The AI Industry: Architecture, Chokepoints, and Deflation (v2.1)

AI disclosure: the research is the author's; this text was drafted with AI assistance and reviewed by the author. The models, and the conflicts they create, are named in the Conflict of interest and scope section. An industry teardown of the artificial intelligence infrastructure buildout, upgrading the qualitative PESTLE scan of v1.0 to a physical chokepoint map, a motivated-claim deflation register and a dated resolution calendar. Version 2.0 is a primary-source rebuild rather than an extension. Its central method choice is to prefer measurement over guidance: combined capital expenditure across Alphabet, Amazon and Meta is reported as 226.0 billion USD of measured cash outflow for the six months to 30 June 2026, taken from the cash flow statements rather than summed from management guidance, with Microsoft's 115.9 billion USD fiscal-year figure held separate because its period boundary is not additive. The paper documents that the same rounded 650 billion USD capital expenditure figure circulates against at least three different cohorts (JPMorgan's hyperscaler estimate, Ofgem's attribution to the big four sourced to Bloomberg, and TrendForce's nine-CSP figure of approximately 830 billion USD), and sets Gartner's AI models and platforms line of 64 billion USD in 2026, an increment of approximately 25 billion USD, against total software spending above 1.4 trillion USD. The chokepoint matrix is rebuilt on primaries: CoWoS advanced packaging sold out through 2026 at 52 to 78 week lead times with customer-side concentration of roughly 60 per cent to a single buyer; ASE Technology capital expenditure raised three times in 2026 to 10.5 billion USD; large power transformer lead times of 128 to 144 weeks from the Wood Mackenzie Q2 2025 survey against up to 60 months for extra-high-voltage units per the US Department of Energy; and the Ras Laffan helium disruption reclassified from a buffered Tier 2 event to a Tier 1 constraint on evidence that repairs are estimated at three to five years. The UK grid queue analysis is grounded in Ofgem's Curate consultation and February 2026 Call for Input, including Ofgem's own implied pipeline capital cost of 693 billion pounds, around 23 per cent of UK GDP in 2025. This version carries three findings that weaken its own thesis rather than burying them: Microsoft is extending reported datacentre and building useful lives from 15 to 25 years from fiscal 2027, lengthening rather than shortening the depreciation clock the overbuild argument relies on; TrendForce reports that HBM per-wafer revenue was overtaken by DDR5 64GB RDIMM in the first quarter of 2026, inverting the margin incentive behind the memory crowding-out mechanism; and the high-purity quartz chokepoint claim is withdrawn because the cited USGS Mineral Commodity Summaries chapter states that world mine production and reserves information was not available. A verification note traces every measured and disclosed figure to a named primary document, with all links resolved on 19 August 2026. Claims that could not be traced are labelled unverified in the same row and flagged at the point of use rather than silently retained. Corrections against the v1.0 record and the unpublished v2.x drafts are enumerated in the method section and include Google Cloud quarterly revenue, a withdrawn Microsoft AI run-rate figure, hyperscaler asset useful lives, the helium disruption date and duration, the turbine count and case number in the Clean Air Act litigation over unpermitted generation, the MOFCOM announcement carrying the 0.1 per cent extraterritorial rare-earth threshold, and the standard governing A-weighted noise measurement. Independent analysis, not investment advice. Version 2.1 adds a reproduction bundle and changes nothing in the analysis. reproduce.py runs offline on the standard library and has three arms: it recomputes every derived figure from the measured inputs in claims.csv, it locates each of those figures in the paper text as printed and reports any that regenerate correctly but appear nowhere, and it reports the resolution status of every source URL from source_verification_log.csv as checked on 19 August 2026. The script exits non-zero on any disagreement. Five derivations are marked bundle-only and are listed as such rather than passed silently. The verification log is shipped rather than summarised: it records, per source, the URL as originally cited, the HTTP result, the verdict, the replacement URL where one was found, the fact as found in the primary, and a recheck horizon. Sixteen of the cited URLs did not resolve or did not support their claim and are recorded with what replaced them; four remain open and are carried in the paper as unverified.