The a16z report reveals the truth about the AI craze: revenue and computing power have hit new highs, but a few leading users and companies are supporting the growth. Read this in-depth analysis to see the truth about layering.
Author: Block Analytics Ltd X Merkle 3s Capital
This article references a16z's "State of Markets II" (September 2026, David George).
About three-quarters of the S&P 500's earnings growth over the past year came from tech companies. AI revenue, compute, and capex are hitting new records almost every quarter. But if you ask the people around you what AI has done for their wages, their businesses, and their accounts, most can't give you an answer. Macro data is this hot, yet personal experience is this cold. How can both things be true at the same time?
The same report, two worlds
Let's look at the hot side first. This 90-page a16z report opens by calling tech the "cycle of everything." The idea is that tech is no longer an industry but the engine of the entire economy. The report argues that AI infrastructure buildout accounts for the vast majority of new construction spending and blue-collar job growth in the United States.
The funders of this buildout are also highly concentrated. The report points out that a handful of hyperscale cloud providers have shouldered the vast majority of capex out of their profits, causing free cash flow to fall sharply and not expected to recover until around 2028. The shortfall is being covered by the bond market. The entire AI infrastructure bill rests mainly on a few giants.
The earnings data is just as hot. The S&P 500 is a stock index composed of the 500 largest listed companies in the U.S. by market cap. According to the report's methodology, as of August 2026, about three-quarters of the S&P 500's earnings growth came from tech companies, whose earnings growth was nearly 4 times that of non-tech companies.
The report also goes out of its way to stress that this rally doesn't look like a bubble. Tech stock valuation multiples are about 20% below their 5-year average, and share price gains have been driven mainly by earnings. Prices were pushed up by profits, not inflated by sentiment.
Retail investors have crowded into the excitement too. Page 13 of the report notes that retail investors remain an important force driving the market, with leverage and chasing momentum amplifying volatility. Unlike the meme-stock craze of years past, this time retail investors are buying a consensus theme. In other words, everyone is buying the same AI story.
Now the cold side. The title of page 27 of the report is "Demand is strong, but it's still early." Nearly 70% of S&P 500 companies say they are using AI. About 30% can report a quantifiable impact. Only about 2% are consistently tracking effect metrics.
The report's assessment of these numbers is just one line: adoption is broad, but shallow. The hot side and the cold side come from the same report, the same authors, the same month. They are not contradictory. The contradiction only arises because we are used to reading them through averages.
Averages lie. Suppose there are 10 people in a room, 1 person earns 10 million yuan a year, and the other 9 earn 100,000 yuan a year. The average annual income in this room is 1.09 million yuan. That number is true, but it cannot describe any real person in the room.
AI is now just such a room. Page 32 of the report puts it bluntly: impressive aggregate data obscures the reality beneath the surface. Growth is driven by a very small number of heavy users. When I read this page, I went back and flipped through the first 30 pages again.
Read from another angle, and every chapter of the entire report is saying the same thing. In the AI era, money and benefits flow to the top first at every layer. This is true at the user layer, the company layer, and the software, venture capital, and startup layers as well. Let's look at this layer by layer below.
AI is hot, but only in a few households.
1% of people use most of the AI
First, let's explain a term. A token is the unit of measurement for text processed by an AI model, and can be roughly understood as the "word count" AI reads and writes. The more tokens a person or company uses, the more work it is having AI do. So token usage is a good indicator of the true intensity of AI use.
Page 32 of the report gives the single most important sentence in the whole text. Growth in token demand is driven almost entirely by the top 1% of users. These people "are only just beginning to touch the edge of what's possible." Note that the report does not say the top half, nor the top 10%, but the top 1%.
The enterprise-layer data turns this sentence into a chart. The report cites a set of enterprise data ranking companies by token usage (based on OpenAI enterprise usage data, April 2025 to June 2026). The top 10% of companies are called "frontier companies," and the companies in the middle 10% are called "typical companies." The baseline is the usage of typical companies in April 2025, recorded as 1x.
By June 2026, output tokens at frontier companies had risen to about 17x, while typical companies were only at about 2x. The gap between them was 8.3x, whereas at the start of the year it had been only 2.6x.
