a16z State of Markets 2026: Tech Is the Everything Cycle, Atoms Are Back

a16z’s State of Markets II, led by David George, argues tech is now the engine of the US economy and that the AI buildout is paying off, so far. The second edition of Andreessen Horowitz’s growth-team chart book covers the first half of 2026 in roughly 90 slides, paired with a short essay on the a16z newsletter and the full deck as a PDF. It runs from S&P 500 earnings to GPU rental prices to venture fund returns, and nearly every chart points to the same conclusion: demand for compute is still outrunning supply.

TLDR

Tech produced about 76% of S&P 500 earnings growth in 2026, and the market’s record highs came from profits while valuation multiples actually shrank. Hyperscaler capex is heading toward $800 billion this year and over $1 trillion a year from 2027, paid for first with free cash flow and now with debt. Money has rotated from software to hardware: chips, memory, power and the rest of the physical supply chain. Software got a “prove it” re-rating rather than an apocalypse, and private markets now hold the biggest companies, the biggest exits and the most extreme power-law returns ever recorded.

Thoughts

The strongest argument in the deck is the “hard to call it a bubble” slide, and it holds up better than most bull cases do. The S&P 500 is up 12.7% year to date while forward earnings are up 27.7% and the P/E multiple has fallen 11.7%. Tech is the extreme case: prices up 22%, earnings up 56%, multiples down 22% and now about 19% below their five-year average. Bubbles are built on prices running ahead of earnings, and this year it went the other way. The weak spot sits a few slides later. Hyperscaler free cash flow, as a share of sales, is projected to fall to roughly zero in 2026 and 2027, and borrowing by the hyperscalers and Nvidia (corporate bonds plus off-balance-sheet vehicles) is running above $300 billion this year, against a decade of mostly under $50 billion. The answer a16z gives, that return on invested capital (about 20%) still beats the cost of capital (about 10%), is right as far as it goes. But that ROIC was earned on the old asset base. The real test is the return on the new spending, which no chart can show yet.

The thesis really rests on one variable: the price of compute. If GPUs became obsolete in three years, as the bears claimed, the depreciation math would sink the buildout. Instead the deck shows rental prices holding or rising across every generation (B200 around $5.69 an hour, H100 at $2.63, even the five-year-old A100 at $1.59), while the price of model output per token keeps falling. That is the Jevons paradox the deck keeps naming: as intelligence gets cheaper, total demand for it rises faster than the price falls. The most telling sign of how central this has become comes near the end of the deck, where Kalshi runs a prediction market on the end-of-month price of B200 compute. When a prediction market is trading on GPU rental rates, the industry is treating compute like oil. Watch that number. The day older GPU rental prices start falling steadily is the day this story changes.

The demand charts are more interesting than the headline adoption numbers. Adoption is wide and shallow: 69% of S&P 500 companies point to a live AI deployment, but only 2% disclose a metric they track over time, and only 2.2% of US households paid for an AI subscription as of April. What drives token demand is a tiny group of heavy users. Ramp data shows the top 1% of companies outspending the top 10% by about 8x on AI tools. OpenRouter shows agent traffic passing human traffic on February 6, 2026, with weekly volume up roughly 27x in a year to 126 trillion tokens. In other words, all of today’s compute demand comes from the top 1%. If even the next 10% of companies start working the way the top 1% do, the capex forecasts that have missed low every single year will miss low again.

The software section is more honest than the “SaaS-prove-it” slogan suggests, because the deck includes the counterargument. One slide shows newspaper stocks selling off about five years before newspaper earnings collapsed. That is the bear case for software in a single chart: the market can price disruption long before the income statement shows it. The deck’s reply is that software fundamentals have been stable, and Stripe’s payment data shows mature SaaS revenue growth recovering to roughly 24% while software companies less than a year old are growing several hundred percent a year. Taken together, I read it this way: incumbents with weak defenses are being priced like newspapers, and the new revenue is going to companies that did not exist two years ago. The real split is between old software and new software, more than between hardware and software.

The final third of the deck makes the point with the most consequence and the least attention: the biggest wealth creation of this cycle is happening where ordinary investors cannot buy in. Active US unicorns are now worth $5.34 trillion, more than the whole Russell 2000 at $3.5 trillion. The five largest private tech companies are worth more than every tech IPO of the past decade combined. In the first half of 2026 the top 1% of exits produced 84% of all exit value, and the top 30 venture portfolios have risen about 23x since 2017, against roughly 5x for the ten largest public companies. Read this deck as an a16z document, too. It is a reasoned case for a cycle a16z is heavily invested in, and many of its “private data” slides (Flock, Anduril, robotaxi data, inference providers) carry the portfolio disclaimer. The charts are useful. The conclusions come from investors who benefit if those conclusions hold.

