The Stagflation Trigger Has Moved: From Hormuz to the AI Capex Cycle

Thesis

Stagflation needs two things at once: inflation the central bank can't ignore and a growth shock it can't fix. Since March the US has had the first. The Iran conflict pushed headline CPI up by a full percentage point, household inflation expectations have jumped, and the Fed has started hiking.

What it hasn't had is the growth shock, and the reason is narrow. AI-related investment and the equity wealth it has generated are carrying an economy whose other parts are close to stall speed. That makes the growth side of the stagflation equation hostage to one variable: whether hyperscaler capital spending keeps accelerating.

Capex doesn't need to fall to hurt GDP. It only needs to stop growing. A plateau in 2027, with oil still near $100 and a Fed that has just committed to tightening, would deliver the stagflationary regime our March piece described, by a different route. We think markets are pricing the 2027 capex path as settled. It isn't.

Six months on: the March scorecard

We owe readers an honest accounting before recasting anything.

Core CPI data note: widely cited secondary sources report August core CPI inflation at both 2.4% and 3.1% y/y. We have not resolved this against the BLS release and do not rely on either figure in this note.

The inflation mechanics worked. The error was assuming the Fed would treat an energy shock as something to look through. Under Chair Warsh it did the opposite, and real yields rose. That removed the single most important support in our gold case, which rested on real rates stagnating or falling. Record central bank buying kept gold from falling further, but a floor is not a return.

The risk we named against the gold thesis in March was de-escalation. The risk that actually materialised was a Fed willing to tighten into an oil shock. We underweighted it.

Two economies, one headline

The US economy is growing, but most of that growth is coming from a single source. Every serious attempt to measure it lands in the same area: AI-related investment accounted for somewhere between a fifth and a half of recent GDP growth, depending on definition and whether imports are netted out.

The upper estimates overstate the net effect. Much of the hardware is imported, and a widening trade deficit driven largely by imported chips subtracted more than a percentage point from Q2 growth. But the pattern outside AI is clear enough without precision. Non-tech business investment fell year-on-year for six straight quarters to Q1 2026. Factory construction fell at a 17.2% annualised rate in Q2 and is 26% below its 2024 peak.

There is a second, indirect channel. ING estimates that $33 trillion of extra household equity wealth since late 2022 has added roughly $500 billion of cumulative consumer spending, about 0.65 percentage points of spending growth. That wealth is concentrated in the top income brackets and heavily exposed to AI equities. EY describes income erosion for lower- through upper-middle-income households and for smaller firms outside the AI investment orbit.

So the answer to "is the US slowing?" is yes, outside the part of the economy that is building data centres or owns the stocks of the companies that are.

The capex arithmetic

AI investment contributes to GDP through its change, not its level. If hyperscalers spend $900 billion in 2027 and $900 billion in 2028, 2028's contribution to growth from that spending is roughly zero. This is the point markets tend to miss when they read flat capex guidance as reassuring.

The bar is also rising. 2026 capex estimates for the largest spenders have been revised up at almost every earnings season this year, and consensus for 2027 already assumes further strong growth.

Consensus for the four largest spenders in 2027 sits at $934.5 billion, against $725 billion planned for 2026. That's about 29% growth on a base that already rose roughly 77%. Every percentage point of shortfall comes straight out of the GDP contribution.

One offset matters. When capex slows, imports of chips and servers slow too, so the net GDP hit is smaller than the gross spending change. The domestic components (data centre construction, power and grid equipment, electrical contractors, and the wealth effect) carry most of the real exposure. They are also where the second-round effects on employment would show up.

The funding strain is already visible

The capex cycle has outrun the cash flow that funded it. Through 2024 hyperscalers paid for their buildout almost entirely from operations. By mid-2026, incremental debt had risen to 32% of capex over the trailing year, up from 9% in FY2024, and equity is back in the mix.

Alphabet's debt has grown from about $16 billion a year ago to roughly $100 billion. That financing now reprices against a 10-year Treasury at 5%. The equity market has noticed. Semiconductor earnings grew 131% in Q2 while the sector ETF fell 23%, because capex growth pushed hyperscalers' near-term free cash flow negative and investors cut multiples. On 14 September the Philadelphia Semiconductor Index fell almost 6% in a session after AI lab leaders called for caution on frontier development, and OpenAI's IPO was delayed.

None of this is a capex cut. No hyperscaler has lowered guidance. But the first cracks are in funding and valuation, which is where capex decisions for 2027 will be made.

