Reading a16z’s State of Markets II AI’s next chapter is physical.

The report opens by showing how large technology has become in the economy's investment and earnings. High-tech categories, which it defines as high-tech equipment, software, and research and development, account for about 55% of U.S. capital spending, in a series compiled from LSEG, Datastream, Yardeni Research and the Bureau of Economic Analysis through May 5, 2026. That is a reported share of business investment under a broad definition. It is not technology's share of GDP. a16z, p. 5 ↗

Earnings tell a similar story. Using S&P Capital IQ estimates as of August 28, 2026, the deck attributes about 76% of S&P 500 earnings growth for 2026 to its technology universe, and shows the cumulative earnings gap between technology and everything else widening for years. Because 2026 is not finished, that figure depends on estimates, and the universe is a16z's own grouping rather than a standard sector. a16z, p. 7 ↗

The most useful chart in this section explains why technology stocks rose in 2026 even as their valuations fell. According to Duality Research data through August 30, the technology sector's price return of 22.1% came from earnings that contributed 55.7% while the price-to-earnings multiple contracted 21.6%. Those two effects multiply rather than add. The sector's P/E also sat about 19% below its five-year average. a16z, p. 11 ↗

The mechanism matters more than the headline. Durable software franchises and digital distribution have produced very large profits for a handful of companies. Those profits can pay for research, data centers and infrastructure beyond software itself, and the effect spreads through suppliers and customers, which is why sector labels are an imperfect guide to who is exposed. What is new is the combination: established companies with heavy earnings and a fresh appetite for physical capital. That gives this cycle a different financing base from booms led by young companies raising new equity.

Strong earnings do not settle the argument. They can coexist with expectations that run ahead of what the businesses will earn later. Earnings-backed price gains weaken the simple speculation story, but they leave open questions about how long margins last, how much the market already prices in, and how concentrated the gains are. The deck also asserts that AI infrastructure accounts for nearly all incremental activity across several economic categories. That claim needs its own series and definitions before it can carry a headline, so we leave it as a16z's interpretation.

What would change the reading: downward revisions to forward technology earnings, falling margins, a pickup in earnings growth outside technology, or rising dependence among suppliers on a few large customers.

For two decades the largest internet companies were prized for needing little physical capital to grow. The second section of the report shows that changing quickly. Its preferred series, Vanguard calculations from Bloomberg data dated July 31, 2026, tracks combined capital spending at Alphabet, Amazon, Meta, Microsoft and Oracle. Spending rose from $97 billion in 2020 to $416 billion in 2025. The consensus forecast for 2026 is $777 billion, about 86.8% above 2025, and the series projects roughly $1.1 trillion to $1.2 trillion a year from 2027 through 2030. a16z, p. 15 ↗

A few definitions make the rest of the section easier. Operating cash flow is the cash a business generates from its normal activities. Capital spending, or capex, is cash spent on long-lived assets such as servers, buildings and power equipment. Free cash flow is what remains after subtracting capex from operating cash flow, and it is the cash a company can return to shareholders, hold, or use to pay down debt.

The mechanism is simple. The cash generated by search, advertising, cloud and software now pays for accelerators, networking, buildings and power. As capex absorbs more of that cash, free cash flow shrinks, and the deck shows that compression clearly in its forecasts. Debt fills part of the gap. The report shows rising bond issuance by the largest technology companies, and project financing and leases can carry spending further, sometimes off the balance sheet of the company that will use the capacity. Those structures matter for anyone asking who bears the risk if the capacity earns less than planned. a16z, p. 17 ↗ p. 18 ↗

The deck's answer to the return question is a chart of return on invested capital, or ROIC, against borrowing costs and the weighted average cost of capital (WACC, the blended return lenders and shareholders require), from BofA Global Research in August 2026. ROIC sits comfortably above both. The slide's left-hand heading refers to return on equity, but its chart and body use ROIC, so ROIC is the measure we discuss. a16z, p. 19 ↗

That chart answers a narrower question than it first appears to. Historical corporate ROIC blends decades of established, highly profitable businesses with past investments. It does not measure the return on the next data-center campus. The return on that project depends on how long the equipment stays useful, how fully it is used, what customers pay, and how quickly depreciation arrives. The recovery in free cash flow the forecasts show around 2028 is a projection, not an observation.

What would change the reading: actual capex against the forecasts, incremental gross profit compared with depreciation growth, utilization of new capacity, and cash flow after financing and lease commitments.

