Fred Wilson and USV's $900M Fund: Why AI Will Obliterate, Not Automate

Fred Wilson, Rebecca Kaden and Michael Mignano of Union Square Ventures explain why USV raised $900 million, its largest fund cycle ever, and how they plan to invest it in AI. In this 85-minute episode of Sourcery with Molly O’Shea, recorded at the firm’s New York office, the three general partners cover the thesis they call “obliterate, don’t automate”, why they think ARR hides weak retention, why developers are leaving closed AI stacks, and how big a company has to get for an early-stage venture fund to win.

TLDW

USV kept its funds small for 20 years and has now raised $900 million, because Series A rounds start at $10 million and AI startups burn cash on compute and on standing out. Fred Wilson says AI in 2026 looks like the internet in 2003: the infrastructure exists and the defining applications do not. The firm is betting on four areas (AI applications, AI infrastructure, physical AI and energy), with a thesis that AI will replace whole professions for the end user instead of selling software to them. Along the way the partners argue that trial is at a record high while engagement is at a record low, that personal agents create a consumer security problem, and that USV does not need trillion-dollar outcomes to return its fund.

Thoughts

“Don’t hire a lawyer, use an AI lawyer” is a stronger claim than most AI investing theses, because it bets against the customer everyone else is selling to. The bulk of legal AI and health AI money has gone to tools for lawyers and doctors. Wilson says the opposite: no human in the loop, and he offers his own firm as evidence, claiming 20 to 30 percent of USV’s legal work is no longer touched by a human lawyer and that term sheets come from a model trained on the firm’s old ones. The incumbent argument is sound. A law firm that bills by the hour cannot give the work away, and Kaden’s line from Brad Burnham about betting against incumbents when the new thing threatens their core business is the cleanest version of it. The weak point is the one Kaden raises herself a few minutes later. If USV can train its own legal model without buying one, so can everybody else, and the thing that kills the lawyer may not be a company at all. The thesis explains why incumbents lose. It does not yet explain who wins.

Two statements near the end sit awkwardly next to each other. Wilson says USV does not need to underwrite to a trillion dollars, that many of its biggest winners landed around $10 billion, and that single-digit billions still return the fund. Kaden, a minute earlier, says value is concentrating in the top one or two players faster than ever and that the third or fourth company in a market is no longer a good bet. Both can be true only if USV keeps finding the eventual number one at seed and Series A prices. That is the real reason for the fund size: checks of $10 million to $30 million are what it now costs to hold meaningful ownership in the company that wins. The partners present $900 million as the same strategy at a higher price. It is also a bet that a firm famous for discipline can pay 2026 prices and keep 2006 judgment.

The ARR discussion is the part founders should hear. Mignano’s point is mechanical: many AI products charge before you can try them, often on an annual plan, so revenue arrives before anyone knows whether the customer likes the product. Kaden sharpens it into a description of the whole consumer market, with trial at an absolute high and engagement at an absolute low, because people feel they have to try every new agent that makes the news. The honest admission comes right after. At USV’s stage they cannot wait for clean retention data, since waiting means writing a much larger check, so they look for proxies: how the product feels in their own hands, and how the team iterates. That is a judgment call, and it is refreshing to hear investors say so.

The “Rebel Alliance” thesis rests on a historical memory few people have firsthand. Wilson watched developers build on Facebook and then get cut out when Facebook wanted those businesses for itself, and he thinks builders on Claude or OpenAI fear the same ending. His best evidence is an anecdote from USV’s CEO summit, where Coinbase’s CTO mentioned that the company had moved off Claude and built its own coding agent harness on Pi, an open source project. Wilson says USV is an investor in the company behind Pi, so this is a man talking his book. It is still a useful signal. When a public company with serious security needs decides it wants a stack it controls, the open tooling is good enough to bet real engineering on.

The comparison to early Twitter flatters today’s AI apps more than the partners intend. Twitter and Facebook gave the product away because scale produced a network effect, and the network effect produced lock-in. USV won that argument inside those boardrooms. Today’s startups are subsidizing inference to reach scale too, but Mignano concedes the products are not very retentive, and Kaden says nobody has yet shown how a single-player agent becomes truly multiplayer. Subsidy without a network effect is just a discount. Until someone solves that, the free tier is a cost of competing with ChatGPT and Gemini, not a moat.