Key Takeaways

Tech contributed about 76% of S&P 500 earnings growth in 2026, makes up about 60% of the world’s 100 largest companies by market cap (ex-China), and now accounts for 55% of all US capital spending.
The rally came from earnings, not rising valuations: tech prices rose 22% while tech earnings rose 56% and multiples fell 22%. A record 93.6% of S&P 500 companies met or beat estimates.
This was not a Magnificent 7 year. The rest of tech is up roughly 40% year to date while the hyperscalers lagged, because they are spending their profits on the buildout.
Hyperscaler capex is projected near $800 billion in 2026 and above $1 trillion a year from 2027, and Wall Street has underestimated it every year since 2021.
Free cash flow at the hyperscalers is expected to bottom near zero until about 2028, so debt markets have stepped in. Hyperscaler ROIC of about 20% is still roughly double their weighted cost of capital.
OpenAI and Anthropic added more revenue in 2026 than all of public software (excluding the clouds) combined, and major cloud revenue backlogs have roughly doubled year over year.
GPUs are not becoming obsolete on schedule: rental rates for A100s, H100s, H200s and B200s have held or risen in 2026, so older chips keep their resale value.
Adoption is broad but shallow. 69% of S&P 500 companies cite a live AI deployment, 29% report a quantified result, and only 2% disclose a metric they track over time.
Agents now consume more tokens than people on OpenRouter, and about 86% of agent tokens come from cached prompts. Frontier labs take about 70% of spend on about 35% of tokens.
Companies that adopt AI heavily added entry-level headcount compared with non-adopters, and above-trend hiring for software developers resumed in 2025 and 2026.
The shift from bits to atoms reaches well beyond chips: an estimated $90 trillion of global infrastructure need by 2040, utility capex heading toward $276 billion a year, rising defense budgets, Waymo at 500,000 weekly rides and a coming wave of robots.
Memory has been the single best AI trade, with DRAM and NAND spot prices up roughly tenfold. Power gear is the bottleneck: transformers average about 128 weeks of lead time, and heavy gas turbines ordered today arrive around 2031.
Software was repriced, not killed. About 75% of public software companies are profitable but only about 30% grow 20% or more. Cybersecurity is the runaway winner and vertical SaaS held up better than horizontal.
VC-backed exit value reached about $2.19 trillion in the first half of 2026, against a prior peak of $865 billion in 2021, and 86% of US venture deal value went to AI-related companies.

Detailed Summary

Tech Is the Everything Cycle

The opening section argues that tech is no longer a sector but the main driver of the economy, taking the place durable goods like houses and cars once held. Tech accounts for nearly half of S&P 500 profits, its earnings have grown about 3.8x more than the rest of the index since 2010, and S&P margins are near a record 17.5%. The deck says the AI infrastructure buildout is responsible for nearly all net-new construction spending, blue-collar job growth and new investment-grade bond issuance. Retail traders have piled into the same themes, with leveraged ETF assets near $192 billion and semis the fastest-growing slice.

AI Capex: Big, Underestimated and Now Debt-Funded

Hyperscaler capex went from $97 billion in 2020 to about $416 billion in 2025, with consensus near $800 billion for 2026. Measured as a share of GDP, it is now at or past the peaks of the railroad, telecom and shale booms. Every year since 2021 the consensus capex forecast has been revised up. With free cash flow consumed, Amazon, Alphabet, Meta, Oracle and Microsoft have turned to bond markets and special-purpose vehicles. The deck also points to wider economic effects: data center jobs pay sizable premiums (facilities managers earn 64% more), and one study finds that states with more data center growth saw residential electricity prices fall.

Capex Is Working: Revenue, Backlogs and GPU Prices

The justification is revenue growth. OpenAI and Anthropic’s combined annualized revenue climbs steeply through 2026, cloud backlogs at Microsoft, Google and Amazon have roughly doubled, and neoclouds like CoreWeave are growing faster than AWS or Azure did at the same age. Earnings-call examples run from Meta (a 15% lift in conversations from its GEM model) to ServiceNow (AI contract value over $1 billion) to Microsoft’s 30 million paid M365 Copilot seats. Enterprise AI spending is still small: only about 20% of organizations name cost as a constraint, and AI is a smaller share of IT budgets than cloud was in its second year.