A token is the unit AI models read and write, roughly a short word or part of one. Prices per token have fallen quickly, and the deck argues that cheaper tokens are expanding total use. Its clearest data point comes from OpenRouter, a service that routes requests to many models: weekly token volume rose from 4.7 trillion in September 2025 to 126.2 trillion in September 2026, about 26.9 times higher. a16z, p. 78 ↗

Two other charts sharpen the point. OpenAI's Signals data, normalized to April 2025 and reported August 12, 2026, show output-token use at frontier firms, the top 10% of users, up 17.1 times across all industries, against 2.1 times at a typical firm in the middle 10%. Use is growing and concentrating in the companies that have rebuilt work around it. a16z, p. 33 ↗

And within OpenRouter's observed activity on August 28, 2026, Anthropic, Google and OpenAI together received about 69% of spending while processing about 35% of tokens. Buyers pay a premium for frontier models on the tasks where quality matters, and route cheaper work elsewhere. That is a picture of one platform's customers, and the remaining labs are not the same as an open-weight category. a16z, p. 36 ↗

The mechanism behind the growth is that lower cost per task makes more tasks worth automating. Agents add to this. They make many model calls per job, keep long documents in context and revisit them, and the deck shows agent-related volume rising sharply alongside a large share of cached prompts, which are reused inputs billed at lower rates. Routing and caching lower the cost of each job while the number and complexity of jobs grows. Falling unit costs and rising total spending can happen together. a16z, p. 35 ↗

The limits are important. Tokens are not uniform: input, output, cached and reasoning tokens are priced and consumed differently across models. A volume chart cannot by itself show revenue, energy use or useful output, and falling token prices next to rising GPU prices cannot pin down the size of an economy-wide rebound effect. The deck's demand story is plausible. It is not yet measured.

GPU rental prices deserve a valuation reader's attention. The deck uses rental rates, and the residual values implied by them, to argue that older chips keep earning. Rental prices are evidence of earning capacity today, under scarcity. They do not settle how fast the equipment depreciates or becomes obsolete once more supply arrives, which depends on hardware improvements, utilization, power costs, contract length and the depth of the secondary market. An appraiser would recognize the problem: a current rent is not a remaining economic life. a16z, p. 26 ↗ p. 81 ↗

What would change the reading: cost per successful task, gross profit per workload, cache-adjusted usage, the quality of agent completions, and compute pricing once new capacity comes online.

If AI makes software cheaper to build and some tasks cheaper to automate, what happens to the companies that sell software? The deck's answer is a careful one. In a selected universe of public software companies tracked by S&P Capital IQ through September 8, 2026, about 75% are profitable under the chart's framework, while only about 30% are growing revenue faster than 20% a year. The universe changes over time, from 148 to 160 software companies and 94 to 101 SaaS companies. a16z, p. 61 ↗

Valuations have fallen much further than the operating numbers. The deck measures software value with EV/TTM revenue: enterprise value (equity value plus debt, minus cash) divided by revenue over the trailing twelve months. Using J.P. Morgan Asset Management data from August 20, 2026, the median vertical software company, the kind built for a single industry, fell from 14.2 times revenue in 2020 to 4.6 times in the first half of 2026, a decline of about 68%. Horizontal software, built for every industry, fell from 12.3 times to 2.7 times, about 78%. a16z, p. 59 ↗

Two forces explain the gap between operations and prices. The first is ordinary finance. Higher borrowing costs pushed software companies to cut losses and run for cash, and slower growth justifies a lower multiple even when revenue keeps rising. The deck's public software revenue and operating-margin series on page 60 look far more resilient than a blanket collapse story would suggest, and Stripe's own September 17 analysis of SaaS platforms on its network ↗ makes a similar case. a16z, p. 60 ↗ p. 83 ↗

The second force is about the future. AI can change how easily a product is replaced, how many people a customer needs, and whether pricing per seat still works when software does part of the job. Page 62 uses the newspaper industry as a warning: share prices began falling well before earnings did, because investors priced the disruption early. The analogy is about timing. It is not a forecast that software will follow newspapers. a16z, p. 62 ↗

Vertical software held up somewhat better than horizontal software, but it still lost about two-thirds of its multiple. Industry focus helps only when it creates something hard to copy: operational value a customer can measure, data a competitor cannot easily assemble, integrations that are costly to replace, or execution a customer trusts with consequential work. Sector averages hide which companies have those things. That question is directly relevant to the software appraisers and lenders use, which we return to in chapter 10.

What would change the reading: net revenue retention, pricing changes, moves from seats to usage-based pricing, customer concentration, inference costs, and the share of product value tied to completed actions and outcomes.