Key Takeaways

USV raised $900 million in new funds, by far the most in its history. Mignano says the number came bottom-up from the strategy (how many companies, what check sizes, which stages) and not from wanting to reach $1 billion.
The fund will back roughly 30 to 40 companies, a mix of seeds and Series A rounds, with more seeds than USV has done before.
A Series A now “starts with 10”, meaning $10 million, and Kaden calls that the small side. USV expects its own Series A checks to run $10 million to $30 million. When the firm began, Wilson was thinking of $3 million checks.
USV invests in four areas: AI applications, AI infrastructure, physical AI and energy. Wilson says energy is 20 to 30 percent of the portfolio and calls it a very levered bet on AI.
Building a product is cheaper than ever, but building a company is not. The money goes to compute, to subsidizing inference so users do not have to pay, and to marketing that cuts through a crowd of lookalike startups.
Wilson expects a third to a half of USV’s seed deals to be “proprietary seeds”: incubations, pure founder bets, or ideas the firm hands to a founder with the first two checks attached. Supertake, an AI agent that turns an investing idea into a managed portfolio, was incubated inside USV.
“Obliterate, don’t automate” means backing companies that replace a profession for the end user. Wilson’s examples are an AI lawyer, an AI accountant and an AI doctor with no human in the loop. Portfolio company Doctronic can already write some prescriptions in Utah, the partners say.
Mignano stands by his earlier claim that ARR is a fake metric. USV looks at usage retention: whether people come back daily or weekly, and whether they are gone after a month.
Every USV fund has produced two or three great companies, Wilson says, and none took a straight line. Abridge began as a consumer app for patients, and Duolingo’s first product translated documents.
Wilson expects people to have one personal agent that talks to the agents of Amazon, Shopify and Etsy. He says Amazon recently told Facebook it would not let the Muse agent in, and he expects more of that before a standard emerges.
Personal agents make security a consumer problem. Mignano notes that an agent with access to your email can reset the password on every account you own. Wilson thinks local models on a phone are part of the answer.
USV does not need trillion-dollar companies. Wilson says winners in the single-digit billions produce healthy returns at this fund size, and that many of the firm’s biggest wins were around $10 billion.

Chapters

14:00 The $900M Fund and Why 2026 Looks Like 2003

After an office tour with Mignano, the three partners sit down and Kaden announces the raise. Wilson goes back to the firm’s first letter to investors. In 2003 the internet’s infrastructure was built and the applications that defined the next decade had not appeared. He sees AI in the same place. Mignano adds that a new wave of agent products, including Instinct, Muse and Grokbot, has arrived in recent weeks.

16:54 Cheap to Build, Expensive to Win

O’Shea asks where the capital goes if open source models make building cheaper. Mignano says it goes to breaking out of the noise, especially in consumer. Kaden says compute costs are in a different hemisphere from anything startups have faced, and that nobody knows when models stall or how low compute prices fall. Wilson points out that ChatGPT, Claude and Gemini all subsidize their chat products, so a startup cannot charge for what a competitor gives away. He recalls Twitter raising $5 million, then $25 million, then $100 million in successive years, and USV arguing against early monetization.

23:54 Personal Agents and Why Amazon Is Shutting Them Out

O’Shea pushes back on agents as a category, since every company will ship one. Wilson agrees that Amazon, Shopify and Etsy will each have an agent, and argues that people will still want their own agent to deal with all of them. Today agents log into services and pretend to be the user, which Amazon dislikes. He expects more blocking, then a standard for agents to talk to each other.

31:02 Why Small Funds Pick Better, and What a Series A Costs Now

Kaden lists the case for a small fund: it is easier to return, the partners can be pickier, and nobody has to be in a set share of the market for the math to work. Limited partners were told about the change over many quarterly calls, and several asked whether USV should raise more. On pricing, the partners say a Series A starts at $10 million and that billion-dollar rounds labelled Series A are outside their zone.

40:17 Is ARR a Fake Metric?

Asked what teams should focus on, Mignano returns to a clip from his last Sourcery appearance. Subscriptions make it easy to collect money before a customer knows whether the product is any good, so USV studies usage retention instead of revenue retention. Wilson shifts to the long view. At USV’s CEO summit, Abridge’s founder told him the company earned all its revenue from one product for five or six years and has launched two more in the last twelve months, which he calls a three-legged stool.

