a16z Partner David George: Technology is the Cycle of Everything, AI Infrastructure's Share of GDP Just Exceeded the Railroad Era
Andreessen Horowitz General Partner David George released the second edition of "Market Status," stating that technology is the cycle of everything, and the share of AI infrastructure in GDP has just surpassed that of the railroad era.
The report and podcast feature over a hundred charts interpreted by George, Sarah Wang, Alex Immerman, and Santiago Rodriguez. High-tech equipment, software, and R&D account for about 55% of U.S. capital expenditures, while technology represents about 40% of U.S. stock market value, with eight of the top ten global companies being U.S. tech firms. As of late August, technology contributed approximately 76% to S&P 500 earnings growth. Alphabet, Amazon, Meta, Microsoft, and Oracle's capital expenditures are projected to be around $780 billion in 2026, up from $416 billion in 2025, with expectations to exceed $1 trillion annually starting in 2027. Model companies have raised over $350 billion cumulatively. Six private companies are valued at approximately $2.4 trillion based on their latest funding rounds, surpassing all public funding in the past decade. About 75% of listed software companies are profitable. The stock market has risen about 90% since ChatGPT, with earnings increasing about 15%, but multiples have not expanded in sync.
The demand-side diffusion is still early: the median AI supplier spending for the top 1% of companies is about eight times that of the top 10%; approximately 69% of S&P 500 companies have deployed AI, about 30% disclose quantifiable impacts, and only about 2% reach a level where "stopping would be immediately noticeable." The number of proxy users is still only in the millions, but it can cover billions of internet users; Meta and Google have monetized U.S. users at over $200 per year, believing there is more room for proxies. Themes for the next five years include consumer proxies, robotics, autonomous systems, AI and biology, personal health, enterprise diffusion, and a new wave of U.S. vitality. Global infrastructure demand is cited as approximately $90 trillion by 2040. Factories or data centers are being written as products themselves.
Funding is shifting from software subscriptions to chips, power, and construction. Buyers aim to convert free cash flow into semiconductor and power capacity in cloud factories, while sellers are enterprises still in shallow deployments of enterprise software. The event is driven by the annual market report. Beneficiaries are hardware and power companies positioned upstream in the capital expenditure chain; pressured are software stocks that have not integrated models into workflows and cannot report quantifiable impacts. Construction has been priced according to railroad standards, while penetration is still measured in single-digit work replacements.
Source: Public Information
ABAB AI Insight
a16z translates "software is eating the world" to "atoms are back": software still consumes profits, while capital expenditures shift to building factories. George uses the railroad's share of GDP as a benchmark, writing the $780 billion of massive cloud as a new track. The path is for cloud factory profits to turn into semiconductor profits, then leverage further. The six unlisted companies valued at $2.4 trillion indicate that the primary market has priced models, data, and mobility as asset classes, with public listings being just the next round of financing.
Capital is shifting from the light asset software cycle to the heavy asset power cycle. The $1 trillion annual capital expenditure expectation writes the discount rate, grid queuing, and construction timelines as constraints on tech stocks. The motivation is that supply still cannot keep up with demand, hence continuing to treat factories as products. Resources are shifting from selling cloud seats to integrating models into company data: 69% have deployments, 30% can quantify, and 2% cannot afford to stop; the gap in between is the battlefield for enterprise software in the next five years. Proxies that monetize advertising at $200 per user are being rewritten as executable task labor prices.
In contrast to the 19th-century railroad's 60% market value drop, the 2000 fiber optics, and the 2021 unprofitable software bubble. The current narrative is driven by profits rather than multiples: what is rising is profit, not valuation bubble labels. The industry phase is supply expansion and shallow demand diffusion: upstream has surpassed railroad standards, while downstream is still in pilot phases. U.S. vitality is included as a theme, equating defense and re-industrialization into the same capital expenditure table.
The essence is the reconstruction of the industrial chain. The bottleneck of the tech cycle has shifted from marginal costs of code to power and facilities. The mechanism is: concentrated profits allow a few companies to qualify to pour cash flow into atoms; the diffusion gap allows software that connects workflows to still have a five-year window. When factories become products, it means the next round of winners will not be those who write one more line of prompts, but those who first turn power and infrastructure into sellable computing power.
ABAB News · Law of Cognition
Profits are in software, capital expenditures have moved to building factories
Deployment does not equal quantification; only stopping counts as penetration
When factories are products, power is more like a bottleneck than code