Andreessen Horowitz’s State of Markets II finds compute demand still ahead of supply, older GPUs holding value, and public software facing a prove-it re-rating rather than a collapse.
Andreessen Horowitz released the second edition of its State of Markets report on Wednesday, arguing that technology now sits at the center of U.S. equity earnings and that capital inside the sector is moving from software toward hardware.
David George, a general partner who leads the firm’s growth practice, wrote the accompanying essay. The deck covers public and private markets and technology through the first half of 2026. George called it the year’s “canonical chartapalooza” for equity markets and tech.
Tech has compounded earnings faster than the rest of the field for years, he wrote, though off a lower base. Since 2023 the gap has widened. As of late August, tech accounted for about 76 percent of the S&P 500’s total earnings growth in 2026.
That figure is the spine of the report. George said technology can no longer be treated as one sector among many. Durable goods once set the cycle — houses, dishwashers, cars. Tech has taken that job. “Tech is the everything cycle, now,” he wrote.
The year’s rotation inside tech, he said, runs from bits to atoms. Software led the last cycle. Hardware is leading this one. The AI buildout has lifted demand in businesses that used to look sleepy, cyclical and capital intensive: semiconductors, plus power and networking.
Hyperscalers’ profits have financed much of that surge, converting their free cash flow into semiconductor free cash flow. Debt is playing a larger role as well.
AI is not the only pull. George pointed to global infrastructure needs measured in trillions, rising defense spending, grids under pressure from electrification, reshoring of manufacturing, and incoming robots and robotaxis. Public and private equity are funding compute, memory, power, robotics, manufacturing and defense “with an intensity we haven’t seen in decades (if not longer).” His line on the shift was blunt: “atoms are so back.”
On capital spending, one claim was unambiguous. Demand for compute still outpaces supply.
A prominent critic of the buildout had questioned whether GPUs would justify the money poured into them if the chips went obsolete in three to four years. Nvidia’s B200 sales are strong. The open question was what that meant for A100s installed a year or two earlier.
George’s answer, for now: the A100s are still useful. Rental rates, and with them residual values, are supposed to fall as chips age. That is not what the report describes. As intelligence gets cheaper, demand for compute is rising. Prices for the newest GPUs have climbed. Older silicon has held up. The A100 is pricing at or above its level from the start of the year.
Neither compute progress nor model progress has been a zero-sum contest so far, he wrote. Cheaper, better intelligence has added value across the stack. Older chips and older models have kept “substantial value well past the expiration dates assigned by the bears.”
All of that is happening while adoption is still immature. Use is broad. Depth is not.
Nearly 30 percent of S&P 500 companies report some “quantifiable impact” from AI. About 2 percent report a tracked metric. Agentic use follows the same pattern. Only a tiny share of the user base is deploying agents at meaningful scale.
Consumer paid use is thin, too. As of April, barely about 2 percent of U.S. households were paying for an AI service. George said the share is higher now and still growing. It remains small against the size of the market.
GPUs are already running hot, the report said, even though mature utilization has not arrived.
Software took a different beating at the start of the year. The early verdict was that enterprise applications would be wiped out by AI and “vibe-coded everything.” George summarized the mood as “SaaSpocalypse was nigh,” then quoted the joke that went with it: “Run for the hills, enterprise SaaS, and don’t let the door hit you on the way out.”
The sell-off was real. It was more selective than an extinction event. AI, and the threat of AI, played a part. It was not the whole story.
A re-rating had been building since the end of the zero-interest-rate period, when companies traded growth for profit. In 2022, public software was crowded with high-growth, mostly unprofitable names. By 2026 that mix had flipped: about 75 percent are profitable, and only about 30 percent are growing 20 percent or more.
Higher rates were meant to make capital scarce. Firms slowed down and aimed for self-funding growth. Slower growers do not keep high-growth multiples for long. The new pace caught up with the sector.
Not every name was repriced the same way. Fast growers still trade near historical-average multiples, though below zero-rate peaks. There are fewer of those companies now, so the group multiple came down. George’s verdict: no apocalypse. There has been a “prove it.”
Looking forward, the firm expects AI to widen demand — deeper use in enterprises and among consumers, and new ground in robotics, biotech, health and AD. The technology is improving at an exponential pace, George wrote. No one can predict the future. Given the speed of the change, he added, “we’re pretty confident that this cycle isn’t going to be like any previous cycle.”
The first State of Markets report was published earlier this year.
No tracking, no middleman. Follow by RSS (nothing is collected) — or add your email to our self-hosted list.
RSS feed →