to between $660B and $775B in 2026 capital expenditure, with the high case approaching $1.0 trillion in 2027. The composition has shifted decisively: roughly 75% of aggregate spend is AI-related, and approximately one-third of all hyperscaler capex now flows directly into memory chips — up from 8% in 2023.

The free-cash-flow consequences are already visible. Capital intensity at the Big Four has reached 45–57% of revenue — levels historically associated with utilities, not technology companies. Amazon's free cash flow is projected to turn negative in 2026. Hyperscalers raised $108B in debt in 2025 alone; Morgan Stanley and JP Morgan project up to $1.5 trillion in technology-sector debt issuance over the coming years to fund this trajectory.

The signal is unambiguous: AI infrastructure has become the dominant determinant of capital allocation at the world's largest technology firms. The question is whether end demand justifies the trajectory.

The chokepoint sits upstream of the GPU.

Memory — not compute, not power, not capital — is now the binding constraint on the speed of the AI buildout. The dynamic is structural, not cyclical.

Three manufacturers (SK Hynix, Samsung, Micron) control over 95% of global DRAM production. All three are deliberately reallocating wafer capacity toward High-Bandwidth Memory (HBM) for AI accelerators, where margins are several multiples higher than commodity DRAM. The result is what industry analysts have begun calling the wafer cannibalization effect: an active starving of the traditional computing market to feed AI infrastructure.

The supplier oligopoly

The most telling signal in the entire AI cycle comes from these three companies. They have lived through multiple semiconductor cycles. They understand that overbuilding leads to catastrophic price collapses. Their capital plans reveal what they actually believe.

Read this carefully: a rational oligopoly with 95% market share is choosing margin over volume and explicitly limiting capacity expansion until 2027–2028. Goldman Sachs projects undersupply persists through 2027. Micron has already sold out its 2026 AI memory contracts and acknowledges it can only meet roughly two-thirds of medium-term customer requirements.

This is not a market that expects insatiable demand to last. This is a market that has learned the lessons of 2018 and 2022 and is determined not to overbuild.

The enterprise monetization problem.

The capex/revenue gap is the single most important divergence in this cycle, and unlike past bubbles, it is widening, not closing. Historical infrastructure bubbles narrowed as adoption caught up to deployment. In AI, hyperscalers are doubling GPU orders while pure-play AI vendor revenue, while growing rapidly, remains a fraction of the infrastructure deployed on their behalf.

Goldman Sachs estimates that for capex returns to match historical norms, the AI hyperscalers would need to generate over $1 trillion in annual incremental profit. The 2026 consensus is approximately $450 billion — less than half. Free cash flow at the Big Four is projected to decline by up to 90% in 2026 as spending outpaces revenue.

The 95% problem — and what it actually means

MIT Project NANDA's 2025 study has become the defining data point of the skeptical case: 95% of enterprise generative AI pilots produced zero measurable P&L impact. The headline is widely misread. The substance is worse than the headline.

The failure is not the technology. The failure is the conversion of pilots into production: approximately 80% of the work required to move from pilot to production is data engineering, governance, workflow integration, and measurement infrastructure — work most organizations have not done. Vendor-led deployments succeed 67% of the time. Internal builds succeed one-third of the time.

The bifurcation is the story. The "AI high performers" — roughly 20% of enterprises using AI for revenue reinvention rather than narrow cost-cutting — are capturing approximately 74% of all economic value created. The remaining 80% are running pilots without measurement frameworks, producing the headline failure statistics. Software development and other narrow, high-specificity domains are clear winners; enterprise-wide horizontal AI deployments are not.

Reading the cycle from both sides.

Three constraints money cannot solve.

Even if demand justifies the spend, three physical bottlenecks will dictate the pace of deployment through 2028. None can be solved by capital alone.

The capital ceiling sits above all three. Big Tech cannot self-finance the physical infrastructure required on balance sheets alone without brutally punishing equity valuations. Total data center infrastructure costs through 2030 are estimated at $6.7 trillion. The arrival of sovereign-level capital — most visibly through the U.S.–UAE Stargate framework with MGX deploying billions into U.S. data centers in exchange for advanced Nvidia chip access — signals that this has already become a geopolitical infrastructure question, not a corporate one.

Two endings, both with precedent.

Historical analogues for this cycle are imperfect but useful. The railway boom of the 1840s, the fiber-optic overbuild of 1999, the cloud buildout of 2010–2020 — each saw excessive spending, each generated lasting infrastructure, but the equity outcomes diverged sharply.

The honest reading: this cycle exhibits characteristics of both. Probability-weighted, we see ~55% probability of a partial mid-cycle correction (significant equity drawdowns in over-leveraged participants, infrastructure absorbed and ultimately monetized) and ~25% probability each of the polar outcomes.

What we're actually watching.

The core asymmetry is this: demand created by the buildout itself currently exceeds demand from end-user value capture. That is the textbook definition of a supply-driven boom that needs end-demand to catch up before financing runs out.

Four indicators will determine which way this cycle resolves. We will be tracking them in subsequent notes:

1. Memory supplier capex behavior. If SK Hynix, Samsung and Micron remain disciplined through 2027, the structural supply story stays intact. The first sign of a serious capacity announcement from any of the three is a leading indicator that even they believe in the demand projections — which would be bullish for hyperscaler valuations and bearish for memory pricing.

2. Free-cash-flow trajectory at Amazon and Microsoft. If Amazon's FCF turns negative in 2026 as projected and the trend extends into 2027, debt markets begin pricing capex risk into the cost of capital. Watch credit spreads on hyperscaler bonds.

3. Enterprise conversion rate. The 20% of high-performers needs to widen to 35–40% within 18 months for the productivity case to hold. Watch survey data from McKinsey, Deloitte, MIT NANDA and the WRITER annual study.

4. Sovereign capital flow. The Stargate framework, UAE MGX deployment and similar structures are now part of the financing stack. If sovereign capital withdraws or terms tighten, the capital ceiling moves closer.

Your intuition — that "consumption for convenience will reach saturation" — is the right framing. The answer depends on whether the transition from convenience-AI to embedded-workflow AI happens before the financing math breaks. The strongest contrarian signal in the data: the people with the most information and the most to gain from the boom continuing — memory suppliers — are voting with their capacity decisions to not believe the projections.

We remain constructive on selective AI infrastructure exposure, cautious on undifferentiated AI software, and watchful on the four indicators above.