Andreessen Horowitz has pulled the returns of 3,000 US venture capital firms over the last two decades. Only 20 delivered a consistent 3x net return.
Most pitches to limited partners are about finding more good funds to spread across. David George and Aram Verdiyan's data says the opposite: spreading a portfolio across dozens of venture firms all but guarantees the average return, because almost all of the excess return sits with a small group that keeps winning.
"We've looked at the data of 3,000 venture capital firms in the US. Only 20 have achieved consistent 3x net returns over the last two decades."
George runs Andreessen Horowitz's Growth Fund, one of the largest pools of late-stage venture capital in the industry. Verdiyan is a partner at Accolade Partners, which allocates more than $3 billion into venture and growth funds including a16z's, and spent years earlier in his career as an investor inside a16z itself.
I listened to the full episode so you can skip it. 49 minutes of audio, 13 minutes of reading.
Here are the 9 numbers that matter.
👤 Guests: David George, General Partner and Head of the Growth Fund at Andreessen Horowitz; and Aram Verdiyan, Partner at Accolade Partners, a venture and growth-equity fund of funds that allocates more than $3 billion and has invested in a16z's own funds
🎙️ Host: Jen Kha, Managing Partner and Head of Global Partnerships at Andreessen Horowitz
📰 Published: 10 September 2026 on YouTube (a16z)
🔴 YouTube | 🟣 Apple Podcasts | ⏱️ 49 min | ✅ Time saved: 36 min
Key Takeaways
Only 20 of 3,000 US venture firms have delivered a consistent 3x net return over 20 years
Access to the category-defining company in every vintage is what separates them, not fund count
Capital itself now compounds a company's advantage, which never used to be true
George says a startup used to break under too much money; now it buys more compute and gets stronger
The best venture funds tolerate a 60% loss rate at the early stage, and treat a lower one as under-risking
Growth-stage loss rates run 10% to 20% instead
Late-stage venture can now return an entire fund on one position
A single company sized at 5% to 10% of a late-stage fund can do it — George says that math didn't exist before
AI spend is still wildly concentrated: $12 a month per employee at the median company, $7,000 at the top 1%
Cursor's own investors doubted it until the morning SpaceX announced a $60 billion deal to buy it
A 1-percentage-point acceleration in a public software company's growth is worth 3 percentage points of EBITDA
Pre-ChatGPT-era LBO software deals were priced at 25 to 32 times EBITDA; George says the same assets trade near 2 times revenue now
The best venture firms had a chance to sell Stripe and Databricks years ago and refused
Family offices among their LPs don't want the cash back, because they don't want to pay the tax on it
Robotics is going to be bigger than the AI language products, in George's view, and it's barely started
Fewer than 10,000 Waymos are on the road in the whole US
The AI bottleneck isn't demand, Verdiyan says — it's energy, data centers and chips
The US has a speed-to-power problem, not a generation problem
1. The power law goes systemic
The two guests opened by restating the thesis that got them booked: capital now compounds an AI company's edge instead of straining it. "Right now, clearly the power law is more extreme than it has been in the last 10 to 20 years of technology investing. For the first time, you can take capital and throw it at a company and it compounds their advantage," George said
His example of the old failure mode: a startup used to break under too much cash, hiring faster than it could coordinate. Now the money buys compute, and compute makes the product better
Verdiyan's framing is that AI is not one more software category but a claim on the whole economy. "AI is attacking every facet of the GDP, transportation, labor, services, capital, coordination. There hasn't been a technology paradigm that hits on 30 trillion in GDP at the same time"
The growth curve, on his numbers, is unprecedented. AI has reached $100 billion in revenue in four years; software-as-a-service took 15 years to get there, and Verdiyan said penetration of actual demand is still nowhere close to done
He argued this makes AI a core position for an allocator, not a satellite one. Coming from private equity and growth equity 18 years ago, he said the venture and growth asset class overall has grown from a cottage industry into one he estimated in the trillions of dollars. Companies are staying private for longer, and on his view "they're not going to reverse"
2. AI's TAM Dwarfs Software
George's example is healthcare, where AI is going after the underlying labor rather than just selling software licenses. "That's claims, billing, administration. That's a trillion dollar industry. So the TAM of AI can be 10x plus bigger than traditional SaaS or healthcare IT"
He said the industry has no real way to size the eventual market: "We have no idea what how big the TAM can get," and admitted, "I've been chronically wrong about how big these outcomes can get"
Verdiyan made the same point about labor generally. He said dollars spent on labor in the US economy run roughly 40 times what is spent on software, and argued AI is not simply "the next evolution of software" because it targets that much larger base
His own firm's general counsel, George said, now spars with lawyers on legal questions instead of deferring to them, using Harvey — and George's own billable hours have gone up as a result, not down, because the new use cases are "expansionary" rather than a replacement for existing work
