The AI Capex Supercycle: $725 Billion, Bottlenecks Everywhere, and ROIC Hanging in the Balance

The AI Capex Supercycle: $725 Billion, Bottlenecks Everywhere, and ROIC Hanging in the Balance

I am long the AI buildout. I am also willing to name the number that nobody is clearly stating.

Goldman Sachs calculates that maintaining current returns on capital would require these companies to realise an annual profit run-rate of over $1 trillion by 2027. Consensus projects $450 billion. That is a 2x gap — and it will not resolve uniformly across all four hyperscalers.

Through my Asian lens is a reader-supported publication. To receive new posts and support my work, consider becoming a free or paid subscriber.

The numbers first:

• $725B combined 2026 capex — Google, Amazon, Microsoft, Meta. Up 77% from 2025’s record.

• 94% of operating cash flow consumed by AI infrastructure before debt financing (Goldman Sachs).

• $121B in hyperscaler debt issuance in 2025 — four times the historical average.

• Google Cloud backlog nearly doubled quarter-on-quarter to $460B. The demand is real.

The physical layer bites back:

The binding constraint is not GPUs. It is electricity. Transformer backlogs stretch 5+ years. Nearly half of the planned 2026 data centre projects are delayed or cancelled due to power shortages. The White House has invoked the Defence Production Act, classifying grid infrastructure as a national security emergency. That is a meaningful policy signal — but policy moves in years, not quarters.

The trade is bifurcated, not homogeneous:

Goldman data: investor correlation across AI hyperscalers dropped from 80% to 20% in nine months. Markets are already differentiating between companies that can tie spending to revenue and those that cannot. Meta dropped 6% on its capex raise. Alphabet traded higher on Cloud strength. These are not the same bet.

The APAC angle that every generic note misses:

US grid constraints and data sovereignty fragmentation are routing capacity to alternative geographies. Singapore sits at the intersection of US-aligned infrastructure, ASEAN digital growth, and regulatory predictability. The bottleneck upstream is an opportunity downstream.

My position: long the buildout, selective on the names, watching the Singapore-Johor data corridor for the APAC overflow play.

Are markets pricing the bottlenecks or just the dream? And which layer - infrastructure, platform, or productivity beneficiary — is where the real money gets made from here?

Full analysis with hyperscaler comparison, ROIC breakdown, DPA implications, and APAC positioning matrix below.

(This is a collaborative document with Grok, Claude and Kimi.)

The AI Capex Supercycle: $725 Billion, Bottlenecks Everywhere, and ROIC Hanging in the Balance

The largest concentrated infrastructure cycle in corporate history is underway. The demand thesis is real. The physical-world constraints are binding. The ROIC math has a $1 trillion gap that nobody is naming clearly. Here is the honest read — bullish on the buildout, sceptical on the timeline, selective on the names, and watching APAC for the opportunity the US constraints are creating.

Author: Benson Kong (KYH) with Grok, Claude & Kimi

Date: May 2026

Confidence: [H] on facts · [M] on ROIC timing

SECTION 00

EXECUTIVE SUMMARY

Problem: The four largest hyperscalers are committing $725 billion to AI infrastructure in 2026 alone. Maintaining current returns on capital would require over $1 trillion in annual profit by 2027. Consensus forecasts only $450 billion. The gap is real, and nobody is naming it.

Observation: The binding constraint is not demand or chips — it is electricity. Transformer backlogs stretch 5+ years. Half of the planned 2026 data centre projects are delayed or cancelled. The White House has invoked the Defence Production Act.

Finding: This is a bifurcated bet, not a sector-level bull case. Investor correlation across AI hyperscalers dropped from 80% to 20% in nine months.

APAC Signal: US grid constraints and data sovereignty fragmentation are routing capacity to alternative geographies. Singapore is the primary beneficiary in Asia.

Call to Action: Stop treating this as a homogeneous sector bet. Map your position — infrastructure, platform, or productivity beneficiary — and ask whether the ROIC math closes for the specific names you hold.

SECTION 01

The Numbers — Staggering, Rising, and Still Underestimated

This is not a 2026 story. It is a structural commitment. Q1 2026 earnings confirmed that three of the four hyperscalers raised guidance. Bank of America estimates that total combined capex will cross $1 trillion in 2027.

The number that isn't being stated clearly enough: Goldman Sachs calculates that maintaining current returns on capital — the returns to which investors have become accustomed — would require these companies to realise an annual profit run-rate of over $1 trillion by 2027. Consensus estimates project $450 billion. That is a 2× gap between what the spending requires and what analysts currently expect. Some of these companies will close it. The math suggests not all of them will.

