Author: Block Analytics Ltd X Merkle 3s Capital. This article draws on a16z’s State of Markets II (September 2026, David George).
Over the past year, roughly three-quarters of earnings growth in U.S. equities came from technology companies. AI revenue, compute demand and capital spending have kept hitting fresh highs almost every quarter. Yet for many people, the effect on wages, business income or personal accounts remains hard to identify. The analysis from Block Analytics Ltd X Merkle 3s Capital argues that these two realities can coexist because averages are masking a sharp concentration of gains.
One report, two very different pictures
a16z opens its 90-page report by calling technology an “everything cycle,” framing it not as a single sector but as the engine of the broader economy. The report says AI infrastructure buildout has accounted for most of the increase in U.S. construction spending and blue-collar job growth.
Funding for that buildout is concentrated too. According to the report, a handful of hyperscale cloud providers have absorbed most of the capital expenditure with their own profits, pushing free cash flow sharply lower and delaying recovery until around 2028. The remaining funding gap is being filled by the bond market. In practical terms, the AI infrastructure bill is sitting on the balance sheets of a small number of giants.
Earnings data tell the same hot story. Using the report’s methodology, technology companies accounted for about three-quarters of S&P 500 earnings growth through August 2026, and profit growth in tech was nearly four times that of non-tech companies. The report also stresses that this rally does not look like a classic bubble. Valuation multiples for tech stocks are about 20% below their five-year average, and the rise in share prices has been driven mainly by profits rather than sentiment.
Retail investors are part of the move as well. Page 13 of the report says retail remains an important force in the market, with leverage and momentum buying amplifying volatility. Unlike the meme-stock phase, this time retail investors are buying into a consensus theme: the same AI story.
Then the report turns and shows a colder picture. Page 27 carries the line “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 measurable impact. Only around 2% are consistently tracking outcome metrics. The report’s summary of that set of numbers is blunt: adoption is broad, but shallow.
Those hot and cold readings come from the same report, the same authors and the same month. The analysis argues that the contradiction is not in the data. It is in the habit of reading them through averages.
The top 1% of users are driving most of the real AI usage
Page 32 contains what the analysis calls the most important sentence in the report: growth in token demand is being driven almost entirely by the top 1% of users. A token is the unit used to measure how much text an AI model processes. In rough terms, more tokens mean more work is being handed to AI, so token consumption is a useful proxy for actual usage intensity.
The report is not talking about the top half of users or even the top 10%. It is talking about the top 1%. That matters because it suggests the headline growth in AI demand is being pulled upward by a very small group of heavy users.
Enterprise data make the gap visible. The report cites OpenAI enterprise usage data covering April 2025 through June 2026 and ranks companies by token consumption. The top 10% are labeled “frontier enterprises,” while the middle 10% are labeled “typical enterprises.” Usage by the typical enterprise in April 2025 is set as the 1x baseline.
By June 2026, output token usage at frontier enterprises had climbed to about 17x, while typical enterprises were only at roughly 2x. That left an 8.3x gap, up from 2.6x earlier in the year. Before October 2025, the two lines moved in a similar way. After that, the frontier line turned sharply upward while the typical line flattened. The title on that page says heavy users are pulling away from everyone else.
The same shape appears across industries. In information services, frontier enterprises reached 32.5x while typical enterprises were at 2.8x, an 11.7x gap. The spread was 6.7x in finance and insurance, and 7.7x in professional and technical services. The analysis says there are no sector exceptions. Each industry is splitting internally.
The divide is not only about volume. It is also about how AI is being used. Plug-ins and skills tools let AI connect to internal systems and repeat tasks through fixed workflows. Among frontier enterprises, weekly active usage shares for those two categories were 21% and 19%. Among typical enterprises, they were just 9% and 3%. The point is simple: frontier companies are not merely using more AI. They are wiring it into operations.
Household data look much colder. By mid-2026, only about 2.2% of U.S. households were paying for AI, and paying households were spending around $31 a month on average. Put differently, about 98 out of every 100 U.S. households had not paid for AI at all. The report’s view on the consumer side is that everything is still at a very early stage.