This gap was not this large from the beginning. Before October 2025, the two lines moved roughly together. After October, the frontier companies' curve shot up sharply, while typical companies were almost flat. The report titles this page "Heavy users are pulling away from everyone else."
By industry, the shape is the same. In information, frontier companies rose to 32.5x, while typical companies were only at 2.8x, a gap of 11.7x. In finance and insurance, the gap between the two was 6.7x, and in professional and technical services, the gap was 7.7x. There is no exception in any industry; within every industry, divergence is underway.
Beyond usage, how AI is used is also diverging. Plugins and skill-type tools are extensions that let AI connect to a company's internal systems and repeatedly perform work according to fixed workflows. Among frontier companies, the weekly active adoption rates of these two types of tools are 21% and 19%, respectively. Among typical companies, they are only 9% and 3%. Frontier companies are not just using more; they have wired AI into their processes.
The household-layer data is even colder. In mid-2026, only about 2.2% of U.S. households were paying for AI. Paying households spent an average of about $31 per month. In other words, about 98 out of every 100 U.S. households had not yet spent money on AI. The report's judgment on the consumer side is: everything has only just begun.
So why can token demand still rise into a parabola? The answer is agents. Agents are AI programs that can break down tasks on their own and continuously call models to complete work. As models become cheaper, agents become more cost-effective, so token usage grows exponentially.
The report calls this phenomenon the Jevons paradox: the cheaper and more efficient something becomes, the greater people's total usage of it. But we should add one thing. Those who are putting this batch of cheap intelligence to heavy use are mainly the same heavy users. Intelligence is getting cheaper, but the number of people who know how to use cheap intelligence has not grown in proportion.
The influence of this group of heavy users has already exceeded themselves. The report points out that surging demand is driving up compute prices. According to the rental data cited in the report, even for old GPUs that have been used for 4-5 years, rents have barely fallen compared with a year ago. In other words, the usage of the top 1% is setting the price of compute for the entire market.
Two kinds of quantities appear in this chapter: enterprise token usage and household payment rates. Their definitions differ, so they cannot be added together or converted into each other. But their shapes are the same: a small number of people stand on the steepest part of the curve, while most people are still near the starting point.
My own observation is the same. Many people around me say they are using AI. The number who truly have AI running their processes every day can be counted on one hand.
Most people use AI like a calculator; a small number use AI like hiring a team.
Company layer: saying you're using it doesn't mean you're earning from it
Go back to those three numbers on page 27. Line them up into a funnel, and it becomes clearer. Nearly 70% of companies say they are using AI. About 30% can report a quantifiable impact. Only about 2% are consistently tracking effect metrics.
With each step down, the number of companies drops sharply. "Consistently tracking effect metrics" means the company regularly measures the revenue, cost, or efficiency changes brought by AI, and keeps measuring them over time. Without this step, "using AI" is just a statement and cannot enter the books.
Page 28 of the report uses an entire page to list return cases, with the title "Anecdotal data on AI returns is everywhere." These cases all come from company earnings reports or public statements (company earnings report data, Q2 2026). We picked out the four most representative ones:
Microsoft's Microsoft 365 Copilot paid seats exceed 30 million.
ServiceNow's annual AI contract value surpassed $1 billion in Q2 2026, with agent deployments growing 9x in 9 months.Meta's ad models increased conversations on Facebook by 15%, and automated placement tools drove a 13% incremental lift in purchases.
Amazon and Walmart both say customers using AI shopping assistants spend about 40% more per order.
These numbers are real. But they share one thing in common: they all come from large companies. The list includes Microsoft, Meta, Amazon, Walmart, and ServiceNow. Nearly all the return cases the report can produce come from top-tier large companies. How much a median company earns from AI is nowhere to be found in the report.
So is it money that's holding other companies back? The report argues no. Only about one-fifth of surveyed enterprises believe cost limits AI adoption (based on a survey methodology from an international consulting firm). By the second year of cloud computing's spread, it accounted for about 4% of corporate IT budgets. Today, AI spending's share is still below that level.