46:00 Twenty Years of Winners, None in a Straight Line

Wilson runs through the stars of more than ten early-stage funds: Twitter, Etsy and Indeed in the first, then MongoDB, Twilio, Stripe, Coinbase, Duolingo and Abridge. Kaden says an investor in early Abridge was really backing a team and a direction, since the original plan was wrong in the details. Wilson says great founders combine a vision with the ability to tell the story and gather people and money, and that USV misjudges this more often than it gets it right. Founders fail when they overthink, micromanage, struggle to recruit, or cannot make a hard change early enough.

57:46 The Rebel Alliance and Why Coinbase Moved Off Claude

USV calls its open AI infrastructure thesis the Rebel Alliance. Wilson says many developers fear building on a closed stack, because the model company may later compete with them, as Facebook did with businesses built on its platform. Wilson tells the story of Coinbase’s CTO revealing on stage that the company built its own coding agent harness on the open source project Pi, and says Shopify has done something similar.

1:01:44 Obliterate, Don’t Automate: AI Doctors and AI Lawyers

Mignano defines the application thesis. The last decade of software automated existing businesses, and AI can restructure whole markets by giving people expertise that institutions used to gatekeep. Kaden says the phrase comes from a blog post Wilson wrote in 2014 and describes what USV has always wanted, which is to move value from the middle to the end user. Wilson gives the AI lawyer example and says the tool will never be as bad as it is today. Incumbents will not follow because their business model depends on people. Google is the exception, he argues, because YouTube, Waymo, Gmail and Android let it cannibalize search.

1:08:29 Who Protects You From Your Own Agent?

Mignano expects security to shift from the enterprise to the consumer as people hand everything to an agent. Wilson mentions a portfolio company building a sandbox for this problem and says his 96-year-old mother is not prepared for a world where she cannot tell what is real. Wilson admits he has been hacked four times in 15 years of owning crypto and still has about 99 percent of it. The group agrees that Coinbase is in many ways a security company, and Kaden wonders whether AI is waiting for an equally trusted brand.

1:16:28 Betting on Markets That Don’t Exist Yet

O’Shea asks whether every company has to become a trillion-dollar company. Kaden says no, but that USV bets on markets that are small today and could be huge by 2040, as crypto was when it backed Coinbase and as nuclear is now with Radiant. She also says value is collecting in the top players, so backing number three or four no longer works. Wilson says single-digit billions are enough. On overpriced companies that fall back to earth, he calls Bending Spoons a good home for them once the venture journey is over. Asked who inspires them, all three say founders, naming Zack Kanter of Stedi and Mikey Shulman of Suno.

Notable Quotes

“The infrastructure was there. We had Google, we had Amazon, we had Yahoo, we had eBay, but we didn’t yet have Facebook, we didn’t yet have Twitter, we didn’t yet have YouTube, we didn’t yet have iPhone.”
Fred Wilson, on why AI in 2026 reminds him of the internet in 2003

“Trial behavior is at an absolute high and engagement behavior is at an absolute low.”
Rebecca Kaden, on why early revenue in consumer AI can mislead investors

“Don’t hire a lawyer, use an AI lawyer. Don’t hire an accountant, use an AI accountant. Don’t hire a doctor, use an AI doctor.”
Fred Wilson, giving the short version of obliterate, don’t automate

“The best time to bet against an incumbent is when them doing what you want to do fundamentally threatens their current business.”
Rebecca Kaden, quoting USV co-founder Brad Burnham on why incumbents will not replace their own people with AI

“If you give your agent access to your email, which is where all your password reset emails go, they just automatically will be able to access literally every account you own.”
Michael Mignano, on the consumer security problem that personal agents create

“I have owned crypto assets for 15 years and I’ve been hacked four times.”
Fred Wilson, on learning security the hard way

“I don’t think a trillion is what we need to underwrite to at USV.”
Fred Wilson, on why winners in the single-digit billions still return a fund this size

Watch the full conversation between Molly O’Shea, Fred Wilson, Rebecca Kaden and Michael Mignano here.

Fred Wilson and USV’s $900M Fund: Why AI Will Obliterate, Not Automate