a16z: AI Infrastructure Pulls Capital Towards Semiconductors, Power, Networks, Construction, Debt Markets, and Skilled Labor

Andreessen Horowitz released State of Markets II on September 30, authored by David George. The report describes this year's theme as a shift from bits to atoms, with AI infrastructure pulling capital towards semiconductors, power, networks, construction, debt markets, and skilled labor, while paid usage remains low.

The expenditure side has already shown macro numbers. The report states that by late August, technology contributed about 76% of the S&P 500's earnings growth, with high-tech equipment, software, and R&D accounting for about 55% of U.S. capital expenditures. AI construction's share of GDP has surpassed that of railroads in previous years. The five major cloud providers are expected to spend about $780 billion on capital expenditures this year, likely exceeding $1 trillion afterward. In 2024, this is projected to be about $241 billion, and in 2025, about $416 billion. In previous years, they primarily used existing free cash flow, but now they are starting to take on debt, effectively converting cloud providers' cash flow into semiconductor cash flow.

Demand still indicates insufficient supply. The report refutes the view that graphics processors become obsolete in three to four years, stating that the rental prices for A100, H100, H200, and B200 will remain stable or rise by 2026, and older chips retain resale value. The median salary for data center facility managers in the data provided rose from about $82,000 to $134,000, a 64% premium. Construction managers have a premium of about 29%. Cumulative utility capital expenditures are projected to exceed $1 trillion by 2030, but electricity growth does not all come from AI.

Paid penetration does not align with this expenditure. The report states that as of April, about 2% of U.S. households were paying for some form of AI service, a figure that has increased but remains small. A circulated chart indicates that the percentage of paying households in April was 2.2%, up from about 1% a year ago. The average monthly spending for subscription users rose to $31 in May, up from about $22 in May 2024, with an increase over the last four months exceeding the previous twenty months. Menlo Ventures estimates global consumer spending on AI will reach $40 billion by 2026, up from $12 billion a year ago, with the number of users remaining nearly unchanged.

The enterprise side is similarly shallow. The report notes that nearly 30% of S&P 500 companies reported some quantifiable impact from AI, but only about 2% disclosed a continuously tracked metric. Among proxy applications, the proportion of users with truly large-scale deployments remains small. Revenue is concentrated in model companies: the report is cited as stating that the new revenue from OpenAI and Anthropic in 2026 will exceed the total of all publicly listed software companies outside of cloud. The backlog orders for Microsoft, Google, and Amazon's cloud services have roughly doubled year-on-year.

The buyers are cloud providers, new clouds, and model labs, while the sellers are chips, power, networks, and facilities. The driving events are training and inference queues, not household subscriptions. Money flows from cloud providers' free cash flow and new debt into semiconductors and construction sites. Beneficiaries include older graphics processors that have not depreciated, data center labor, and the two major labs. Those under pressure are publicly listed software companies that convert growth into profit, as well as enterprise budgets that have not yet entered workflows and are only mentioned. a16z itself just closed an $1.1 billion Machine Age fund in August, investing only in hardware layers.

The report's original wording indicates it is still early. Graphics processors are already at full capacity, and data on mature adoption and utilization remains in the early stages.

Source: Public Information