The report concludes from this: there is still substantial room for growth in enterprise AI spending. We agree with that direction. But if money isn't the bottleneck, the bottleneck lies elsewhere. Our judgment is that the bottleneck is "knowing how to use it."
Frontier companies have integrated AI into their processes and quantified the results. Typical companies are still stuck at letting employees each give it a try. Budgets can be approved overnight, but usage has to be changed process by process. This is exactly why the funnel narrows the further down you go.
Large companies have a natural advantage here. They have existing data, engineers who can change processes, and sufficiently large business volumes. The same 5% improvement in conversion rate translates into substantial revenue on a large platform. At a small company, it might not even cover the AI bill.
Employment data points in the same direction. The title of page 31 of the report is "AI intensity brings more hiring, especially software engineers." At companies with high AI usage intensity, the share of entry-level positions rose by about 1 percentage point (based on hiring and corporate spending data methodology). At companies with low AI usage intensity, the share of entry-level positions fell by about 0.5 percentage points.
This result contradicts the popular narrative that "AI is taking jobs." But it is equally a form of divergence. Companies that use it deeply are expanding, while companies that use it shallowly are contracting. AI has not affected employment evenly; it has first widened the distance between companies.
For ordinary readers, this chapter has one practical conclusion. Next time you see "Company X's AI revenue hits a record," first check whether it is among the roughly 2% at the bottom of the funnel. For most companies, AI today is still written in the cost column, not the revenue column.
Many companies say they are using AI; few can produce the books.
Software layer: Same label, threefold valuation gap
At the start of 2026, the U.S. software sector experienced a round of selling that the market called the "SaaS apocalypse." SaaS is cloud software that charges subscription fees monthly or annually. The fear at the time was simple: AI would eat these software companies' lunch.
The report titles this chapter "Rather than an apocalypse, it's more like 'prove it to me.'" The meaning is that the market has not abandoned software. The market is simply demanding that every software company provide evidence that it remains useful in the AI era.
Stock prices were the first to show divergence. From January to August 2026, cybersecurity stocks rose 85-90%, while horizontal SaaS stocks fell 28-30% (S&P Capital IQ methodology, as of end-August 2026). Horizontal software refers to general-purpose tools usable across industries, such as office collaboration software. Vertical software is software deeply customized for a specific industry. The report's defensiveness ranking is: cybersecurity stronger than vertical software, vertical software stronger than horizontal software.
Valuation divergence is even more pronounced. Here we need a metric: EV/Revenue multiple. It is the result of dividing enterprise value by revenue, representing how much the market is willing to pay for every $1 of revenue. The higher the multiple, the more bullish the market is on the company's future.
Page 59 of the report divides software into five categories and lists their respective valuation multiples (based on a top asset management firm's methodology, median enterprise value/trailing 12-month revenue multiple, first half of 2026). Cloud, data, and AI infrastructure software is 9.1x. Security and identity software is 6.8x. Vertical software is 4.6x, consumer and transaction platforms are 4.0x. Horizontal software is only 2.7x.
Between the highest and lowest, the gap is more than threefold. Compared with their respective historical peaks, the gap is even larger. Infrastructure software's peak was 17.9x, and it has now fallen by about half. Horizontal software's peak was 12.3x, and it has now fallen by nearly 80%. In the same downturn, some software has merely cooled off, while other software has been repriced.
What mainly determines the multiple is growth. The chart in Section 5 of the report shows that software companies with revenue growth of 20-40% have forward multiples of 9-13x. Companies with growth of 10-20% have forward multiples of only 4-5x. Forward multiples use expected revenue for the next 12 months, which is a different methodology from the group above using trailing 12-month revenue. The two sets of numbers cannot be directly compared in magnitude, but the direction is consistent: the faster the growth, the higher the multiple.
The problem is that high-growth software companies are becoming fewer and fewer. About three-quarters of listed software companies are already profitable, and only about 30% have growth exceeding 20%. The entire industry has traded growth for profit. The report's phrasing is: if you are still a high-growth company, your multiple is "still quite good."
