The AI Buildout Passed the Railroads, and the Debt Came With It

Two documents landed on September 30 that describe the same boom from opposite ends. Andreessen Horowitz published State of Markets II, a deck of more than 100 charts from growth general partner David George's team, arguing that AI is now the engine of the American economy and that the numbers justify the spending. The same day, the Bank of England's Financial Policy Committee released the record of its September 25 meeting, which spent much of its attention on what happens if those numbers disappoint, and on who is holding the debt when they do.

Read together, they are the clearest picture yet of the AI trade in late 2026: real earnings, real adoption, and a financing structure that is increasingly borrowed.

The bull case, by the chart

a16z's opening claim is scale. High-tech equipment, software and R&D now make up about 55% of all U.S. capital spending, per a chart built on LSEG, Yardeni Research and Bureau of Economic Analysis data. In its own summary of the report on X, the firm says the AI buildout has just passed the railroads as a share of GDP. The deck's text goes further, saying AI infrastructure accounts for all or nearly all of net-new construction spending, blue-collar job growth, durable goods imports and net-new investment-grade issuance, and calling it "arguably the only pro-cyclical impulse firing in the economy right now."

The spending is enormous. a16z's chart, based on Bloomberg data and consensus estimates, has hyperscaler capex approaching $800 billion this year and topping $1 trillion annually from 2027. Crypto Briefing's read of the report puts the 2026 figure at roughly $780 billion, up from $416 billion in 2025.

The firm's answer to the bubble question is earnings. Tech contributed about 76% of the S&P 500's total earnings growth in 2026 as of late August, according to the report. Tech stocks are up more than 20% on the year even as their multiples sit about 20% below the five-year average. "Hard to call it a bubble," one slide is titled, on the grounds that profits rather than prices are doing the work.

Wide, shallow, and very uneven

The more interesting half of the deck is the admission that adoption is broad but thin. Citing Apollo, a16z shows 69% of S&P 500 companies pointing to a live AI deployment in the second quarter, 29% reporting a quantified result, and just 2% disclosing a metric they track over time. About 2% of U.S. households were paying for an AI service as of April.

Spend is small relative to the cost base. A Goldman Sachs chart in the deck puts AI inference at roughly 0.1% of S&P 500 expenses as a share of revenue. a16z notes that cloud reached about 4% of IT budgets in its second year and says most estimates of AI spend sit below that. Only 20% of organizations cite cost as a constraint on AI use, per McKinsey data the firm reproduces. Its reading: enterprise AI budgets have plenty of room to grow.

The usage that does exist is concentrated. YipitData figures show the top 1% of AI spenders outspending the top 10% by roughly eight times. OpenAI Signals data shows output-token use at frontier firms up 17.1 times since April 2025, against 2.1 times for typical firms.

Agents are where the curve bends. On OpenRouter, agentic tokens overtook human ones on February 6, 2026, according to the deck, and the seven-day average for agents has reached 7.3 trillion tokens, a 14-fold increase. More than 85% of that agentic token burn comes from cached prompts, the efficiency that makes long-running agents cheap enough to run. a16z frames this as Jevons Paradox in real time: cheaper intelligence, more demand. The deck notes even the A100 is pricing at or above where it started the year.

The financing problem

The deck does not hide where the money is coming from. Surging profits funded most of the capex so far, but a16z shows hyperscaler free cash flow dropping steeply and expected to stay depressed until around 2028. "With hyperscaler cash fully-tapped, credit markets have sprung to life," one slide reads. The firm's defense is that hyperscaler returns on invested capital remain well above their cost of borrowing.

That is the exact point the Bank of England picked up. The FPC record cites a Morgan Stanley estimate that global AI-related debt issuance hit about $450 billion by early September, more than double the total for all of 2025, and says 2026 issuance is expected to exceed that of countries such as the UK. AI hyperscalers accounted for 47% of sterling corporate bond issuance so far this year. JPMorgan analysts estimate debt-financed AI capex at around $4.1 trillion between 2026 and 2030, and private credit is expected to fund $700 billion of data center capex from 2026 to 2028.

The committee flagged the opacity and, at times, "circular arrangements" in that financing as complicating any risk assessment. It noted that AI equity valuations fell sharply in July as leveraged investors unwound stretched positions, without spilling into core markets. The warning is about the next one. Growth forecasts and fiscal outlooks, the record says, depend partly on expectations of large AI productivity gains, and "a reassessment of those expectations could therefore affect not only AI-related asset valuations but also sovereign debt markets." The FPC judged that the risk of several vulnerabilities crystallizing at once had risen since July, also citing renewed Middle East conflict and higher bond yields, and held the UK countercyclical capital buffer at 2%.

Two readings of one chart

The two documents do not actually disagree on the facts. Both see capex at record levels, both see debt filling the gap left by cash flow, and both tie the outcome to whether AI earns its keep. a16z reads the 2% of companies tracking AI metrics as early innings. A central bank reads the same thinness as the distance between expectations and proof.

For investors, the practical signal sits in the middle of the a16z deck rather than its headline. The bull case depends on the 69% becoming the 29%, and the 29% becoming the 2%, faster than the bond market loses patience. The agent token curve suggests demand is real. The Bank of England's record is a reminder that demand has to show up as cash before the debt comes due.