So why is token demand still rising in a near-parabolic way? The report points to agents. These are AI programs that can break down tasks and call models repeatedly to complete work. As model costs fall, agents become more economical, and token usage rises with them. The report describes this through Jevons paradox: when something becomes cheaper and more efficient, total consumption can increase rather than fall.
The analysis adds an important qualifier. The users deploying those cheaper agents at scale are still mostly the same heavy users. Intelligence is getting cheaper, but the number of people who know how to use cheap intelligence effectively is not rising at the same pace.
The effect now extends beyond those users themselves. The report says surging demand is pushing up compute prices. Based on rental data cited in the report, even older GPUs that have been in use for four to five years have seen little to no decline in rental prices from a year earlier. In that sense, the top 1% of usage is helping set compute pricing for the wider market.
Enterprise token usage and household payment rates are different measures and cannot be added together or converted into each other. Still, they show the same pattern: a small group is standing on the steepest part of the curve, while most people remain near the starting line.
Company adoption does not automatically mean company profits
The three numbers on page 27 look clearer when arranged as a funnel. Nearly 70% of companies say they use AI. About 30% can point to a measurable impact. Only around 2% consistently track outcome metrics. At each step down the funnel, the number of companies drops sharply.
Consistently tracking outcome metrics means regularly measuring how AI changes revenue, costs or efficiency and continuing to do so over time. Without that step, “we use AI” remains a statement rather than an accounting entry.
Page 28 lists a full page of return examples under the heading “anecdotal evidence of AI ROI is everywhere.” These examples come from company filings or public statements, using second-quarter 2026 company reporting. The analysis highlights four of them:
- Microsoft said paid seats for Microsoft 365 Copilot had surpassed 30 million.
- ServiceNow said annual contract value for AI topped $1 billion in the second quarter of 2026, while agent deployments rose 9x over nine months.
- Meta said its ad models lifted conversations on Facebook by 15%, and automated ad tools generated 13% incremental purchases.
- Amazon and Walmart both said customers using AI shopping assistants spent about 40% more per order.
Those figures are real, but they all come from large companies: Microsoft, Meta, Amazon, Walmart and ServiceNow. The analysis notes that the report’s clearest ROI examples are concentrated among major firms, while it offers no clear answer to how much a median company is earning from AI.
Is cost the main obstacle for everyone else? The report says no. Based on a survey from an international consulting firm, only about one-fifth of respondents said cost was limiting AI adoption. It also notes that cloud computing represented about 4% of enterprise IT budgets by its second year of adoption, while AI spending today is still below that level.
From that, the report concludes that enterprise AI spending still has substantial room to grow. The analysis agrees with the direction of that argument but says the bottleneck may lie elsewhere if money is not the main issue. Its view is that the real constraint is whether companies know how to use AI well. Frontier enterprises have integrated AI into workflows and measured the results. Typical enterprises are still letting employees experiment on their own. Budgets can be approved quickly. Workflow redesign takes much longer.
Large companies have structural advantages here. They already have data, engineers who can change processes, and enough business volume to make small improvements matter. A 5% lift in conversion can produce meaningful revenue on a large platform. In a smaller company, it may not even cover the AI bill.
Employment data point in the same direction. Page 31 is titled “AI intensity drives more hiring, especially software engineers.” In companies with high AI intensity, the share of entry-level roles rose by about 1 percentage point. In companies with low AI intensity, the share of entry-level roles fell by about 0.5 percentage point. That runs against the popular line that AI is simply taking jobs away, but it still describes a split: companies using AI deeply are expanding, while those using it lightly are shrinking.
The practical takeaway in the analysis is straightforward. When a headline says a company’s AI revenue hit a record, the first question is whether that company belongs to the roughly 2% at the bottom of the funnel. For most companies, AI still sits more clearly in the cost column than in the revenue column.
Software is not being priced as one category anymore
At the start of 2026, U.S. software stocks went through a selloff that the market dubbed “SaaS doom.” SaaS refers to cloud software sold through monthly or annual subscriptions. The fear was simple: AI might eat into the business models of software vendors.