Power Users, Agents and the Jevons Paradox

The deck’s most original section shows how lopsided usage is. A small group of power users and “frontier firms” use orders of magnitude more AI than everyone else, and the gap is widening. Price per unit of intelligence fell faster in three years than PC prices did in fifteen, and demand rose faster still: agents passed humans in token volume, open-weight models are gaining share, and older models still take a large share of enterprise spend. On the consumer side, paid subscriptions are up about 5x since 2025, retention curves are rising rather than decaying, search referrals are falling, and Meta’s Muse app has taken a growing share of US downloads.

Atoms Are So Back

Hyperscaler free cash flow has effectively become semiconductor free cash flow. A Sankey chart shows how each $100 of AI capex splits across chips, networking, power, cooling and construction. Asset-heavy stocks have sharply outperformed asset-light ones after a decade of losing to them, and memory makers have been the biggest winners. The deck asks whether semis are still cheap, since many large chipmakers trade at forward multiples well below what their expected earnings growth would normally justify. Outside AI it points to manufacturing output, electrification, cheaper rocket launches, robot parts supply chains, robotaxis and defense orders, all heading higher.

SaaSpocalypse? More Like SaaS-Prove-It

The 2026 software sell-off was selective. Horizontal software multiples fell from about 12.3x to 2.7x revenue, while cloud and data infrastructure held 9.1x and cybersecurity and observability stocks rallied. a16z frames it as the delayed bill for the end of near-zero interest rates (ZIRP): companies traded growth for profits, and slower growth earns lower multiples. Revenue growth and operating leverage across software have been stable or improving. The section closes with two open questions: will software follow print media, or will AI raise returns on equity for capital-light businesses too?

The Private Side and the Power Law

Private companies now reach mega-cap size without listing: the deck shows SpaceX at $1.78 trillion, Anthropic at $965 billion and OpenAI at $852 billion in their latest disclosed valuations. Secondary markets show almost no discounts to the last funding round. Concentration also applies to managers: the gap between top-decile venture funds and the rest has never been wider, and recent vintages show 90th-percentile net IRRs of 30-40% against negative returns at the 25th percentile. New unicorns are younger, and post-AI startups are reaching year-four revenue on a curve about 3x steeper than earlier cohorts.

Private Company Data Is Telling the Macro Story

The last section uses data from private companies to answer macro questions: OpenRouter for token demand, OpenAI for usage distribution, Databricks for model routing (its Smart Router solved 92.3% of coding tasks at $2.13 each, about 35% cheaper than the top single model), Kalshi for forward GPU pricing, and Stripe for SaaS revenue. Specialized inference providers such as Fireworks, Baseten, Together AI and Modal added about three quarters as much revenue as Snowflake and Datadog combined, at about a fifth of the valuation. The deck ends on seven open opportunities: heavy-use consumer AI, robotics, autonomy, AI x bio, personal health, enterprise adoption beyond coding, and American Dynamism.

Notable Quotes

“Tech is everywhere and in everything. Tech is the everything cycle, now.”
David George, State of Markets II essay, on why tech can no longer be called just a sector

“AI infra is arguably the only pro-cyclical impulse firing in the economy right now.”
State of Markets deck, on how much of US growth rests on the buildout

“It’s not chicanery, so much as simple unit economics, that have made all this capital allocation look pretty smart, thus far.”
State of Markets deck, answering the GPU depreciation bears

“Somehow the GPUs keep arriving on time, but every other piece of the supply chain continues to have unusually long lead-times.”
State of Markets deck, on the semiconductor and power backlog three years into the buildout

“If that’s the case, then future investors are going to have a hard time looking back on those multiples . . . between all the tears, that is.”
State of Markets deck, on whether chipmakers are still cheap

“There’s been no apocalypse for software, but there has definitely been a ‘prove it.’”
David George, State of Markets II essay, on the 2026 software sell-off

“We’ve entered a new cycle of real assets, and while tech is running through all of it, capital-light bits is no longer the (only) name of the game.”
State of Markets deck, on the rotation from bits to atoms

Read the full State of Markets II essay here, and download the complete 90-slide deck here.