Andreessen Horowitz’s growth team has released the second edition of State of Markets II, a report that looks at market conditions through at least the first two quarters of 2026 and lays out its view of what comes next. Its central argument is direct: technology is no longer just a sector inside the equity market. It has become the market’s “everything cycle.” Within tech, attention has also moved away from software and toward hardware and infrastructure.
The report adds that fears of rapid GPU obsolescence have so far been weakened by real-world pricing data, that software is not heading for extinction but into a phase where it has to prove itself, and that AI adoption still appears early.
Technology, not just another sector
One of the report’s core calls is that technology has evolved from what it describes as a cottage industry into a primary driver of the global economy. In a16z’s framing, nearly every company is now a technology company in some sense, and truly paper-and-pencil businesses are increasingly hard to find. Compared with other industries, tech does more to drive investment, research and development cycles, earnings growth, and margin expansion. The world’s largest, most profitable, and fastest-growing companies are, in many cases, technology companies.
The report says the sector’s rise as a capital-markets force truly accelerated after the global financial crisis. After the housing collapse and credit tightening, technology offered growth-oriented returns with a lighter capital profile for investors who had become more cautious about asset-heavy businesses.
a16z argues that technology also carries built-in operating leverage because addressable markets remain far from saturated while software’s marginal cost is close to zero. In hindsight, the report says, investors who doubted that loss-making tech companies could become high-margin machines missed one of the defining themes of the past decade.
It says software did eat the world, but the centrality of technology in capital markets has moved to another level. Since 2023, tech has been the earnings growth story. Through the end of August 2026, technology accounted for about 76% of total S&P 500 earnings growth.
That leads to a broader conclusion in the report: technology can no longer be described simply as a sector. Durable goods such as homes, dishwashers, and cars once defined cyclical turns. Now, a16z says, technology has taken that role and become the “everything cycle.”
Inside tech, the market is rotating from software to hardware
At the same time that tech has become the “everything cycle,” the structure inside the sector is changing. The report says the annual and multi-year theme is a shift from bits to atoms. Software led the last technology cycle. Hardware is taking the lead now.
AI buildouts are driving a surge in demand for semiconductors, power, and networking, sectors that were traditionally slower-growing, cyclical, and capital-intensive. According to the report, a large share of that demand is being financed by the historic profits of the largest technology companies, effectively turning hyperscaler free cash flow into semiconductor free cash flow. Debt markets are also playing a larger role.
a16z says the move goes beyond AI alone. It points to trillions of dollars in global infrastructure demand, rising defense spending, power grids dealing with broad electrification, manufacturing reshoring, and the approach of robotics and robotaxis.
After years of trailing software, hardware and infrastructure have become a market focus again, the report says. Public and private capital is being directed toward compute, storage, power, robotics, manufacturing, and defense at an intensity not seen for decades, and perhaps longer. Its shorthand for that change is simple: atoms are back.
AI as a generational platform shift
The report describes AI as a generational platform transition and uses hyperscaler capital expenditures to show the scale of this cycle. Those expenditures rose from about $97 billion in 2020 to $241 billion in 2024 and $416 billion in 2025. For 2026, the report estimates $777 billion, and from 2027 onward it expects spending to exceed $1 trillion a year.
a16z argues that every major computing cycle has been larger than the one before it. From mainframes, minicomputers, and PCs to the desktop internet, cloud and mobile, and now AI, device counts, user bases, and capital spending have all stepped up by an order of magnitude.
The report also says annual market forecasts have repeatedly underestimated capital expenditures, as demand and pricing kept pushing investment higher. Capex as a share of GDP is now near, or above, historical peak levels seen in railroads, oil and gas, and telecommunications. Hyperscaler free cash flow is being consumed by capex, and the report expects that pressure to last until around 2028.
As cash resources are pushed harder, debt markets have started to fill the gap. AI-related borrowing rose sharply in 2026 and maturities have gotten longer, the report says. Even so, hyperscaler return on invested capital remains above both the cost of debt and weighted average cost of capital, which, in a16z’s view, still gives the current borrowing-and-investment cycle a financial basis.
The spending wave is also spilling into the real economy. The report says data center construction has added more than 300,000 jobs across construction and skilled trades, with wage premiums for roles such as facility managers, construction managers, and network engineers relative to similar non-data-center jobs.
