The Bull Case Arrives Late: What a16z State of Markets II Changes, and What It Does Not
Part 2 of Capital Misallocation, Subsidized Compute, and the Anatomy of the Artificial Intelligence Bubble
Four days after the first essay in this series, Andreessen Horowitz published State of Markets II, a hundred-chart deck from the growth team led by David George. It is the cleanest institutional bull case yet for the infrastructure cycle the first essay called unsustainable. It does not kill that case. It raises the spending number, knocks down one accounting claim, and leaves the unit-economics claim untouched.
Read it as a positioned argument. a16z is long the buildout. A deck from the owner of the trade is evidence about the trade, not a census of it.
What the deck actually says
The headline is that technology is no longer a sector. George calls it the everything cycle. By a16z’s count, tech contributed about 76 percent of S&P 500 earnings growth in 2026 through late August. High-tech equipment, software, and research now account for about 55 percent of United States capital spending. The AI buildout, they say, has just passed the nineteenth-century railroads as a share of GDP.
Inside that cycle the rotation is from bits to atoms. Software dominated the last decade. Hardware, power, and networking are the bid now, funded in large part by hyperscaler free cash flow converted into semiconductor free cash flow, and increasingly by debt.
The spending path in the deck is steeper than the one used in Part 1. Hyperscaler capital expenditure runs from about $241 billion in 2024 to $416 billion in 2025 to about $780 billion in 2026, with the five largest hyperscalers heading past $1 trillion in 2027. Demand, on their telling, still outstrips supply at every point in the chain.
The market claim is that this is an earnings story, not a multiple story. Stocks up about 20 percent while multiples are down about 20 percent. The S&P under 20 times earnings. The market up about 90 percent since ChatGPT, on about 15 percent earnings growth.
The adoption claim is that diffusion is early. Sixty-nine percent of the S&P 500 has a live AI deployment. About 30 percent report a quantified result. About 2 percent have AI doing a job they would notice if it stopped. The top 1 percent of companies outspend the top 10 percent on AI by about eight times. Paid consumer penetration was about 2 percent of United States households as of April 2026.
The silicon claim is the one that collides with Part 1. The deck’s heading is that reports of GPU obsolescence have been greatly exaggerated. Rental rates were supposed to fall as new chips arrived. They have not. Pricing on the latest chips is still climbing, and older A100 clusters are still earning, with rental rates at or above where they started the year. Cheaper intelligence, they argue, is raising compute demand rather than retiring the last generation. Jevons, not scrap.
Three revisions to Part 1
First, the capex figure was low. Part 1 used an aggregate of $500 to $527 billion. If the deck’s 2026 path is right, the live number is closer to $780 billion, and the 2027 number is a trillion. That does not shrink the gap between infrastructure and downstream cash. It widens it. A larger build requires a larger terminal revenue base, not a smaller one. The railroad comparison cuts both ways. a16z uses it as scale validation. Part 1 used it as the precedent in which the overbuilders take the write-down and the next operator buys the asset cheap. Both readings survive the chart. The chart does not pick a winner.
Second, the obsolescence claim has to be marked contested, not settled. Part 1 argued that stretching server lives from three or four years to five or six years cuts depreciation in half and inflates reported earnings, and that Blackwell-class throughput would strand H100-class book value. The second half of that argument is a forecast. Current rental markets say the forecast has not arrived. Residual values are holding because utilization is holding. A six-year accounting life can still be too long for a three-year economic life. It cannot be treated as an impairment that has already been recognized. Write the sentence as timing risk, not as a loss already on the books.
Third, the index is not the labs. Part 1 described an economy propped up by speculative balance-sheet allocation. At the S&P level, a16z’s earnings-versus-multiple split is a real objection. If tech is producing three-quarters of the index’s earnings growth and multiples have compressed, the public-market bid is not a pure story stock. That objection stops at the hyperscaler income statement. It does not reach the foundation-model cash ledger. Part 1’s operating ratios, roughly $2.60 to $2.75 of cash expended per dollar collected at the frontier labs, are not in this deck. Neither is the twenty-dollar subscription that can carry five hundred to eight hundred dollars of heavy-user compute, nor the loop in which vendor equity is spent back onto the vendor’s own cloud. An index earnings boom and a lab cash burn can be true in the same quarter. They describe different balance sheets.
