AI Capex Cycle — Where Are We in May 2026
Hyperscaler infrastructure spend is decelerating sequentially for the first time since 2022. Why it matters for NVDA, the power complex, and the speculative quantum/space tail.
The defining trade of the last 24 months — long AI infrastructure, short almost everything else — is showing its first cracks. Not collapse. Deceleration. The distinction matters.
The setup
Microsoft, Meta, Alphabet, and Amazon together guided combined FY26 capex of roughly $240 billion, up from $190 billion in FY25 and $130 billion in FY24. That is still a parabolic curve in absolute terms. But the sequential growth rate — the second derivative most equity markets actually price — is decelerating.
NVIDIA's most recent quarter (Q1 FY27, ending January 2026) printed +19.5% QoQ revenue growth. The prior quarter was +22%. The one before, +27%. The trend line is unambiguous: the rate of acceleration is fading, even as absolute dollars keep ripping.
This is not bearish. It is the shape of a maturing capex cycle — the same shape we saw in dotcom 1999, in fracking 2013, in semis 2018. The market historically pays the highest multiple right at the moment of peak second derivative, then derates even as numbers continue to grow.
What's confirming the shift
Three independent signals worth tracking:
-
Insider behavior at NVDA. $165m+ of net executive sales in a single two-week window in March 2026 (Kress, Puri, Stevens). One window means nothing. A pattern means something. We are watching for the next print.
-
Dark pool ratios across semis. Off-exchange volume as a percentage of total has crept up from ~38% in late 2024 to ~44% in Q1 2026 for the top five AI infrastructure names. This is institutional hedging, not retail capitulation.
-
The power complex is leading. VST, CEG, OKLO, SMR — the second-derivative beneficiaries — are outperforming the primary beneficiaries (NVDA, AVGO) on a 90-day basis. That historically happens at the back half of a thematic cycle, not the front.
What this implies for positioning
-
Trim AI infrastructure megacap exposure into strength. Not zero. Trim. Keep the multi-year compounders (NVDA, ASML, TSM) on a longer leash; tighten stops on the second-tier (AMD, SMCI, ORCL) where margin of safety is thinner.
-
Stay long the power and fuel cycle. Uranium spot remains constrained on supply, not demand. Nuclear utilities have multi-decade tailwinds from AI load growth. This trade has years left, not quarters.
-
Reduce speculative tail. The quantum names (IONQ, RGTI, QBTS) and the smaller space SPAC graduates (ASTS, LUNR) are levered beta to the broader risk-on regime. In a cycle deceleration, they crack first and crack hardest.
-
Optionality in defense. KTOS, AVAV — drones and autonomy — are not in the AI capex narrative but are in the geopolitical bid that typically intensifies when the AI mania cools. Cheap relative hedge.
What would change the view
Three things would force us back to "full risk-on":
- Hyperscaler FY27 capex guides materially above $300bn combined → reaccelerating second derivative
- China export curbs ease (low probability under current administration, but the China tech setup gets very interesting if so)
- A clean break of NVDA above $245 on rising volume — the technical confirmation of a new leg
Until then: accumulate quality on weakness, trim quality on strength, avoid junk on either.
This post is opinion grounded in observable data. It is not a recommendation to buy or sell any security. Position sizes, stop-losses, and risk tolerance are yours alone.