The report also emphasizes that software companies' fundamentals have remained very stable. Revenue growth and operating efficiency have barely deteriorated. So the market is not punishing the "software" industry as such. What the market is punishing is software without growth and without a moat. A moat refers to long-term advantages that make it difficult for competitors to take away customers.
The report leaves an open question at the end of this section. Will software be eliminated by new technology like print media? Or will AI, like a rising tide, lift the returns of all software companies? The report does not draw a conclusion. Our understanding is that both outcomes may occur simultaneously, just landing on different companies.
The top of the software layer is not a few companies, but a few types of companies. One type is software that grows alongside AI infrastructure. The other is security software that becomes more needed the stronger AI becomes. If a software company cannot clearly explain why AI makes it stronger, the market will most likely price it at horizontal software multiples.
The market no longer pays for the word "software"; it pays only for growth.
Venture capital layer: The good-looking returns are pulled up by the top 10%
First, explain a term: power law. A power law is an extremely concentrated distribution in which a very small number of individuals capture the vast majority of outcomes. The venture capital industry has always believed in the power law. A fund invests in 30 companies, and often relies on just one or two of them to earn back all its principal and profit.
The title of page 68 of the report is "The power law persists, and is even more pronounced." This statement holds true at both the company and fund ends. Let's look at the company end first.
The most valuable private companies are larger than ever before. They go public later and later, and the report's phrasing is "skip small caps and become giants directly." More and more value creation is happening in the primary market. By the time these companies go public, the stretch of gains that ordinary investors can participate in has already become shorter.
The secondary market tells the same story. The secondary market is a market where investors transfer shares of private companies or fund interests among themselves. The report points out that there is almost no discount in this market, and the best companies are still trading at a premium. It's not that there isn't enough money; rather, the money is all crowding into the same few names.
Divergence at the fund end is even greater. Here we need two metrics. IRR is the internal rate of return, which converts every cash inflow and outflow during the investment period into an annualized rate of return. A fund's "vintage" refers to the year the fund began investing.
Fund data cited in the report shows that venture funds of the same 2024 vintage differ enormously from one another (based on Carta fund data methodology):
Top 10% of funds: net IRR of about 40%.
Median fund: net IRR of about -3%, slightly below zero.
Funds in the bottom 25%: net IRR of about -15%.
The title of page 71 of the report is: the gap between top 10% funds and the rest "has never been so large." On one side is an annualized return of about 40%, on the other is a median slightly negative. Averaging them together produces a number that does not represent any real fund.
The second metric is more direct. DPI is cash already returned to investors divided by invested principal, measuring money actually received. For funds established after 2021, median DPI is about 0. In other words, more than half of these funds have not yet returned a single dollar of cash to their LPs.
IRR includes paper valuations, while DPI counts only cash received. On paper there are both gains and losses; in cash, the median is zero. The venture capital industry's beautiful returns mostly remain on paper, and belong mainly to top-tier funds.This matter is not far from ordinary people. Many of the limited partners in venture capital funds are pension funds, university endowments, and insurance companies. Paper returns only become cash when a company goes public or is acquired. As mentioned earlier, the best companies are going public later and later, so that cash will also arrive later and later.
There is one more point to clarify. a16z is itself one of the top venture capital funds. This report was written by an institution standing at the very top of the curve. That does not mean the data is wrong; it only means that when we read it, we should remember where the authors stand.
The numbers in this chapter are fund returns. They are a different quantity from the token usage and valuation multiples discussed earlier. But their shape is still the same: the top pulls up the average, and the median is far from the average.
The beautiful average in venture capital is carried by the top 10% of funds.
Startup layer: new companies run fast, old companies are gasping by year two
First, some background. The zero-interest-rate era refers to the years when the United States kept interest rates near zero for a long time, when financing was very cheap. After the zero-interest-rate era ended, startups, like public companies, began trading growth for profit.
But page 73 of the report points out an exception. Companies that recently raised money successfully are still growing at the same pace as at the peak of the zero-interest-rate era, or even faster. The reason is simple: without high growth, they cannot raise another round.
