Author: Shou Ziwei, investor internal reference
Recently, the a16z Growth team released the second edition of "State of Markets II". The report reviews the market status in 2026 (at least the first two quarters) from the perspective of the stock market and technology, and gives an outlook for the future.
Its core content can be summarized as follows: technology is no longer just a sector, but an "all cycle" for capital market profit growth; within technology, the market focus has shifted from software to hardware and infrastructure; the GPU obsolescence theory has been weakened by reality; the software industry has not reached its end, but has entered the stage of "proving itself"; AI adoption is still in its early stages.
Technology has become “all cycles”, a high-growth, capital-light alternative
The core judgment of the report is that technology has transformed from a "workshop industry" into the main driving force of the global economy. The report said that every company today is in some sense a technology company, and it is difficult to find a business that truly operates only on pen and paper. Technology drives investment, R&D cycles, profit growth, and profit margin expansion more than any other industry; most of the world's largest, most profitable, and fastest-growing companies are technology companies.
The report points out that after the global financial crisis, technology really took off as a capital market force. In the wake of the housing crash and credit crunch, technology offers high-growth, capital-light alternatives to asset-wary investors.
What's more, technology has huge built-in operating leverage because the serviceable market is far from saturated and the marginal cost of software is close to zero. In hindsight, those who doubted the ability of money-losing tech companies to grow into high-margin machines missed the theme of the previous decade.
It’s true that software is eating the world, but the centrality of technology in capital markets has reached another level. Technology profits continue to grow at a compounded rate. Although the base is low, it is more obvious than other fields. Since 2023, technology has been almost the earnings growth story: as of the end of August 2026, technology contributed about 76% of the S&P 500's total earnings growth.
The report believes that technology can no longer be referred to as just one sector. Durable goods used to define cycles—houses, dishwashers, cars, etc. Now technology has replaced them. Technology is everywhere and has become the “everything cycle.”
Software dominated the last technology cycle, but now hardware has become the protagonist
While technology has become an “everything cycle”, the internal structure of technology is also changing.
The theme for the year and beyond is the shift from bits to atoms, the report said. Software dominated the last technology cycle, and now hardware has become the protagonist.
AI construction drives a surge in demand in traditionally slow-growing, cyclical, capital-intensive industries such as semiconductors, power, and networking. These needs are largely financed by the historic profits of the world's largest technology companies, functionally converting the free cash flow of hyperscale cloud vendors into semiconductor free cash flow; at the same time, they are increasingly relying on debt.
But it’s not just AI. Global infrastructure needs are measured in trillions of dollars, defense spending is rising, the power grid is grappling with the electrification of everything, manufacturing is reshoring, and robots and self-driving taxis are on the horizon.
Hardware and infrastructure have become the focus of the market after years of lagging behind software, the report said. Public and private equity are investing in computing, storage, power, robotics, manufacturing and defense at an intensity not seen in decades, if not longer. In the words of the report: The Atom is back.
AI is an intergenerational platform shift with unprecedented scale of capital expenditures
The report describes AI as a generational platform shift. Capital expenditures by hyperscale companies will rise from approximately US$97 billion in 2020 to US$241 billion in 2024 and US$416 billion in 2025; it is expected to be US$777 billion in 2026 and exceed US$1 trillion per year from 2027. The report states that each round of computing cycles is larger than the last round: from mainframes, minicomputers, PCs, desktop Internet, cloud and mobile to the AI era, the scale of equipment, users and capital expenditures have all increased by an order of magnitude.
The report pointed out that the market's forecasts for capital expenditures are low every year, and demand and prices continue to push up investment. The ratio of capital expenditure to GDP has been close to or exceeded the historical cycle peaks of railways, oil and gas, telecommunications, etc. The free cash flow of very large companies is being swallowed up by capital expenditures, and the report predicts that this pressure will continue until around 2028. As cash is fully mobilized, the debt market begins to step in: AI-related borrowings rise significantly in 2026 and have longer maturities. The report also stated that the ROIC of very large enterprises is still higher than the cost of debt and WACC, which makes their borrowing and investment still have a financial basis at present.