Scenario framework

In March the regime variable was the oil price. The regime variable now is the growth rate of AI capex in 2027. The figures below are Gallium estimates, built from the contribution ranges above, and are illustrative rather than forecasts.

The plateau case deserves the closest attention, for the same reason the $110 case did in March. It requires nothing dramatic. A few hyperscalers guiding 2027 capex flat, citing funding costs or slower enterprise monetisation, is enough. We put the probability of a plateau or worse through 2027 at roughly one in three. That's still below our base case, but well above what equity and credit prices imply.

The regime map

Oil hasn't gone away as a variable. It now sits alongside capex, and the combination determines the regime. A Hormuz settlement, which Washington has signalled could follow the November midterms, moves the economy across one axis. A capex plateau moves it across the other.

Two of the four quadrants involve the Fed cutting or being forced to, and both are on the capex-plateau side. That's the asymmetry the market is mispricing.

Where the market is mispriced

The Fed's own projections assume the build continues. The September median has real GDP growth of 2.3% in 2026 and 2.4% in 2027, with unemployment flat at 4.1% through 2029. Those numbers are hard to reach without AI capex growing at something close to consensus.

Equities have priced the same thing. The S&P 500 is up nearly 12% this year despite a 5% 10-year yield and $100 oil, because earnings keep rising, and a large share of that earnings growth belongs to companies selling into the capex cycle. Credit is priced for calm. Investment-grade spreads of 77bp offer little compensation for weaker growth, persistent inflation or heavy corporate issuance, and data-centre financing is a growing share of that issuance.

The inflation risk from March hasn't gone away. It has simply been joined by a growth risk that the market treats as remote because it has been deferred for three years running.

Portfolio implications

Gold: hold the core, don't add yet

Gold's March case was right about the drivers and wrong about the sequence. It needs falling real yields, and those arrive when the Fed stops hiking, most likely in a capex-plateau quadrant. Until then the central bank bid provides a floor, not a catalyst. We would keep existing strategic allocations and wait for evidence of a capex slowdown before adding.

Duration: extend

In March we favoured a shorter-duration bias. With the 10-year at 5%, that call should reverse. Duration now pays in both plateau quadrants and in the reflation quadrant, and a 5% yield provides carry while waiting. We would move from underweight toward neutral, adding in the 5–10 year segment rather than the long end, where term premium remains the risk.

AI exposure: sort by funding quality

The question is no longer whether to own AI. It is who funds the capex and who depends on its growth rate. Companies spending from free cash flow are better placed than those funding with debt. Suppliers whose earnings depend on capex growth (chip designers, memory, neoclouds) carry more risk in a plateau than suppliers tied to the installed base or to long-dated contracts. We would reduce exposure to debt-funded builders and neoclouds first.

Credit: underweight long-dated IG tech and data-centre paper

Spreads at current levels don't compensate for issuance supply or for a plateau scenario. Shorter maturities and higher-quality issuers outside the AI supply chain offer better risk-adjusted carry.

Asia: the supply chain carries the second derivative

Taiwan and Korea are the most exposed economies to a slowdown in capex growth, through semiconductor, memory and server exports, and their currencies with them. A plateau would hit Asian export earnings before it shows up in US GDP. Singapore's exposure runs mainly through data-centre development and power infrastructure, which are tied to contracted capacity and should prove steadier.

Indicators to monitor

• 2027 capex guidance on Q3 earnings calls from late October, particularly any hyperscaler guiding flat.
• Free cash flow after capex at Alphabet, Meta and Microsoft, and the share of capex funded externally.
• Funding events for AI labs: IPO timing, private round valuations and compute-lease terms.
• Information-processing equipment and software contributions in the Q3 GDP advance estimate.
• Investment-grade issuance tied to data centres, and IG spreads against their 77bp level.
• Brent relative to $100, and the Michigan year-ahead inflation expectation relative to 4.6%.

Conclusion

The March piece identified the right regime and the wrong trigger. Oil delivered the inflation. The Fed's willingness to hike absorbed it, and AI investment kept growth positive. That leaves the US with an inflation problem already in place and a growth engine that depends on one sector continuing to accelerate. If AI capex growth stalls while oil stays near $100, the Fed will face the trap we described six months ago, with rates already higher and less room to move. That combination remains below our base case. It is closer than equity and credit prices suggest, and the next two earnings seasons will tell us which side of the map we're moving toward.