Revenue growth is starting to support the spending case
a16z says AI revenue growth is arriving at a pace and scale the market has not seen before. Combined annual recurring revenue at OpenAI and Anthropic has climbed rapidly since the fourth quarter of 2024 and is projected to approach the $150 billion range by the third quarter of 2026.
The report says OpenAI and Anthropic together are on roughly $100 billion of revenue in 2026, above the roughly $63 billion generated by public software companies excluding hyperscalers. Cloud revenue backlog has doubled year over year, while hyperscaler free cash flow is expected to recover materially in 2029 to 2030.
a16z also points to rapid growth among neocloud companies, naming CoreWeave, Nebius, and Applied Digital as examples with steep revenue curves. On concerns about GPU depreciation and obsolescence, the report says rental rates and residual values for older GPUs have not collapsed. Instead, they have held steady or moved higher. Data on leasing prices and residual values for the B200, H200, H100, and A100 are cited in support of that view.
Adoption is still early and often shallow
At the adoption layer, the report says AI remains early. Among S&P 500 companies, 69% say they have live AI deployments, but only 2% disclose the metrics they track, and 0% break those metrics out and track them separately.
There is still room for enterprise AI spending to rise. Only 20% of organizations say AI-related costs limit usage, according to the report, and most industries expect to raise AI spending. Inference spending currently accounts for only about 0.3% of S&P 500 revenue. By comparison, cloud reached roughly 4% of IT budgets in its second year, and a16z argues AI’s total addressable market is larger than the IT budget alone.
On the consumer side, paid AI subscriptions have grown about fivefold since 2025, and average household monthly spending has increased, though search referral traffic remains small. The report compares the current AI buildout to the mobile internet sequence: semiconductors first, then infrastructure, then software and services. It says AI may follow a similar order.
Why a16z says the GPU obsolescence case is overstated
The sharpest debate around hardware and capex centers on compute supply and demand and on GPU life cycles. The report says compute demand still exceeds supply. Well-known bears had argued that GPUs become obsolete in three to four years and questioned whether so much capital spending made sense. They also asked what the strong demand for the B200 meant for A100 deployments made one or two years ago.
a16z’s answer, at least for now, is that the demand curve for AI compute has shifted upward and the A100 remains useful. In a normal cycle, rental rates and GPU residual values would decline over time. That is not what the data show, the report says. As intelligence becomes cheaper, compute demand rises. As newer chips get more expensive, older chips have also stayed firm. The price of the A100 is equal to, or above, where it stood at the start of the year.
The report says the story is far from over, but so far progress in models and progress in compute are not a zero-sum trade. Better and cheaper intelligence is creating value across the ecosystem, and older chips and models are still retaining meaningful value after the expiration dates imagined by bears.
At the same time, AI adoption still looks immature: broad, but shallow. Nearly 30% of S&P 500 companies say AI has had some “quantifiable impact,” yet only about 2% report any trackable metric. In agentic use cases, only a very small share of users have deployed at scale.
On the consumer side, only about 2% of U.S. households were paying for some kind of AI service as of April. The report says that figure is now higher and still rising, but remains small overall. Its conclusion is that GPUs are already running hot while mature AI adoption and utilization are still early.
SaaS is not dead, but it has to prove itself
The report then turns back to software. Earlier this year, some market commentary argued that software was dead, that AI could generate everything through “vibe coding,” and that SaaS faced an endgame. a16z says reality is more nuanced. There has been a sell-off, but it has been more selective than the “death of software” thesis suggests. AI, or the threat of AI, mattered, but it was not the only reason.
Part of the repricing in public software, the report argues, was overdue. After the zero-interest-rate era ended, technology companies shifted from trading growth for profitability toward more durable earnings profiles. In 2022, markets were full of high-growth but mostly unprofitable software companies. By 2026, that picture had flipped: roughly 75% were profitable, but only about 30% were growing above 20%.
Higher rates were meant to make capital relatively scarcer, so it was rational for companies to move from loss-funded growth toward slower but more self-sustaining expansion. But slower-growth companies cannot keep high-growth valuation multiples forever, and that new normal eventually caught up with the software sector.
