AI Infrastructure
Portfolio

July delivered the first genuine de-rating of the AI infrastructure trade since the current buildout began. The repricing was not driven by weakening demand; it reflected lower tolerance for elevated valuations, increasingly concentrated ownership and the growing financing burden attached to unprecedented capital expenditure. The Nasdaq-100 declined approximately 7%, the semiconductor complex fell more than 20% from its June peak and over US$1 trillion of semiconductor market value was erased, while equal-weighted equities materially outperformed as investors rotated toward value and cyclicals.

Several company-level events changed the market’s interpretation of the compute layer. Meta’s plan to resell surplus AI capacity raised the prospect that a major customer could become a competitor to neocloud providers. Weakness across Korean memory names, Intel’s foundry-yield concerns and a negative share-price response to higher TSMC capital expenditure reinforced the same message: spending that had previously been rewarded as evidence of demand was increasingly treated as a cost, capacity and return-on-capital risk.

The physical buildout nevertheless continued to accelerate. Combined 2026 capital expenditure across Microsoft, Alphabet, Meta and Amazon is tracking toward approximately US$760 billion, compared with about US$413 billion in 2025, with all four companies signalling further increases. The market’s concern is therefore not whether infrastructure will be built, but whether monetisation can rise quickly enough to justify the cash flow being consumed. Alphabet’s share-price decline after raising its capex guidance illustrated how sharply investor tolerance has narrowed, although strong Azure and cloud results late in the month provided partial reassurance.

Macro conditions amplified the de-rating. The Federal Reserve held rates at 3.50–3.75% on a 9–3 vote, with three policymakers dissenting in favour of a hike. The 30-year Treasury yield moved above 5.2% and the 10-year exceeded 4.7%, increasing the discount applied to long-duration AI cash flows. At the same time, renewed Middle East conflict pushed Brent crude above US$85 per barrel, adding an energy and input-cost shock to the higher discount-rate environment.

The key portfolio divergence was between electrons and electronics. Semiconductor and merchant-compute exposures absorbed the most severe repricing, while the power and grid layer held up better, supported by firmer order books, shorter-duration revenue visibility and continued data-centre demand. GE Vernova’s margin expansion across Power and Electrification reinforced this distinction. FERC’s requirement that data centres bear the grid-upgrade costs associated with their interconnections should also accelerate behind-the-meter generation and co-location structures rather than weaken underlying power demand.

Regionally, Asia carried the greatest exposure to the memory and foundry correction, while Europe’s lighter technology weighting provided relative insulation. For portfolio positioning, July narrowed the bear case to three issues: hyperscaler capex is consuming the free cash flow investors historically valued; vertical integration is threatening margins in the merchant compute layer; and 5% long-term yields materially reduce the present value of distant cash flows. The investment implication is to favour the constraint layer—power equipment, generation, grid, cooling and contracted infrastructure—where order books are harder and cash flows are nearer, while applying greater valuation discipline to merchant compute and semiconductor exposure.

AI Infrastructure Q2 2026: The Capex Revenue Gap Is Now the Central Risk Variable

The bull case for AI infrastructure has never depended on whether demand exists. The central question is whether monetisation can catch up with spending before investors lose patience. Q2 2026 was the quarter in which that gap became impossible to ignore.

The scale of spending is unprecedented. The four largest US hyperscalers are guiding to approximately US$725 billion of capex in 2026, up roughly 77% year on year from about US$410 billion in 2025. Amazon is the largest at around US$200 billion, Microsoft is near US$190 billion, Google is guiding to approximately US$175 185 billion, and Meta to US$115 135 billion.

Against that spending base, current AI related cloud revenue remains materially smaller. Google Cloud is running at roughly US$80 billion annualised, AWS near US$150 billion annualised, and Azure AI around US$37 billion annualised. On some estimates, the gap between hyperscaler AI infrastructure spending and ecosystem revenue is now approximately US$600 billion per year and it is widening in 2026 rather than narrowing.

The return on capital hurdle remains demanding. Assuming hyperscalers require a 25% return on AI specific capex, the industry would need to generate approximately US$169 billion of annual AI attributable revenue by the end of 2028. Current AI cloud revenue is estimated near US$150 billion annualised. That is a credible gap to close, but it is still a shortfall and one that equity markets are not yet fully pricing as risk.

This is increasingly a financing story, not only a spending story. Hyperscalers are leaning more heavily on debt markets to bridge the gap between AI capex and internal free cash flow, marking a structural departure from historically cash funded models. More than US$100 billion of hyperscaler debt had reportedly been issued by mid March 2026, compared with roughly US$80 billion during the whole of 2025.