What the deck confirms
The enterprise chapter of Part 1 does not need a rewrite. It needs a second clock.
a16z’s own adoption table is broad and shallow. A live deployment at 69 percent of the S&P, a quantified result at 30 percent, a load-bearing job at 2 percent. That sits next to the field finding that 95 percent of enterprise initiatives do not move profit and loss, and that 88 percent of pilots die before production. Same distribution, opposite caption. a16z calls it the early innings of a five-year diffusion. Part 1 calls it a pilot-to-production chasm. The power-user gap decides which caption ages better. If the top 1 percent, already spending eight times the top 10 percent, pulls the median with it, diffusion was early. If the median stays at a pilot, the 2 percent figure was the market.
The financing chart supports Part 1 more than the essay admits. Readings of the deck put capital expenditure growth near 70 percent a year against operating cash flow growth near 23 percent, with capex projected to overtake operating cash flow around the third quarter of 2026. Free cash flow at the hyperscalers is expected to bottom near zero until about 2028, with debt filling the gap and a recovery penciled in for 2029 and 2030. George’s own line is that the build was funded by profits and is increasingly funded by debt. A return on invested capital that still clears the cost of capital, if it holds, is the bull’s answer. It is an answer about the hyperscaler, not about the lab that rents the cluster, and not about the mid-market firm renting the seat.
Token volume is the other bull exhibit. One reading of the deck has OpenRouter weekly tokens moving from about half a trillion at the start of 2025 to well over a hundred trillion by September 2026, with agent-generated tokens passing human-generated tokens. Usage is not imaginary. Usage is also not cash. A token that is subsidized at the seat, or recycled from a vendor’s own equity check, counts in the usage chart and does not count in the amortization chart.
What the deck does not touch
State of Markets II does not price the twenty-dollar tier. It does not reconcile GitHub Copilot’s early loss per seat, the heavy-user compute gap on entry subscriptions, or the concession that uncapped reasoning tiers lose money on the users who actually run them. It does not open the Anthropic or OpenAI cash ledger. It does not trace equity injected by a cloud provider into a lab that is contractually obliged to spend that equity on the same cloud provider. It does not test whether a doubled cloud backlog is third-party demand or affiliated demand.
Those omissions are the load-bearing part of Part 1. A bull case that argues from rental rates, backlog, and index earnings, and is silent on seat-level unit economics, has not answered the piece. It has changed the subject from the application layer to the infrastructure layer. The infrastructure layer can be scarce and still be a bad purchase at the current cost of capital if the scarcity is produced by subsidies.
The open question, and the operator response
The live question the deck leaves on the table is a timing question. Does backlog convert to cash before capital expenditure eats operating cash flow? Watch four series, not the adjective “bubble.”
Capex growth against operating cash flow growth. If the crossover holds and free cash flow stays near zero past 2028, the debt that “stepped in” becomes the constraint.
Older-GPU rental rates. If A100 prices break while Blackwell supply arrives, Part 1’s impairment forecast moves from contested to observed. If they hold, the depreciation critique stays an accounting point, not a write-down.
The power-user gap. If the multiple between the top 1 percent and the top 10 percent closes because the median spends more, diffusion was early. If it closes because the top 1 percent cuts, the build was concentrated demand.
Backlog quality. A doubled backlog that is cancellable, affiliated, or credit-financed is not the same asset as a doubled backlog of arm’s-length, take-or-pay enterprise contracts.
For a mid-market operator between one and twenty-five million in revenue, none of these series is a reason to change the four mandates from Part 1.
Audit the seats before the subsidy ends. The deck does not say the twenty-dollar price survives contact with unsubsidized inference. It says the factories are full. Full factories and cheap seats are different sentences.
Document the workflow before automating it. A 2 percent load-bearing rate inside the S&P is not a recommendation to install a probabilistic tool in an undocumented process.
Tie every integration to a sixty- to ninety-day economic milestone. Quantified impact at 30 percent, and a tracked metric at 2 percent, is the corporate version of the same rule.
Build the operating system a buyer can underwrite. Premium multiples in this tape are being assigned to earnings, not to tool spend. A founder-dependent firm with a stack of subsidized seats is still a founder-dependent firm.
The bull case arrived with better charts than the bear case had last week. It raised the spend, delayed the scrap value of the chips, and showed that the index is being carried by earnings. It did not show that a twenty-dollar seat pays for itself, or that a lab spending more than two dollars to collect one has found operating leverage. Until those two ledgers move, the infrastructure cycle is a larger version of the cycle Part 1 described, not a different one.