Growth is concentrated in the youngest companies. Payment data cited in the report (based on Stripe payment data) shows that SaaS companies less than 1 year old have revenue growth above 500%, and at the peak it once exceeded 600%. For companies more than 1 year old, growth fell from about 30% to about 20%, then rebounded to about 20-25%. SaaS refers to subscription software charged monthly or annually.
The report uses this set of data to rebut the "SaaS apocalypse": software revenue actually accelerated in the new year. That judgment is correct. But the same chart also shows that the highest growth belongs to newly founded companies. Once a company survives its first year, its growth curve often turns from a steep slope into a gentle one.
Page 75 of the report describes the new generation of unicorns this way: bigger, faster, younger. A unicorn refers to a private company valued at more than $1 billion. The new generation of AI companies has achieved larger revenue in less time and with fewer people.
The situation of existing unicorns is completely different. Data cited in the report (based on Silicon Valley Bank data, 2026 U.S. unicorns) shows that about 40% of unicorns have revenue growth between 0-20%. About 15% are in negative growth. More than half of the companies have less than 2 years of cash runway. Cash runway refers to how long the cash on the books can last at the current burn rate.
Less than 2 years of runway means these companies will soon need to raise again. A financing round usually takes several months from launch to funds arriving. If growth is only 0-20%, they will find it hard to present the growth story investors want to see.
Financing is also concentrating into a small number of companies. Page 74 of the report summarizes this cycle in one sentence: fewer funds, writing bigger checks, investing in fewer companies. These companies are increasingly concentrated in Silicon Valley, and almost all are related to AI.
The report also points out one detail. The industries money flows into are changing. In addition to enterprise software and consumer applications, service industries such as law and accounting, as well as asset-heavy industries such as semiconductors, power, and defense, are also beginning to receive venture capital. But a broader set of tracks has not changed the direction of concentration.
So the divergence at the startup layer is not only between AI companies and non-AI companies, but also between new companies and old companies. New companies are taking off at the top. Among old unicorns, a considerable number are racing against time.
There are two kinds of unicorns: one is taking off, and the other is counting how many months of cash are left.
Next time you see good AI news, ask one question first
Let us look at all five layers together. At the enterprise layer, token output at frontier enterprises is 8.3 times that of typical enterprises. At the company layer, nearly 70% are using AI, and about 2% are continuously tracking results.
At the software layer, the valuation multiple of infrastructure software is more than three times that of horizontal software. At the venture capital layer, the return gap between top funds and median funds has set a record. At the startup layer, new companies are growing above 500%, while more than half of old unicorns have less than 2 years of cash runway. All five layers describe the same shape.
Token usage, valuation multiples, and fund returns are three completely different quantities. We do not add them together or convert them into one another. Different calibers, same shape: at every layer, a small number of top players carry the whole.
Some will say that top-heavy concentration is only an early-stage phenomenon. The report itself says this: others will eventually catch up, or be left behind. We agree that this is possible. Both the internet and cloud computing went through a path from the top spreading to the masses.
But "eventually" has no date. Before others catch up, whether a number describes the top or the majority leads to completely different conclusions. So we believe the power law is not a list of opportunities, but a ruler. It does not tell you what to buy; it only helps you measure clearly how far a piece of good news is from you.
Next time you see good news about AI, you can first use this ruler to ask three questions:
Does this number describe the top, or the median? The top 1% of users and the top 10% of funds do not represent most people.
Does this number describe "using" or "earning"? Nearly 70% of companies are using AI, but only about 2% can show the books.
Does this number describe paper value, or cash in hand? Fund paper returns look good, but the median DPI is about 0.
These three questions require no professional background. For example, suppose a news story says "AI boosts enterprise efficiency by 30%." First question: does it count frontier enterprises, or all enterprises? Second question: does it measure the hours employees are using it, or the extra money the company is making? Third question: has this money already entered the income statement, or is it still a forecast?
We do not give any buy or sell conclusions. If a number's answers to all three questions are "top," "using," and "paper," then it is good news about a small number of people. If the answers are "median," "earning," and "cash," then it is truly relevant to most people.
For good news about AI, ask one question first: is it talking about the top 1%, or you?