Section 02

The Uncomfortable Math — ROIC, Depreciation, and the Timing Mismatch

The briefing note’s ROIC framework is correct as far as it goes: $1 of capex needs to support $0.30–0.50+ in annual high-margin revenue over time to deliver 15–25% unlevered ROIC. For $725B in 2026 capex alone, that implies hundreds of billions in incremental AI-driven annual revenue. Current AI revenue segments are growing at 30–100%+ and the backlogs are real. But the timing mismatch is structural, not temporary.

The depreciation mismatch — the issue nobody is leading with

Princeton CITP research finds that GPU hardware has an effective economic life of one to three years. These companies depreciate it over five to six years. The difference is not a footnote. It means the ROIC calculations that analysts present — and that appear in every bullish deck — are systematically overstated. When hardware that generates revenue for two years is depreciated over five, the reported returns look better than the economic returns. For investors evaluating real capital efficiency, the gap between accounting ROIC and economic ROIC on AI hardware is a material consideration.

The honest framing: the 15–25% ROIC range cited widely in AI infrastructure analysis assumes the 5–6 year depreciation schedule. Corrected for real hardware life, the ROIC corridor narrows — and the revenue ramp required to close the gap becomes steeper. This does not invalidate the bull case. It tightens the conditions under which it works.

“The magnitudes of current spending and market caps alongside increasing competition within the group suggest a diminishing probability that all of today’s market leaders generate enough long-term profit to justify the investment.” - Goldman Sachs Research · January 2026

Section 03

The Physical Layer Bites Back

The binding constraint is not GPUs. It is not capital. It is electricity — and the infrastructure needed to deliver it at the scale AI demands. This is the chapter every frontier infrastructure cycle has. Railroads, electrification, and internet backbone — they all had this chapter. The overbuild creates capacity; the bottlenecks create volatility; the resolution creates the value.

Power and grid infrastructure have emerged as the single largest constraint on the AI buildout. US data centre power demand is surging, but grid interconnections face years-long delays. Transformer backlogs stretch five-plus years. Nearly half of the planned 2026 data centre projects have been delayed or cancelled due to power shortages and electrical equipment constraints.

The White House has now invoked the Defence Production Act (Section 303) to address this crisis explicitly: “Grid infrastructure and its associated upstream supply chains, including transformers, transmission lines and conductors, substations, high-voltage circuit breakers, power control electronics, protective relay systems, capacitor banks, electrical core steel, and related raw materials... are industrial resources, materials, or critical technology items essential to the national defence... without Presidential action... United States industry cannot reasonably be expected to provide these capabilities in a timely manner due to limited domestic production capacity, extended procurement timelines, foreign supply dependence, and insufficient capital investment.” This is the US government treating AI infrastructure demand as a national security emergency. That is a meaningful policy signal — but policy moves in years, not quarters.

Section 04

The Trade Is Not Homogeneous — Markets Are Already Differentiating

I am long the buildout. The balance sheets are deep enough, the demand is real enough, and the long-term optionality is asymmetric enough to warrant conviction. But “long the buildout” is not the same as “long every hyperscaler at current multiples.” The easy part of this trade is over.

Goldman Sachs data point that should be on every investor’s dashboard: the average stock-price correlation across large public AI hyperscalers fell from 80% to 20% between June 2025 and early 2026. The market is no longer treating this as one trade. It is beginning to discriminate between companies that can tie AI spending to revenue and those that cannot. The $1 trillion profit gap will not affect all four hyperscalers equally — and the divergence is already priced in.

Meta dropped 6% after its capex raise at Q1 2026 earnings, while Alphabet traded higher on Google Cloud strength. Meta’s AI investment pays back internally through an advertising platform — the payback path is real but harder to verify quarter-by-quarter. Amazon’s free cash flow is projected to turn negative in 2026 despite its commanding AWS position. Microsoft’s capex includes $25B in component price inflation — a real cost but not a volume signal. The four are not in the same position, and treating them as equivalent is the analytical error most widely made.

Section 05

The APAC Signal — What US Bottlenecks Create in This Region

This is the angle that every generic investor note misses — including both sources this article synthesises. The US grid constraints and the data sovereignty fragmentation created by US-China tensions are routing infrastructure investment to alternative geographies. Singapore is the primary beneficiary in APAC, and the window is opening now.

The sovereignty window: hyperscalers cannot route Chinese-originated data through US facilities — and increasingly cannot route US-allied APAC data through facilities subject to Chinese jurisdiction. Singapore’s data centre market sits at the intersection of US-aligned infrastructure, ASEAN digital growth, and regulatory predictability. The US grid emergency the DPA invocation, creates a natural demand overflow that has nowhere to go in the US for 3–5 years. It needs an alternative geography with reliable power, skilled labour, and regulatory stability. Singapore is that geography — but with land scarcity, water constraints, and a moratorium history that limits how much capacity can be built.