Companies now stay private far longer, and the deck shows how much value has accumulated outside public markets. According to PitchBook-NVCA data as of June 30, 2026, active U.S. unicorns, private companies valued at $1 billion or more, carry an aggregate post-money valuation of $5.34 trillion. The deck sets that beside the Russell 2000 index of smaller public companies, at about $3.5 trillion on April 30, 2026. The comparison is striking but uneven: the dates differ, and a last-round private price is a different kind of number from a public market capitalization. a16z, p. 67 ↗

The fund table is the most useful chart in this section, because it puts three measures side by side. The median 2021 fund shows a net IRR of 1.3% and a TVPI of 1.04 times, which means its holdings are marked at roughly what investors paid in. Its DPI rounds to zero: almost no cash has come back. Paper value and cash realization are separate questions, and this table shows why.

Several valuation bases get mixed in private-market coverage: last-round post-money prices, enterprise value, public market capitalization, tender-offer prices, fund net asset value and distributed cash each answer a different question. Preferred shares often carry protections that common stock does not, and private marks can be stale. An exit value measures a company at an event. It is not the cash investors receive.

The deck also shows how concentrated outcomes are. In PitchBook data for the first half of 2026, covering 197 valued exits, the top 1% of exits accounted for 84% of exit value and the top 10% for 94%. Those groups overlap rather than add up, and a few very large transactions can dominate a six-month window, so we would not read it as a durable change without seeing the underlying deals. Secondary markets and the dispersion between venture managers, on pages 69 to 71, point the same way: investors with access to a few leading companies have a very different experience from those holding the broad market. a16z, p. 68 ↗ p. 69 ↗

What would change the reading: actual distributions to fund investors, the terms of new rounds, the size and coverage of tender offers, time to exit, and whether returns spread beyond the largest companies.

Venture money has moved almost entirely toward AI. J.P. Morgan Asset Management data on page 74 put AI-related companies at 86% of U.S. venture deal value in 2026, up from 65% in 2025. That is a share of dollars, not of deals, and the 2026 figure is a partial-year snapshot. A handful of very large rounds can move it. a16z, p. 74 ↗

Below that, the startup population is more constrained. SVB research on page 75 shows the newest unicorns reaching that valuation younger while the overall cohort of active unicorns gets older, a sign of companies waiting longer for an exit. The same page carries a favorable heading about unicorn profitability that its own chart does not support: the margin bins shown beside it put about a quarter of companies in positive-margin categories, 4% at margins of 25% or more and 21% between zero and 25%. We could not tell whether the heading uses a different measure, so we leave its claim out. a16z, p. 75 ↗

The fastest growth sits in a selected group. For top AI applications, page 76 reports median year-over-year ARR growth of 423% for companies starting between $1 million and $50 million of ARR, and 150% for those starting between $50 million and $200 million. Those figures are proprietary, approximate, and based on inconsistent reporting windows, and the sample is explicitly the leaders. They describe winners, not the typical AI company. a16z, p. 76 ↗

New technology is attracting the large majority of venture dollars, and fundraising itself selects for companies investors believe can keep scaling. So company formation, funding, growth, profitability and survival need to be read separately. A market can be producing record-setting leaders and a strained median at the same time.

What would change the reading: net retention, revenue quality, contribution margin after compute costs, runway, customer acquisition efficiency, and how much of the funding total comes from a few large rounds.

The last part of the deck, pages 77 to 88, draws on companies that see activity directly: OpenRouter, OpenAI, Databricks, Kalshi, inference providers, Stripe, Protege, EliseAI, defense manufacturers, Flock and Whatnot. Platforms observe economic activity close to the transaction, so their data can show change months before official statistics. They can also describe mainly their own customers, their own product adoption or their own fortunes. a16z, p. 77 ↗

A few questions sort the useful from the misleading. Who is in the sample, and who has left? Has the definition changed? How was the measurement collected? Is there a fair comparison group? And is the growth shown the platform's market share or the market itself?

The other cases show the report's breadth: a healthcare figure, policing tools, defense manufacturing and collectibles commerce. Each needs its own framework. The healthcare chart on page 84 supports an association in the cited data, not a claim that AI reduced mortality. Vendor testimonials, earnings-call quotes, usage charts and outcome correlations are all different from independently replicated causal evidence, and it helps to know which one is on the slide. a16z, p. 84 ↗

What would change the reading: growth among a stable set of customers, wider coverage, independent replication, and business outcomes beyond usage and engagement.