The report frames this chapter differently. Its title is closer to “show me” than “doom.” The message is that the market has not abandoned software. It is asking each company to prove it still matters in an AI era.
Price action showed the split first. From January through August 2026, cybersecurity stocks rose 85% to 90%, while horizontal SaaS stocks fell 28% to 30%, based on S&P Capital IQ data through the end of August 2026. Horizontal software refers to general-purpose tools used across industries, such as workplace collaboration software. Vertical software is built deeply for a specific industry. The report’s ranking of defensiveness is clear: cybersecurity above vertical software, and vertical software above horizontal software.
Valuation dispersion is even sharper. On page 59, using data from a major asset manager, the report lists median enterprise value to trailing 12-month revenue multiples for the first half of 2026:
- Cloud, data and AI infrastructure software: 9.1x
- Security and identity software: 6.8x
- Vertical software: 4.6x
- Consumer and transaction platforms: 4.0x
- Horizontal software: 2.7x
The spread between the highest and lowest categories is more than threefold. Compared with prior peaks, the gap looks even starker. Infrastructure software peaked at 17.9x and is now down by about half. Horizontal software peaked at 12.3x and is now down by nearly 80%. In the analysis, that means some software categories have merely cooled off, while others have been repriced.
Growth remains the main driver of multiples. A chart in section 5 shows that software companies growing revenue at 20% to 40% trade at forward multiples of 9x to 13x, while those growing at 10% to 20% trade at only 4x to 5x. Those forward multiples use expected revenue over the next 12 months, so they are not directly comparable with trailing multiples. The direction is the same, though: faster growth commands higher valuations.
The problem is that fewer software companies still have that kind of growth. About three-quarters of listed software companies are already profitable, but only around 30% are growing faster than 20%. The sector has traded growth for profits. The report says that if a company is still a high-growth name, its multiple remains “quite good.”
It also says software fundamentals have stayed relatively stable, with little deterioration in revenue growth or operating efficiency. So the market is not punishing software as a whole. It is punishing software companies that have neither growth nor a moat.
The report leaves one open question at the end of the section: will software be displaced by new technology the way print media was, or will AI lift returns across the whole sector like a rising tide? It does not settle the issue. The analysis reads the answer differently: both outcomes may happen at once, just to different companies. The market is no longer paying for the word “software” by itself. It is paying for growth.
Venture returns look strong mainly because the top decile is pulling up the average
Page 68 says “power laws persist, and even more so.” A power law is an extremely concentrated distribution in which a very small number of outcomes account for most of the total result. Venture capital has long operated on that logic. A fund may invest in 30 companies and still rely on one or two winners to generate the entire return.
On the company side, the most valuable private companies are larger than ever and are staying private longer. The report says they are “skipping small cap and going straight to giant.” More value creation is happening in private markets, which means the portion of upside available to public investors is getting shorter by the time these companies list.
The secondary market tells the same story. This is the market where investors trade stakes in private companies or fund interests with each other. According to the report, there is almost no discounting there, and the best companies still trade at premiums. The issue is not a lack of money. It is that capital is crowding into the same few names.
Fund-level dispersion is even wider. Using Carta fund data, the report says 2024 vintage venture funds are showing huge differences from one another:
- Top 10% of funds: net IRR of about 40%
- Median fund: net IRR of about -3%
- Bottom 25% of funds: net IRR of about -15%
Page 71 says the gap between the top 10% of funds and the rest has “never been wider.” One side is producing annualized returns near 40%. The median is slightly negative. Averaging those outcomes does not describe any real fund.
A second metric makes the point more directly. DPI, or distributions to paid-in capital, measures actual cash returned to investors divided by invested capital. For funds launched after 2021, the median DPI is about 0. In other words, more than half of those funds have not yet returned any cash to limited partners.
IRR includes paper valuations. DPI counts only realized cash. The analysis argues that much of venture’s attractive performance still exists on paper, and that paper performance is concentrated among top funds.
This is not far removed from ordinary savers. Venture fund investors often include pension funds, university endowments and insurance companies. For paper gains to become cash, portfolio companies need to go public or get acquired. As the report has already shown, the best companies are listing later, so the cash comes later too.