Actual adoption is still tiny relative to the opportunity. George put the median US company's AI spend at $12 per employee per month; the top 1% of the companies in a16z's data set was spending $7,000 per employee per month
He estimated the most advanced banks are running AI on only about 1% of headcount cost today
3. High Risk Is the Strategy
George said a16z deliberately tolerates loss as the cost of catching category winners. "If we're not losing money in a given fund on a given amount of investments, we're not taking enough risk." He put the loss rate on the firm's best-performing early-stage funds at 60%, against roughly 10% to 20% at the growth stage
What's new is that late-stage venture can now return a whole fund on a single position. George said a late-stage fund's best company should be 5% to 10% or more of the fund, sized so it alone can return the fund — something he said "didn't exist before. It now does"
Winner-take-all inside a category, not across the whole market, is how they see it. George pushed back on the idea that AI produces one winner per layer: "There wasn't winner take all right now," he said of past tech cycles, describing instead a steady expansion in the number of viable categories — CRM, he noted, barely existed as a category 20 years ago and is now massive
4. Venture's Middle Is Squeezed
The two firms that are thriving, on George's account, sit at opposite ends of scale: deep domain specialists at seed, and the handful of firms big enough to invest from seed through a public listing. He said a16z sits in the second group, funded in part by "a tremendous amount of resources" — "that's why we have 700 employees" — deployed to help portfolio companies win rather than only to write checks
Founders pick partners who they believe can de-risk the outcome, and that preference compounds. A firm that helps a company succeed gets referred to the next founder, which George called the flywheel behind persistent returns and behind LPs wanting into the fund in the first place
Pre-seed and seed specialists can coexist with the large multi-stage firms, Kha and George agreed, by moving a round earlier than the big funds are willing to commit — but only if they hold their fund size and strategy steady rather than creeping upmarket into direct competition
5. GPs and LPs Diverge
A GP is fireable for missing a winner; an LP almost never is for a bad pick. Kha put it as an asymmetry built into the two jobs: a general partner can be fired for missing the next Facebook or the next Uber. "Like that is the error of omission and like that is fireable." A limited partner, by contrast, keeps their job by tracking the benchmark, even from below it
"You don't get fired for investing in IBM if you're an LP," Kha said, framing the safe, unambitious choice most LPs default to
George's framework for an LP's job is access, selection and sizing. He said the same 20-of-3,000 concentration applies to LPs choosing managers, so a portfolio should hold 15 to 20 funds consistently rather than 50 to 70, where he said it becomes very hard to imagine the blended result beating the asset class average
He described a common mistake: an LP finds a strong fund, puts in only 1% of its own capital, gets a 10x return on it — and the win moves the overall portfolio by 10% of 1%, not enough to matter
6. Traction Is Hard to Read
AI has made judging a company's growth harder, not easier, in George's telling — rounds move faster and revenue signals are noisier. He described startups raising "off of that traction at huge multiples" on ARR figures with no renewal cycle yet behind them, some of it revenue generated by companies selling to each other inside the same accelerator cohort
He pointed to Cursor as the example everyone got wrong for months. "Even the morning of the acquisition announcement people were still saying that cursor is dead," George said — hours before "they just announced that they were going to be acquired by SpaceX for 60"
The company's own early metrics, on his account, were modest: "$3 million ARR, $400 million round or somewhere maybe around there"
His own portfolio company Harvey is the clearest example of the signal flipping. Early on, Harvey signed marquee law firms but usage stayed weak; after reasoning models improved, adoption took off. "It almost became a flip from what was previously like, oh, we're scared of things like hallucinations to, no, no, no, every client is actually demanding the law firms use the product," George said
His conclusion: "Come back to, you know, is the market demanding more of your product? Like that is always the question" — something he said only shows up in customer conversations, not in a cohort spreadsheet
7. AI Reprices Private Equity
A software company's growth rate now moves its valuation far more than its margin does. "Our data shows 1 percentage of growth in the public markets is equivalent to three percentages of EBITDA," George said — a reversal, he noted, from the COVID-era premium on profitability
Pre-ChatGPT-era software buyouts look expensive against that new math. George said 2021-to-2022 leveraged buyouts of software companies were priced at 25 to 32 times EBITDA against 200-plus billion dollars of debt; he said the same assets now trade closer to 2 times revenue in the public markets, which is pressuring the private-credit funds that financed them
Simply layering AI onto a private-equity-owned company doesn't work, on his account. "Just because you put Sears on a website didn't make it Amazon" — the fix has to be structural, not an easy win like hiring an AI customer-service agent, which he said can tank a company's Net Promoter Score and revenue if the underlying workflow isn't rebuilt