Capital spending also spills over into the real economy. According to the report, data center construction has increased hiring in construction and skilled trades by more than 300,000 people, and has brought about wage premiums: Data center positions such as facilities managers, construction managers, and network engineers have higher salaries than similar non-data center positions.
Capital spending is producing results
According to the report, AI revenue is growing at an unprecedented speed and scale. The combined ARR of OpenAI and Anthropic will rise rapidly from the fourth quarter of 2024, and is expected to be close to US$150 billion by the third quarter of 2026. The report shows that the revenue of OpenAI and Anthropic will be approximately US$100 billion in 2026, exceeding the revenue of public software (excluding very large-scale enterprises) of approximately US$63 billion. Cloud revenue backlog has doubled year over year, and hyperscaler free cash flow is expected to recover significantly in 2029-2030.
The report also mentioned that Neoclouds are growing rapidly and companies such as CoreWeave, Nebius, Applied Digital and others have steep revenue curves. Regarding concerns about GPU depreciation and obsolescence, the report states that the rental and residual values of old GPUs have not collapsed, but have remained stable or increased: The rental price and residual value data of B200, H200, H100, and A100 support this judgment.
At the adoption level, the report believes that AI is still in its early days. 69% of companies in the S&P 500 say they have real-time AI deployed, but only 2% disclose the metrics they track and 0% break out and track the metrics individually. There is still plenty of room for enterprise AI spending: only 20% of organizations limit their use of AI due to costs associated with it; most industries expect to increase AI spending. With AI inference spending currently accounting for only about 0.3% of S&P 500 revenue and cloud reaching about 4% of IT budgets in its second year, the TAM for AI is larger than IT budgets, the report said. On the consumer side, paid AI subscriptions have grown approximately 5 times since 2025, and average monthly household spending has increased, but search recommendation traffic remains small.
The report also uses mobile Internet as a reference: semiconductors will lead first, followed by infrastructure, and then software and services. The report suggests that the AI version may present a similar sequence.
The theory of GPU obsolescence is exaggerated
Directly related to hardware and capital expenditures are computing power supply and demand and the GPU life cycle. The report notes that computing demand still exceeds supply. Previously, well-known bears questioned whether GPUs would become obsolete in three to four years and whether it was worth investing so much money; what does the B200's sales mean for the A100 deployed one or two years ago? At least for now, the AI computing power demand curve is shifting upward, and the A100 is still quite useful.
Typically, rental rates and GPU residual values decrease over time, but this is not the case. As smartphones become cheaper and demand for computing power rises, prices for the latest chips are rising, while prices for older chips remain firm. A100 is priced at or above year-end levels. The report believes that the story is far from over, but the computational and modeling advances so far are not zero-sum. Better, cheaper smartphones are accumulating value across the ecosystem, and older chips and models retain considerable value beyond the “expiration date” set by the bears.
Meanwhile, AI adoption remains relatively immature: widespread, but shallow. Nearly 30% of S&P 500 companies report some “quantifiable impact” from AI, but only about 2% report any trackable metrics. Proxy-based use cases are deployed at scale for only a very small number of users. On the consumer side, as of April, only about 2% of U.S. households paid for some kind of AI service; now that number is higher and growing, but the overall figure is still small. The report concludes that GPUs are already running hot, while the data shows that mature AI adoption and utilization is still early.
SaaS: Prove yourself, not doom
Returning from hardware to software, at the beginning of the year there were voices claiming that software is dead, that AI can generate everything with "atmosphere programming", and that the end of SaaS is coming. The reality is often more nuanced, the report said. The sell-off did occur, but it was more discriminating than “doomsday” suggests; AI or AI threats play a role, but are not the only factor.
A partial revaluation of publicly traded software companies is long overdue. After the zero interest rate policy ended, technology companies and others exchanged growth for profits. In 2022, the market will be filled with high-growth, mostly unprofitable software companies; by 2026, the story will be reversed: about 75% are profitable, but only about 30% have grown by more than 20%. High interest rates are meant to make capital relatively scarce, so companies would be wise to move away from loss-making growth toward more self-sustainable, slower but steady growth. This in itself makes sense, but slow-growth companies don't earn high-growth valuation multiples for long, and the new "slow-growth" normal eventually catches up with software. This is not the case for all companies: fast-growing companies are still trading at multiples in line with historical averages, albeit not at peak zero rates; but there are fewer such companies and the sector is being re-rated as a whole. Summary of the report: Software is not doomed, but it does have to “prove itself.”