PUKKA’s muse — Through My Asian Lens

I Am Long the Buildout. I Am Selective on the Names. I Am Watching Singapore.

The demand thesis is real. Generative AI reached 53% adoption within three years of the first widely available product — faster than the PC and internet adoption curves. McKinsey reports 88% of respondents use AI regularly in at least one business function. Google Cloud backlog nearly doubled quarter-on-quarter. The infrastructure being built will be used. The question is not whether the demand exists. The question is whether the revenue ramp can catch the spending curve, and on what timeline.

The Goldman $1 trillion gap is the number to track above all others. Maintaining current ROIC requires $1T+ in annual profit by 2027. Consensus projects $450B. That 2× gap will resolve in one of two ways: revenue accelerates beyond current consensus (the bull case), or multiple compression occurs as the market re-rates companies that cannot demonstrate monetisation. Neither is guaranteed, and they will not resolve uniformly across the four hyperscalers. I hold the conviction that AWS and Google Cloud have the most legitimate path to closing the gap. Meta’s internal payback thesis requires more patience than most institutional investors have. Microsoft is a genuine conviction position, but the depreciation mismatch requires careful modelling.

The physical bottlenecks are not a bear case. They are the feature of frontier infrastructure investing, not the bug. Every transformative infrastructure cycle — electricity, railroads, internet backbone — had exactly this chapter. The constraints create the pricing power that justifies the premium. The DPA invocation is a bullish catalyst — not because it solves the transformer problem tomorrow, but because it signals the US government will not allow the problem to persist. Policy timelines are years; the investment thesis is 5–10 years; the timelines are compatible.

For Singapore and APAC specifically, the bottlenecks upstream are an opportunity downstream. US grid constraints cannot be resolved in 12–24 months, regardless of DPA authority. Demand that would otherwise flow to US facilities needs somewhere else to go. Singapore is the governance-stable, US-allied, ASEAN-connected answer to that question — but the land and water constraints mean Singapore cannot capture all of it. Johor captures the overflow. The companies positioned along the Singapore-Johor data corridor, with compliance-grade governance and long-term power agreements, are holding a strategic position that the US grid emergency will only make more valuable over the next 3–5 years.

My question to this market: are you pricing the 2026 capex number, or are you pricing the architecture that converts $725 billion in infrastructure into $1 trillion in profit? Those are different bets, on different timelines, with different winners. The market has started to make that distinction — the 80% to 20% correlation drop is the evidence. The rest of us should, too.

The Question Worth Discussing

Are markets pricing the bottlenecks or just the dream? And which layer — infrastructure, platform, or productivity beneficiary — is where the real money gets made from here?

Share your read. Especially interested in APAC investor perspectives on the Singapore/Johor data corridor and the sovereignty infrastructure thesis.

@TigerStars

$Meta Platforms, Inc.(META)$

$Alphabet(GOOG)$

$Amazon.com(AMZN)$

$Microsoft(MSFT)$

References

1. Yahoo Finance / Reuters (May 2026). Hyperscalers Hit $700 Billion in 2026 AI Spending Plans. Q1 earnings summary. finance.yahoo.com

2. Tom’s Hardware / Financial Times (May 2026). Google, Microsoft, Meta, Amazon capex spending to hit $725 billion in 2026, up 77%. tomshardware.com

3. CNBC (April 30, 2026). Big Tech AI spending approaches $1 trillion in 2027. Bank of America analysis; individual company guidance. cnbc.com

4. Fortune / Goldman Sachs (January 2026). Tech companies may only get half the profit they need to justify their AI investment. $1T profit requirement vs $450B consensus. fortune.com

5. Goldman Sachs (December 2025). Why AI Companies May Invest More than $500 Billion in 2026. Correlation drop 80%→20%. goldmansachs.com

6. Prism News / Goldman Sachs (May 2026). Goldman Sachs sees enterprise AI spending hinging on productivity gains. Stanford HAI adoption data; McKinsey survey. prismnews.com

7. Medium / Truthbit.ai (December 2025). The $2 Trillion Question: Can AI Revenue Catch Up to Capex? Princeton CITP GPU depreciation mismatch; Goldman 94% cash flow figure. medium.com

8. White House Executive Order 14156 (2026). Defence Production Act Section 303 — Grid Infrastructure as National Security Emergency.

9. Futurum Group (February 2026). AI Capex 2026: The $690B Infrastructure Sprint. APAC context; Oracle; Chinese hyperscalers. futurumgroup.com

10. Pukka Intelligence (May 2026). AI That Cannot Learn Cannot Compound — The Gartner ROI Reckoning. Cross-reference: 80% cut jobs, zero ROI correlation.

11. Pukka Intelligence (May 2026). The Fearless Partner — PwC AI Barometer through the AIHI Lens. Cross-reference: 3× revenue per employee; 56% wage premium.