Intercom is his example of a company that took the harder path. "It's almost like you're suiciding your existing business, which in private equity is really hard to do," George said of the decision to bring back Intercom's founder and rebuild the product natively around AI
8. Liquidity Now Takes Longer
The average unicorn stays private more than 10 years, and an IPO doesn't deliver cash quickly even once it happens. George said it can take 12 to 24-plus months to get real liquidity out of a public listing, especially for a holder of 10% to 15% of the company at IPO
Given the choice, the firm has chosen to hold rather than sell. Asked whether they'd have sold Stripe or Databricks three or four years ago, George said, "the answer is unanimously no." He described a fund-one vote on Stripe: sixteen years into the fund, every LP asked said they'd rather let the position keep compounding than take liquidity, and a16z waited another year before finally exiting
He said the incentive varies by LP type: "Family offices, quite frankly, don't want the money back because they don't want to pay taxes on it. They'd rather have it continue to compound"
Venture's own exits now regularly beat private equity's. George compared this year's two largest private-equity buyouts — EA and Medline, both around $50 billion — against the $60 billion sale of Cursor to SpaceX
9. The Next $100T Company
Both guests said the next giant company is likely one that doesn't exist yet. George called a $10 trillion-plus outcome "probably two tech cycles away, not just one," and pointed to categories barely touched by AI: consumer products beyond a chatbot interface, robotics, autonomous vehicles, healthcare, and physical infrastructure
He rates robotics above the current language-model wave. "We are ... nowhere on robotics, but I think robotics is going to be bigger than the language stuff," George said, putting the timeline at roughly the next 10 years
Autonomy is still a rounding error. He said there are fewer than 10,000 Waymos operating in the entire US, with far fewer robotaxis from anyone else
Healthcare is the biggest untouched category by his own count. "Healthcare is 18% of GDP," he said, and argued AI has barely scratched either care delivery or drug discovery
Institutional allocators are still catching up to the last cycle, let alone the next one. George cited CalPERS converting its portfolio from 91% to 58% in fixed income and from 9% to 43% in venture and growth, after the pension fund's well-known losses from staying out of venture too long
Bonus Insights
Verdiyan closed by naming the constraint he thinks matters more than any single company bet. "The bottleneck in AI today. It's not demand, it's on the supply side" — he laid it out as energy and the grid, then data centers, then chips, then frontier models and apps, arguing the US is strong on the right side of that chain and weak on the left
"The US doesn't have a problem with energy generation. It has a problem with speed to power," he said, pointing to permitting and transmission as the actual bottleneck, and noted other countries are adding renewable capacity roughly 10 times faster per year
He does not think this is a repeat of 2000 or 2020. Responding to LPs who compare the moment to the dot-com bubble or the COVID run-up, Verdiyan said, "the traction is real and it's not ephemeral revenue like COVID"
A newer wrinkle George flagged: even AI-native private equity isn't automatically insulated, because portfolio companies that bolt on an AI customer-service layer without rebuilding the underlying workflow can lose customer trust and revenue quickly once a competitor does it properly
Verdiyan marked the 10-year anniversary of leaving a16z to become an LP by having Kha's team dig up an old branded keepsake from his time there — a small, human reminder of how long he has been on both sides of the table he and George were discussing
Products, Companies & Tools Mentioned
Andreessen Horowitz (George's firm; the source of the 3,000-firm, 20-firm return study behind the episode's opening claim)
Accolade Partners (Verdiyan's fund-of-funds, which allocates more than $3 billion into venture and growth managers including a16z)
Stripe and Databricks (The two companies George says the firm's LPs unanimously would not have sold three or four years ago)
Cursor (Anysphere) (The AI coding tool investors doubted for months before SpaceX agreed to buy it for $60 billion)
SpaceX (The acquirer of Cursor, and the company George says will likely go public soon after Anthropic)
OpenAI and Anthropic (Named alongside SpaceX as the three frontier companies Kha says represent $3.5 trillion to $5 trillion of potential enterprise value)
Harvey (The legal AI tool George cites as the clearest example of usage flipping from weak to explosive once reasoning models improved)
Intercom (George's example of a legacy software company rebuilding itself around AI after bringing back its founder)
Workday and Silver Lake (George's example of private equity getting more aggressive about reshaping software incumbents around AI)
Electronic Arts and Medline (This year's two largest private-equity buyouts, each around $50 billion — both smaller than Cursor's sale to SpaceX)
CalPERS (The pension fund George says is still catching up on venture and growth exposure after missing the last cycle)
Airtable (Cited as a company benefiting from the current cycle, including a planned spinoff of its AI agent product)
Waymo (George's example of how early autonomy still is — fewer than 10,000 vehicles operating in the US)
If this was worth your time, send it to someone closer to the industry than you are.
Get the latest market chatter as it happens: