The slowdown is a gift to the hyperscalers
6 Nov 2026 · News
Yesterday we said a permit is the scarce asset. The same bottleneck is also cover for Google, Meta, Amazon and Microsoft to cut next-year CapEx without looking like they blinked first.
A slower US data center buildout helps the hyperscalers more than it hurts them. Political pauses and construction delays give Microsoft, Amazon, Alphabet and Meta a reason to guide CapEx down, protect free cash flow, and keep the gap over debt-funded neoclouds. The building continues under tougher terms. The companies with balance sheets and existing sites keep the lead.
CapEx already slipped on delays
The miss showed up before the midterms. On its late-October earnings call, Microsoft CFO Amy Hood said capital expenditures were $45.2 billion "as some third-party capacity deliveries we expected in Q1 were delayed." She cut the calendar-2026 CapEx figure to about $170 billion from about $175 billion for the same reason, and said demand still exceeds supply. CNBC reported Alphabet also came in under Street CapEx estimates and held its $195 billion to $205 billion guide. Alphabet's finance chief said outside compute "came online only at the very end of Q3," per Quartz.
That is the pattern. Spend is limited by halls, power and third-party deliveries, not by a sudden drop in AI demand. When the binding constraint is construction timing, a company can miss a CapEx target and still say the book of business is full.
The market split the trade
CNBC's Jenny Lee described what happened the day after the midterms. Microsoft and Amazon each rose more than 6%. CoreWeave fell about 10%. Nvidia fell about 3%. JPMorgan's James Sullivan told CNBC the split reflects a view that "existing capacity is significantly more valuable now because of incremental supply constraints," and that neocloud borrowing spreads have widened far past what investment-grade hyperscalers pay.
Nomura's post-election note, cited in the same piece, said most of the capacity due in 2027 is already under construction. The bigger delay risk sits in 2028 projects that still need permits. That matches what we wrote yesterday: steel in the ground keeps going; greenfield sites wait.
New deal terms, same owners
Two days after the vote, OpenAI published a National Community Compact for future compute campuses it owns and operates. The company says it will pay the full cost of electricity and the infrastructure needed to serve it, skip NDAs with public officials except for narrow security and trade-secret cases, and post financial assurances so ratepayers are not stuck if a project slips. Data Center Dynamics notes the scope: owned and operated. Leased and partner halls may sit outside.
That posture fits the slowdown thesis. Campuses that clear get written under public, Compact-style terms. Many planned ones slip. Hyperscalers can guide CapEx lower and call it prudence. Neoclouds stay in the credit vice we mapped on 3 November.
Then and now
On 5 November we wrote that the scarcest thing in AI is now a permit. That post was about who gets a yes. This one is about what that scarcity does to CapEx. Permission is still scarce. For the companies that already own sites and cash flow, the scarce permit is also cover.
On 22 October we wrote that whoever owns the slow part sets the pace. On 4 November we wrote that flat AI plans were not built for agents that never log off. If new capacity arrives slower than planned, the meter on agent work is not going away, and the owners of live megawatts keep pricing power.
Our call
By 31 March 2027, at least two of Microsoft, Amazon, Alphabet or Meta will cut their calendar-2027 CapEx outlook relative to what they signaled on their Q3 2026 earnings calls, and will cite permitting, power or construction delays as a reason. We're wrong if fewer than two of those four do that, or if a true federal build ban (not a state pause) lands and the CapEx story becomes a demand story instead.