Bill Maris opens with a self-introduction: founder of Section 32, previously the founder and CEO of Google Ventures, and Google's VP of Special Projects where he incubated Waymo, Google X, Calico, and many other projects. Before all that, he founded a web hosting and data center company. Today he wants to share four lessons he's learned along the way.
Going back to 1997: fresh out of college with a neuroscience degree, he somehow landed a job on Wall Street. He was miserable β wearing a suit, trudging to work in the heat. But one day he looked in the office closet and saw a server. He asked, "What is this thing beneath our jackets?" They said: "Well, that's where our email and websites live."
I had a moment where I felt like I was bathed in the light of inspiration. If you can have our website and email in your closet, how many websites and emails could I put in MY closet? I quit my job immediately. I had glimpsed the future through a keyhole.
He started a hosting business with credit cards. Eventually he had five servers. This wasn't a data center at all β it was his Vermont apartment. Servers lived in one room. Work happened in the other room. It got very hot in that room, so he opened the windows. Then it got very cold. So cold that by noon, a glass of water on the desk would ice over. And that was also his bed β a mattress with a Home Depot rug for warmth.
One day there was a thunderstorm and the roof started leaking. Water and servers don't mix. The landlord said "well, that happens sometimes." So Bill went to Home Depot, got a bucket of tar and a mop, and climbed up on the roof in the lightning and rain to tar it. He didn't know you're supposed to start at the far corner and work toward the door, so he tarred himself into a corner. His choice was either the servers get electrocuted or he gets electrocuted. As an entrepreneur, he took the risk. He survived. (His shoes are still stuck on that Vermont roof.)
To see the future, sometimes you need to be a little bit insane. It may appear to those around you that you were tarring the roof in the thunderstorm.
He shares slides his friend Stewart Butterfield gave him. 1989 inauguration: one person taking a film camera photo. 2005: not much different. Then 4 years later, another inauguration β everyone has a camera. But the most interesting figure in that photo is someone who, to his friends, must have looked insane: a man recording or live-streaming the inauguration on his laptop. He knew a secret about the future that those around him didn't believe.
"One of the things I've always looked for in entrepreneurs: they know a secret about the future that most of us don't believe."
Fast forward to 2007. Bill finds himself at Google. He's given a challenge: Google needs a venture fund. They've been making some investments, no coherent strategy, no budgets. He partners with Rich Miner, co-founder of Android, and they go up and down Sand Hill Road talking to anyone willing to talk.
The plan they come up with: obtain all the data of venture they can find β and being Google, that's a lot of data. Then use AI on it. But at the time, Google would not let them use the term "AI."
Bill, AI is science fiction. It's a hundred years away if it's ever going to happen. Let's stick to machine learning. When you say AI, it freaks people out.
So they called it machine learning. They used ML to do two things: design the ideal portfolio construction by running millions of simulations and backtests, and determine the ideal fund size. TechCrunch ran headlines. People inside Google got excited. Bill admits it seemed crazy at the time.
But over 2009β2018, top quartile and top decile VC returns played out a certain way. Using only publicly available information, Bill estimates Google Ventures' returns at about 4.1x. And the investments he personally led β adhering more closely to the strategy β turned out even better.
Don't bet against computer science. I've seen it happen many many times in many many fields. If you apply the right kind of computer science at the right time to the right problem, you will get to the right answers. I would not bet against it β even if it looks like you're tarring the roof in a thunderstorm.
Fast forward to 2017. Bill decides to start his own fund. Again, those around him said: "You're insane. Why would you do that? You're in the warm womb of Google. Lunch is free, the massages are plenty." Then the advice shifted: "Raise as much money as possible. You'll get a big management fee. You'll be happy."
He decided not to take that advice. Over his time at Section 32, they've had six funds, invested in companies like CrowdStrike, Cohere, and Coinbase. All six funds have averaged about $400 million in size and all are performing in their top decile.
To the extent there is DPI to measure β that's the only measure as far as I'm concerned in venture that counts. DPI.
Lesson four β this will be heresy to some, but small funds outperform large funds. This is simply the math. Not an opinion. There are many reasons: smaller funds allow more focus. Bill has already managed a multi-billion dollar fund with hundreds of employees β it's distracting. You can't give the attention to founders that you'd like to.
The data: funds smaller than $750M had average top-decile DPI return of 4.76x. Funds larger than $1B: 2.42x. Funds below $750M represented 95% of top decile performers, with discontinuous return compression above $750M.
If you have a $500M fund and you own 10% of a company, you need $5B of exits to get your money back. If you want to be in this business long-term and target 3x, you need to return $15B of exit value. Now if you have a $7B fund, you need to return $210B β which exceeds the total venture-backed M&A and IPO exit value in most years.
And remember: the 75th percentile of venture loses money. There is persistence of performance in the top quartile. Bill closes his four lessons.
