the $1.1 trillion bar no software cohort has ever cleared

the number

morgan stanley quietly raised its 2026 hyperscaler capex forecast for amazon, alphabet, meta, microsoft, and oracle from $765b to $805b, and lifted 2027 from $951b to $1.1 trillion. that 2027 figure is, give or take, equal to the entire non-tech s&p 500 capex from last year. five companies. one year. one technology stack.

this is not a software cycle. software cycles do not require building the equivalent of the rest of corporate america's physical plant. this is closer to the buildout of the interstate highway system, except instead of being financed over thirty years by the federal government, it is being financed over thirty-six months by the operating cash flow of five companies whose investors still expect margin expansion.

the bar

call it the monetization bar. every prior compute platform shift, mainframe, pc, mobile, cloud, eventually generated more revenue than capex required to build it. but those buildouts were paced over a decade or more, and the marginal capex was funded by the marginal revenue. the ai buildout has inverted the order. the capex is landing in 2026 and 2027. the revenue has to show up by, generously, 2028.

the consensus read is that this is fine because ai will be "bigger than cloud." maybe. but cloud capex peaked at roughly $200b/yr across the same cohort, and cloud revenue today is somewhere in the $300b range. the proposed 2027 capex line is roughly 5x peak cloud capex. the implied 2030 ai revenue, just to clear the same payback ratio, is north of $1.5t. that is more than the entire global enterprise software market today.

where the money is actually going

the capex is not abstract. it is power, land, hbm, and nvidia gpus, in roughly that order of constraint. tsmc allocation used to be the binding input. it is now substation queue position, with pjm and ercot interconnect waits stretching four to seven years. microsoft, amazon, and google are signing multi-decade ppas not because they want to be in the energy business but because the alternative is a stranded data center.

and the revenue side is starting to bifurcate in a way the capex models have not absorbed. linkedin's agentic hiring tools just hit a $450m run rate, monetized through model-integrated workflow rather than seat licenses. google and meta are running an ad revenue boom on ai-automated targeting and creative. ai traffic to us retailers was up 393% yoy in q1. these are real dollars. they are also concentrated in two companies that already had distribution.

for everyone else, the value capture story is harder. the nyt's reporting on tech layoffs makes the corollary explicit: ai capex bills are coming due, headcount is the only flexible line item, and meta is targeting may 20 for its first wave. the layoffs are not ai replacing workers. they are workers funding the gpus.

the steelman

the counter is real. ai is showing up in pharma r&d as j&j halving lead generation time. gpt-5.4 is hitting 83% on gdpval and 75% on osworld against a 72.4% human baseline. if you believe the productivity surface is genuinely broad, then $1.1t is not absurd, it is underbuilt. the internet buildout was "overbuilt" in 2001 and looked badly underbuilt by 2010.

the honest version of the bull case is that capex this concentrated only works if ai becomes load-bearing infrastructure for the entire economy, not just a feature shipped inside existing software. that requires the agentic workflow story to actually clear the integration tax, which is the part gdpval scores do not measure.

what to do with this

two things follow. first, the value capture inside ai is going to compress toward whoever owns distribution and whoever owns power. model labs without a distribution partner and infra plays without a power story are the squeezed middle. the q1 funding numbers already reflect this, $300b into ai startups with $152b going to two labs is not a boom, it is a barbell.

second, the capex line is now the dominant variable on every hyperscaler income statement, which means the next two years of big tech earnings will be dominated by the question of whether ai revenue is ramping fast enough to justify the depreciation that is about to hit. the answer is knowable on a quarterly cadence. watch the gap between capex guidance and ai-attributed revenue, not the model benchmarks.

the five hyperscalers are about to spend more in one year than the rest of the s&p 500 ex-tech spends combined. either ai is the largest business in the history of business, or somebody is about to learn what stranded capex feels like at trillion-dollar scale.