Oracle is the clearest stress case. Its large compute agreement with OpenAI drove a substantial increase in capex guidance to approximately US$50 billion, creating a funding gap of more than US$27 billion. Oracle’s five year credit default swap spread has more than tripled since September, with trading volumes well above prior norms. Credit markets, rather than equity markets, appear to be where the first meaningful scepticism is emerging.

Why can the hyperscalers not stop, even if ROI remains uncertain? Pulling back carries its own strategic risk. The companies that build the largest and most efficient data centres first gain asymmetric advantages in GPU access, training speed and partnership economics. Hyperscalers are not spending US$725 billion because returns are already proven; they are spending because being short of compute is the one mistake none of them can afford to make.

That creates a coordination game dynamic. Rational individual behaviour continue spending can still produce a poor collective outcome if enterprise monetisation disappoints. Because four companies are making the same bet simultaneously, the downside is highly correlated. If enterprise AI adoption stalls, the capex stack is likely to re rate across the board rather than gradually.

The risk is therefore asymmetric. As the capex number rises, the bridge between spending and eventual ROI becomes more fragile. In more aggressive scenarios, industry capex could approach US$1.4 trillion by 2027, which would deepen the funding cycle rather than resolve the monetisation question.

Portfolio positioning should distinguish between layers of the stack. The picks and shovels layer chips, power, cooling and data centre REITs has the clearest near term revenue visibility because it is supported by signed capex commitments. These businesses are paid regardless of whether enterprise AI applications monetise successfully.

The application and software layer is where the US$725 billion ultimately has to convert into durable revenue, and it is therefore the layer most exposed if the gap does not close by 2027 2028. This is a useful lens for ECATS Momentum weighting: infrastructure layer momentum is currently backed by contracted spending, while application layer momentum still depends on monetisation that has not yet been fully proven.

Q3 watch list: hyperscaler earnings calls for any change in ROI language or capex discipline; further widening in Oracle style CDS spreads as a leading indicator of credit market concern; and whether Azure, AWS and Google Cloud AI revenue growth can remain in the current 48 123% year on year range while the capex base continues to grow faster than revenue.

Reading the Market’s Second Layer

The index stayed calm while the AI stack fractured.

July was a clear example of why headline volatility can understate portfolio risk. The VIX declined from 16.45 to 15.99 and rose materially only once, reaching 20.66 on FOMC day. Yet the Nasdaq-100 fell approximately 7%, the semiconductor complex lost more than 20% from its June peak and individual AI infrastructure holdings experienced much larger moves. Risk was concentrated within the theme rather than expressed as a persistent market-wide volatility event.

The underlying regime was one of extreme dispersion. Merchant compute, semiconductors and neocloud exposures repriced sharply as investors questioned monetisation, capital intensity and competitive structure, while power, grid and selected infrastructure beneficiaries proved more resilient. Broad equity indices muted this divergence because gains in energy, financials and value-oriented sectors offset losses across long-duration technology.

For an AI Infrastructure portfolio, that distinction is critical. A broad Nasdaq or S&P 500 put can protect against a correlation shock, but it is less effective when portfolio holdings decline for stock- and sector-specific reasons while the wider market remains supported. July therefore favoured protection placed closer to the underlying risk: single-stock puts, collars and put spreads on the highest-beta semiconductor, networking and merchant-compute positions.

Dealer positioning and event risk

Dealer gamma remained an important timing input around earnings and the FOMC. Positive gamma can dampen index moves and help explain why headline volatility remained contained, but it does not prevent large gaps in individual securities after company-specific information. Meta’s compute-capacity plans, memory-supply commentary and hyperscaler capex guidance all carried the potential to move single names independently of the index.

The brief VIX spike on 29 July showed that systemic protection still has a role, particularly when dealer positioning turns negative and hedging flows risk amplifying a broad decline. The sizing lesson is to separate two exposures: maintain modest, inexpensive index convexity for correlation shocks, while allocating the larger share of the active hedge budget to the portfolio’s actual concentration risks.

Where the carry opportunity moved

Single-stock implied volatility and downside skew became materially richer as dispersion expanded. That created selective opportunities to harvest premium through covered calls on positions where upside targets were already extended, while using the proceeds to finance collars or put spreads on more vulnerable holdings. The opportunity was not indiscriminate volatility selling: rich premium around earnings and strategic announcements often represented genuine gap risk.

The preferred July structure was therefore layered rather than directional: targeted stock-specific protection for semiconductor and merchant-compute risk; selective call writing where implied volatility compensated for surrendered upside; and smaller index hedges retained as inexpensive insurance against a future rise in correlation. The portfolio’s volatility framework must follow where the risk is priced, not where the index suggests it should be.