The hosts pile on. Sacks recalls that when Bill started GV, Sacks's startup Yammer was the first ex-Google company GV invested in. Then Climate Corp β a billion-dollar exit to Monsanto. "Back then, a billion dollars was a lot of money."
Friedberg pushes back: doesn't the existence of these multi-trillion-dollar exits β billion to 10 billion to 100 billion to trillion β justify an alternative strategy? Maybe a barbell: small venture vehicles plus very large late-stage vehicles that bet on "sure things" with compounding advantage?
Bill responds with two observations. First, he hasn't seen the data science to support that the late-stage trend is an ongoing one, rather than a weird moment in time around these multi-trillion-dollar exits. Second β and this is where he gets sharp β if you're an RIA collecting assets, that is not venture. Venture, as he practices it, is concentrated bets of your time and capital on entrepreneurs, helping them build a business.
I have an objection to companies that wrap themselves up in public benefit language and then keep the value creation to themselves and an elite group of investors through a big part of the curve, and then say "well, we're here to benefit humanity." What humanity needs is money. It might be better to go public sooner.
And then comes the headline moment. Bill pivots to the AI pricing question:
If I'm Google β and I don't speak for Google β and I decide to arbitrarily cut the cost of tokens by 80%, I'm going to cut them in half again. What happens to the business models of OpenAI and Anthropic at that point?
The hosts pause. "What happens? Tell us." Bill lays out the mechanism: Google has the cash flow from advertising, owns its TPU infrastructure, and has every incentive to use capital as a weapon. OpenAI and Anthropic would have to follow on price β but they're already burning enormous capital relative to revenue. The margin compression would hit a critical point.
Sacks adds the math: Anthropic has committed to roughly $1 trillion in capex, against revenue around $6 billion. OpenAI is in a similar capex-vs-revenue gap. Token deflation closes that gap from the wrong direction.
The conversation turns to who actually ends up holding the bag when these late-stage AI companies finally go public. Bill is direct: by staying private longer, the value-creation curve is captured by an elite group of pre-IPO investors. By the time the company hits the public markets β and S&P 500 inclusion rules often let it into index funds quickly β the ordinary 401k holder buys at the top.
Sacks reinforces the structural critique: there's a mismatch between $1T capex commitments and $6B revenue. Even if model performance keeps improving, the unit economics depend on token prices NOT falling. And Bill's whole point is that Google has every motive to make them fall.
It might be better to go public sooner because we'll see how these multi-trillion-dollar IPOs go. The lockup expiration six months after IPO is going to be very telling.
On the VC side, this creates a bimodal market: a small number of funds (Founders Fund's $200M-into-$100B-paper-return type stories) print spectacular numbers, but the average and the median tell a very different story. Paper gains have to be realized eventually, and the public market will reprice on future cash flow, not on the last private round.
This is why Bill comes back to DPI as the only metric that matters. TVPI is a marketing number.
Asked where AI is on its development curve, Bill offers his favorite metaphor of the conversation.
We're at the Atari command-line stage of AI. Think Zork, Planetfall β text adventures in the early 1980s. The current tools are brittle. They forget. They lose context. They reset. All of that gets solved. The gaming industry took 40 years to go from Zork to PlayStation 5. AI will compress that into about 5 years.
Critically: he is NOT investing in the large foundation models themselves. Capitalism's leveling pressure will commoditize that layer β which is exactly what the Google-80%-token-cut scenario foreshadows.
What he IS investing in: the infrastructure layer. "Just like the gaming industry's leap wasn't really about better stories β it was about GPUs and physics engines β AI's leap is about infrastructure, not models." Physics engines, controllers, GPUs, ambient computing.
He's also bullish on deep tech in the Elon sense. AI-enabled simulation and physics engines make previously intractable deep tech bets newly tractable. SpaceX-style and Neuralink-style problems become more solvable when you can simulate at human-cell or planetary-scale.
Human biology and healthcare, Bill argues, is the largest TAM in the world. Every human ends up a patient. He mentions New Limit β the longevity company co-founded by Blake Byers and Brian Armstrong β as an example of what excites him.
I avoid therapeutics that depend on running clinical trials β the timelines and binary outcomes are brutal. What I'm excited about is computational biology. The moment we can do in-silico simulation of human cells, the risk-adjusted return profile of bio investing changes fundamentally.
The final theme of the conversation turns darker. Bill expresses deep concern about the weakening of the CDC and NIH, the broader anti-science mood, and pressure on the H-1B visa program. The combination is eroding America's traditional ability to attract the world's best scientific brains.
China is actively recruiting top scientists from Europe and India. India and Europe are increasing R&D funding. The pipeline that brought the world's best minds to the United States isn't natural β it was a conscious choice. And that choice is now being reversed.
Bill points to his own immigrant family background and warns that if the U.S. continues on this trajectory, American leadership in AI, bio, and every other frontier science will erode faster than people realize. The episode ends on that note β a reminder that the capital-and-tokens story Bill opened with is downstream of a much bigger question about who gets to do science, and where.