a16z's market thesis: big buildout, shallow adoption, huge private companies

a16z released State of Markets II with a companion podcast. These are the numbers in it that matter most for investing:

• Capex. High-tech equipment, software and R&D are now about 55% of US capital spending. The five largest hyperscalers will spend about $780B on capex this year and more than $1T in 2027. a16z says demand still outstrips supply at every point in the chain @a16z.
• Who keeps the profits. a16z estimates hyperscaler free cash flow fell from about $275B in 2022 to roughly $0 today. Semiconductor free cash flow rose from about $50B to $400B over the same period @a16z.
• Enterprise adoption is broad but shallow. 69% of the S&P 500 has a live AI deployment. 30% report a quantified result. Only 2% have AI doing a job they would notice if it stopped. The top 1% of AI spenders spend eight times as much as the top 10% @a16z. David George sees connecting models to company data and workflows as the opportunity, and calls diffusion into companies one of the main themes of the next five years @DavidGeorge83.
• Agents now use more tokens than people. Agents consume nearly 5x the tokens humans do, up 14x since February @a16z. George says agent products have only a few million users today, against a possible audience of billions @DavidGeorge83.
• Bubble question. The partners argue this is an earnings story. Stocks are up about 20% while multiples are down about 20%, and the market trades below 20x earnings with 15% growth. They contrast that with 2000 and 2021 @a16z.
• Private-market scale. Anthropic, OpenAI, Databricks, Stripe, Waymo and Revolut add up to about $2.4T at their last-round valuations. That compares with $1.7T for all IPOs of the past decade, excluding SpaceX @a16z. George's view is that founder-led companies can make bets with longer paybacks while private. He calls the founder "the asset class" @a16z.

a16z's five-year themes are consumer agents, robotics, autonomy, AI x bio, personal health, enterprise diffusion and American Dynamism @DavidGeorge83.

Factory, Cognition and investor conflicts

Factory CEO Matan Grinberg says Factory removed Chris Degnan as board observer and advisor. Grinberg says a conversation Degnan first described as casual turned out to be "formal and recurring" talks with Cognition executives, held while Degnan was sitting in Factory board meetings. Factory says it does not know how much Degnan shared @matanSF. Factory also claims Cognition engineers "feigned interviews" to learn about its product @matanSF. It reports 10x team growth and 100x year-over-year revenue growth @matanSF.

Vinod Khosla replied publicly, calling Factory a "struggling second tier competitor" and saying it was "straight out lying" about Degnan being fired @vkhosla. The facts are disputed. Two investors staked out norms the same day without naming anyone:

• Keith Rabois: "It is unethical per se to even interview at a competitor while attending Board meetings" @rabois.
• Leo Polovets says his firm does not invest in direct competitors @lpolovets.

Funding and valuations

• ElevenLabs closed an employee tender at a $22B valuation, 2x its February Series D. Wellington and T. Rowe Price led it @mati.
• Flow raised a $50M Series B at a $750M valuation. Valor's Antonio Gracias and Atreides' Gavin Baker co-led, with Sequoia participating @parisingh. Flow says it is the default requirements and verification platform for Anduril, Joby, Stoke Space and Rivian, and that GM and the Volkswagen–Rivian joint venture use it @parisingh. Roelof Botha led the Series A, invested personally in this round, and joined the board as an independent member @roelofbotha.
• Neko Health, Daniel Ek's body-scan company, raised a $700M Series C, according to TechCrunch's Equity podcast TechCrunch.
• Instinct. My First Million hosts say Instinct launched about two months ago and went through rounds at $2.5B, $5B and $10B in a span of three to four weeks My First Million. They relay the founder's plan to keep it free and take a cut of transactions. They also relay his claim of more than $1B in transactions while still invite-only, 40–50% of it travel My First Million. These are figures relayed secondhand on the podcast.

Runway: a marketing product and a robotics model

Runway launched Runway Ads, an autonomous performance-marketing engine. It says that using the product internally since July, it scaled ad volume 1,000% while doubling return on ad spend, cut cost per subscriber 41%, and added $100M ARR. These are Runway's own results. Access is limited to select enterprise partners for now, with a wider rollout planned in the coming weeks @runwayml.

Separately, Runway released Praxis-1, an open-weight "World Action Model" for controlling robots, trained on its video pre-training. The bet is that video can make up for scarce robot demonstration data. Runway is testing it with Noble Machines, Standard Bots and Ultra, and says it will release the weights publicly in the coming months @runwayml.

Defense procurement offers startups a faster path

At YC's Startup Industrial Base event in Washington, Emil Michael, Under Secretary of War for Research & Engineering, said the Pentagon's 300-page RFPs were "designed for a lawyer and a lobbyist and a prime contractor." He proposed one-page requirements with a 90-day target @garrytan. He described Drone Dominance as a fly-off: the top 20 applicants compete on objective metrics, and he says it ends with about half a billion dollars in orders split among two or three winners @garrytan. He is funding directed-energy and high-power microwave systems that cost thousands of dollars per shot, not millions @garrytan.

Transportation Secretary Sean Duffy cited SMART, an air-traffic prediction tool from Air Space Intelligence, a small company he said beat the largest vendors. Controllers still make every call @garrytan.

Agent commerce: demand is real, controls are early

At Newcomer's Machine Earning summit, Browserbase CEO Paul Klein said a customer's agent made a $300,000 credit-card purchase Newcomer. Sapiom's Ilan Zerbib estimated an always-on assistant costs about $10 a day to run, and said that has to fall substantially before such products can scale Newcomer. Zerbib also said his own agent booked three luxury hotel nights without permission.

Controls are starting to appear. OnePay caps agent spending with "spend pockets." Browserbase is building a policy engine that lets an agent browse a site without being able to buy Newcomer. a16z's Anish Acharya expects the next unicorns to come from industry-specific personal agents, coding agents for small businesses, and AI for entertainment. He also predicts horizontal agents will "hire" vertical ones, and says too few people are building there Newcomer.

Models and infrastructure

• Embeddings. Perplexity is open-sourcing its contextual embedding models @AravSrinivas. It says pplx-embed-v2-context-9b-preview sets a new state of the art on ConTEB and on turbopuffer's context-bench @perplexity_ai. Cohere launched Embed 5 in Pro and Fast versions @cohere.
• Gemini Argon. Bindu Reddy says Google's Gemini Argon is 5x cheaper than Astra, but is skeptical of its benchmarks @bindureddy.
• OpenAI vs. Jev. OpenAI's Decisions API was explicitly inspired by Jev's fast-classification model, per an OpenAI team member speaking to Latent.Space Latent.Space. That puts direct pressure on TypeSafe.
• FTC. A headline from The Independent says the FTC is investigating OpenAI and Anthropic one day after a White House meeting. No details were available in the material reviewed r/artificial.

Robotics economics and fusion supply chain

According to a Reddit summary of Anthropic's study, existing robots can perform tasks that make up 34% of US working time in at least some settings, but are cost-competitive with human labor for only 0.3% today. The poster says they have not independently checked the estimates r/artificial.

Commonwealth Fusion Systems agreed to buy more than 10,000 km of high-temperature superconducting tape from Fujikura for its ARC plants. It is the largest agreement of its kind for HTS tape. CFS says building ARC depends on its SPARC machine first achieving Q>1 r/Futurology.

Hiring

Greylock's talent lead calls 2026 one of the hardest hiring markets in 15 years, with many companies competing for a small talent pool. The advice is to keep in touch with strong candidates for years after they say no Greylock.