Three companies, one question, told the same way three times. Alphabet's annual report describes investment "in our technical infrastructure, including servers, network equipment, and data centers, to support the growth" of its business (Alphabet Form 10-K, FY2024, filed 2025-02-05), and the FY2025 filing repeats the framing around servers and datacenters (filed 2026-02-05). Microsoft ties its capacity to "demand for fast access to Microsoft services provided by our network of cloud computing and AI infrastructure and datacenters" (Microsoft Form 10-K, FY2024). NVIDIA, on the supply side, frames its platform around AI "model training and inference" (NVIDIA Form 10-K, FY2026).
Follow the money and you hit the same wall at each company: there is no standalone "AI capex" line. The spend lives inside technical-infrastructure investment (Alphabet) or cloud-and-AI infrastructure (Microsoft). That's not evasion; it reflects that the physical datacenters serve AI and non-AI workloads from the same racks. Anyone quoting a precise per-company AI spend number is estimating, then presenting the estimate as disclosure.
So what do the filings settle? Direction and magnitude. Every one of these reports describes infrastructure investment as large and tied to demand growth, year over year. The buildout is disclosed and rising. The debate the filings cannot resolve is the payback — none of them disclose a return attributable specifically to AI, because none of them can cleanly separate it.
The cross-check that does work is structural: read the demand side (Microsoft, Alphabet buying capacity) against the supply side (NVIDIA selling the accelerators that fill it). When the buyers describe rising infrastructure investment and the seller describes rising training-and-inference demand, the two filings corroborate each other. That mutual confirmation is more reliable than any single point estimate.
The honest conclusion for a reader: the capex is real, the labels are fuzzy, and "is it paying off?" is a question the 10-Ks deliberately leave open. The right move is to track total infrastructure spend against total cloud revenue over time — both disclosed — rather than chase an AI-only number that the filings never actually report.