The report also says this does not apply equally across the board. Faster-growing companies still trade at multiples in line with historical averages, though no longer at zero-rate peaks. The problem is that there are fewer of them, which has pulled down the sector as a whole. a16z’s summary is blunt: software is not dead, but it is in a prove-it phase.
Private markets are holding more of the value creation
a16z argues that private markets are accumulating value on a historic scale. Year to date in 2026, venture funding stands at about $72 billion, while venture-backed exit value is about $2.188 trillion.
The largest private companies are also bigger than ever. The report names SpaceX, Anthropic, OpenAI, Databricks, Stripe, Revolut, Anduril, Cursor, Ramp, and Waymo, saying their combined valuations run into the trillions of dollars. It adds that the top five private companies are now worth more than the total value of tech IPOs over the last decade.
a16z describes the pattern as “private-for-longer,” meaning companies are listing later and more value is being created before the IPO. Typical time to IPO has lengthened from about three years in 1999 to 2005 to more than 10 years in 2020 to 2025.
Active U.S. unicorns are now worth $5.34 trillion in aggregate, above the Russell 2000’s $3.5 trillion, the report says. Power-law dynamics in exits have also intensified, with the top 1% of exits accounting for 84% of exit value and the top 10% accounting for 94%.
Secondary markets show the same pattern. Employee secondary participation is below 60%, but subscription rates are at historical highs. Discounts in secondaries are limited and small. Median pricing in the secondary market is at a 0% discount relative to the previous round, while the 75th percentile trades at a 23% premium and the 90th percentile at a 79% premium.
The report also compares performance between private and public market winners. From 2017 to 2026, the Top 30 VC portfolio index rose from 100 to 2,333, while the Top 10 public companies rose only to 542. Venture concentration has increased as well: the top 10 VC names account for 17.7% of NAV, or about 24% including SpaceX. At the same time, dispersion in VC fund performance has widened to a record, with the gap between top-decile funds and the rest growing further.
AI now captures 86% of venture deal value
The report says startups, like public companies, shifted from a growth-first mindset to a profitability focus after the zero-rate era ended. But the companies that are still getting funded continue to grow quickly, in many cases at rates comparable to, or faster than, those seen at the zero-rate peak.
AI’s share of venture deal value has risen sharply, from 15% in 2016 to 86% in 2026, according to the report.
New unicorns are also getting younger. The median age of new unicorns has fallen from about seven years in 2019 to about four years in 2026, while the median age across all active unicorns has climbed to about 15 years.
a16z adds that among U.S. venture-backed tech unicorns in 2026, most generate less than $500 million in revenue, are growing no more than 20%, and have shorter cash runways. Post-AI startups look different. For companies founded in 2022, fourth-year revenue growth is about three times steeper than for cohorts from earlier years.
The report also gives application-level growth figures. Top AI applications have grown about fivefold on small bases and about 2.5-fold on large bases. Median year-over-year growth is 423% for the smaller-base group and 150% for the larger-base group.
Private-company data are becoming macro signals
a16z says private technology company data are increasingly acting as leading indicators for larger macro questions. OpenRouter data show weekly token usage rising from 0.5T at the start of 2025 to 4.7T in September 2025 and then to 126.2T in September 2026. That is roughly 27x year-over-year growth, with usage doubling twice since June.
The report again describes AI adoption as broad but shallow. Top-decile users make much heavier use of plugins and skills than typical enterprises. Plugin adoption stands at 95% for OpenAI, 21% for frontier companies, and 9% for typical companies. Skills adoption is 93%, 19%, and 3%, respectively.
It also cites Databricks’ Smart Router, which solved 92.3% of coding tasks at a cost of $2.13 per task, cutting cost by 35% and raising scores by 1.3 points. Kalshi has started pricing forward GPU compute. Inference providers, the report says, added roughly three-quarters of new revenue while accounting for only about one-fifth of enterprise value.
a16z’s outlook
After reviewing the market, the growth team says AI will widen the demand base from here. Enterprise and consumer adoption should mature further and extend into robotics, biotech, health, and AD.
The closing line of the report is that technology is still improving at an exponential rate. No one can predict the future with precision, but given the speed and tempo of change, a16z says this cycle will not look like any cycle that came before it.