Private markets: Value accumulates on the private side
The report believes that the private equity market is accumulating historic value. Year to date in 2026, venture capital financing has been approximately US$72 billion, and venture capital-backed exit value has been approximately US$2.188 trillion. The largest private companies are bigger than ever: SpaceX, Anthropic, OpenAI, Databricks, Stripe, Revolut, Anduril, Cursor, Ramp, Waymo and others have combined valuations in the trillions of dollars. The top five private companies are already worth more than the combined value of technology IPOs over the past decade, the report said.
The report calls this phenomenon "Private-for-longer", that is, companies go public later and more value is accumulated during the private placement stage. The typical IPO time has extended from about 3 years from 1999 to 2005 to more than 10 years from 2020 to 2025. The total valuation of active U.S. unicorns reached US$5.34 trillion, surpassing Russell 2000’s US$3.5 trillion. The power law effect on exit returns is also stronger: the top 1% of exits account for 84% of exit value, and the top 10% account for 94%.
The secondary market also reflects this trend. The report stated that the employee participation rate in the transfer of old shares was less than 60%, but the subscription rate reached a historical high; the discounts for the transfer of old shares were few and small; the median secondary market price was 0% discount compared to the previous round, the 75th percentile was a 23% premium, and the 90th percentile was a 79% premium. The top VC portfolios have grown in value far more than the top public companies: Between 2017 and 2026, the Top 30 VC portfolio index rose from 100 to 2,333, while the Top 10 public companies only rose to 542. VC concentration has increased, with the Top 10 VCs accounting for 17.7% of NAV, and approximately 24% if SpaceX is included. At the same time, the dispersion of VC fund performance is at its widest ever, with the gap between the top deciles and the remaining funds widening.
After the era of zero interest rates ends, startups, like public companies, will shift from growth to profitability, the report said. But companies that successfully raise capital still maintain rapid growth, with growth rates comparable to or faster than the peak of zero interest rates. AI’s share of VC deal value will rise from 15% in 2016 to 86% in 2026. New unicorns are younger: The median age of new unicorns dropped from about 7 years in 2019 to about 4 years in 2026, while the median age of all active unicorns rose to about 15 years.
The report also stated that most of the technology unicorns supported by U.S. VCs in 2026 will have revenue of less than US$500 million and a growth rate of no more than 20%, but have a short cash runway. Post-AI startups show a different growth curve: For companies founded in 2022, revenue growth in the fourth year is about 3 times steeper than in previous years. Top AI applications grew approximately 5 times on small bases and approximately 2.5 times on large bases; the median small base growth was 423% year-over-year, and the large base median growth was 150%.
Private company data is telling the macro story
The report said private technology companies are increasingly becoming a leading signal of important economic issues. OpenRouter data shows that weekly token usage rose from 0.5T in early 2025 to 4.7T in September 2025, and then to 126.2T in September 2026, a year-on-year increase of approximately 27 times, and has doubled twice since June.
The report believes that AI adoption is still “widespread but shallow.” The proportion of top decile users using plug-ins and skills is much higher than that of typical companies: plug-in adoption rates are 95% in OpenAI, 21% in cutting-edge companies, and 9% in typical companies; skills adoption rates are 93%, 19%, and 3%. Databricks’ Smart Router solved 92.3% of coding tasks at $2.13 per task, a 35% cost reduction and a 1.3-point improvement in score. Kalshi begins pricing forward GPU computing power. The report also said that inference providers added about three-quarters of new revenue but accounted for only about one-fifth of EVs.
After the review, the report also gives an outlook. The a16z Growth team expects that AI will expand the demand side: more mature enterprise and consumer adoption, and push to new frontiers such as robotics, biotech, health and AD. Technology in general is improving at an exponential rate. No one can predict the future, but given the speed and pace of change, this cycle will not be like any before, the report said.