The analysis also notes that a16z itself is one of the leading venture firms. That does not make the data wrong, but it does matter when reading a report written from the top of the curve.
Startups are splitting between very young winners and aging unicorns under pressure
Page 73 highlights an exception to the post-zero-rate environment. After the end of near-zero interest rates, startups, like public companies, broadly shifted from growth toward profits. But companies that have recently managed to raise money are still growing as fast as, or faster than, they did at the peak of the zero-rate era. The reason is simple: without high growth, they do not get the next round.
That growth is concentrated in the youngest companies. Using Stripe payment data, the report says SaaS companies less than one year old are growing revenue by more than 500%, with peaks above 600%. Companies older than one year saw growth fall from about 30% to around 20%, then recover to roughly 20% to 25%. The report uses that chart to push back on the “SaaS doom” narrative, saying software revenue accelerated again in the new year. But the same chart also shows where the fastest growth sits: in the newest companies.
The analysis puts it this way: once a company gets past its first year, the growth curve often shifts from a steep slope to a gentler one.
Page 75 describes the new generation of unicorns as bigger, faster and younger. A unicorn is a private company valued at $1 billion or more. The report says the new AI cohort is reaching larger revenue numbers in less time and with fewer people.
Older unicorns are in a very different position. Using Silicon Valley Bank data on U.S. unicorns in 2026, the report says about 40% are growing revenue at 0% to 20%, about 15% are shrinking, and more than half have less than two years of cash runway. Cash runway measures how long a company can keep operating at its current burn rate before running out of money.
Less than two years of runway means many of these companies will need to raise again soon. A financing round usually takes months from launch to close. If growth is only 0% to 20%, it becomes much harder to present the kind of story investors want to hear.
Funding itself is becoming more concentrated. Page 74 sums up the cycle in one line: fewer funds are writing larger checks into fewer companies. Those companies are increasingly concentrated in Silicon Valley, and almost all of them are tied to AI.
The report also notes that the sector mix is changing. In addition to enterprise software and consumer applications, venture money is now reaching legal and accounting services, as well as capital-intensive industries such as semiconductors, power and defense. The field is broader, but the direction of concentration has not changed.
That means the startup split is not only between AI and non-AI companies. It is also between new companies and older ones. New companies are taking off at the top end. A meaningful share of older unicorns are racing against time.
The same structure appears across five layers of the market
The analysis closes by putting five layers side by side. In enterprise usage, frontier companies produce 8.3x the tokens of typical companies. In company adoption, nearly 70% say they use AI, but only about 2% consistently track outcomes. In software, infrastructure names trade at more than three times the revenue multiple of horizontal software. In venture, the gap between top funds and the median has reached a record. In startups, new companies are growing above 500%, while more than half of older unicorns have less than two years of runway.
Token usage, valuation multiples and fund returns are completely different measures. They should not be added together or converted into one another. But they share the same shape: at each layer, a small group at the top is carrying the aggregate number.
The report itself says this concentration may be an early-stage feature and that others may eventually catch up or be left behind. The analysis accepts that possibility and notes that both the internet and cloud computing spread from leaders to the mainstream over time. The problem is that “eventually” has no date attached to it. Before that diffusion happens, the answer changes dramatically depending on whether a number describes the leaders or the majority.
That is why the article proposes three simple questions for reading AI headlines. Does the number describe the top end or the median? Does it describe usage or earnings? Does it describe paper value or cash in hand? The article gives an example: if a headline says AI lifted enterprise efficiency by 30%, the next questions should be whether the sample covers frontier companies or all companies, whether the metric is employee usage time or actual profit, and whether the gain has already reached the income statement or remains a projection.
The piece does not offer a trading call. Its framework is narrower than that. If a number points to the top end, to usage, and to paper gains, it is mostly good news for a minority. If it points to the median, to actual earnings, and to realized cash, then it is much closer to the experience of the majority. The final question it leaves readers with is direct: when AI headlines look strong, are they talking about the top 1%, or about you?


