Bill Maris: How Google Could Crush AI Competitors, Why Small Funds Win, and AI's Atari Stage
In a Nutshell
Bill Maris argues small venture funds outperform large ones because they can make concentrated bets and deliver higher DPI returns—data shows funds under $750M averaging 4.76x versus 2.42x for those over $1B. He warns Google could crush AI competitors like OpenAI by slashing token prices 80%, using capital as a weapon to grab market share. Maris sees current AI as being in its "Atari stage" and predicts rapid progress toward ambient computing, with his Section 32 fund focusing investments on enabling infrastructure like GPUs and physics engines rather than model development.
These notes were generated by AI and may contain inaccuracies.
Bill Maris, founding CEO of Google Ventures, has raised $150 million for his new fund called Section 32. With a smaller fund, he has the advantage to be very selective in the companies invested in and the people hired. The fund will invest for a financial return because any other metric is impossible to measure and therefore won't succeed.
Bill Maris founded and was CEO of Google Ventures. He was also Google's vice president of special projects where he incubated Whimo, Google X, Calico, and many other projects. Before that, he founded a web hosting and data center company. He will discuss four lessons from his experiences, beginning in 1997.
After graduating college with a degree in neuroscience, Maris worked on Wall Street but was miserable wearing a suit. He discovered a server in the office closet and learned it housed email and websites. This sparked inspiration that he could build a business around hosting websites and email. He quit his job immediately, glimpsing through a keyhole what he thought was the internet and a data center.
In 1997, a state-of-the-art data center consisted of three servers: small, medium, and large. The business grew to five servers. This "data center" was actually Maris's apartment in Vermont, founded with credit cards. Servers occupied one room; work happened in the other. The room would get extremely hot, so windows were opened, making it so cold that water glasses would ice over by noon. Maris slept there on a Home Depot rug for warmth.
During a thunderstorm, the roof began leaking. Maris called the landlord who dismissed it as normal. He bought tar and a mop at Home Depot and went onto the roof in lightning and rain to fix it. He did not know to start at the far corner and work toward the door, so he tarred himself into a corner. The choice was whether the servers or he would get electrocuted. As an entrepreneur, he took the risk and survived, though his shoes remain 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 are tarring the roof in a thunderstorm. Maris shared photos from Stuart Butterfield showing the 1989 inauguration with someone taking a picture, the 2005 inauguration similarly, and the 2009 inauguration where everyone had cameras. One man was live-streaming or recording on his laptop, knowing a secret about the future that most others did not believe.
In 2007, Maris found himself at Google tasked with creating a venture fund with no coherent strategy or budgets. He partnered with Rich Miner, co-founder of Android. They spoke with everyone on Sand Hill Road willing to talk. Their plan was to obtain all available venture data and use AI (called "machine learning" because Google would not allow the term AI, saying it was science fiction a hundred years away) to design ideal portfolio construction through millions of simulations and back-testing, and to determine ideal fund size.
From 2009 to 2018, top quartile VC returns and top decile returns were compared. Using publicly available information, Google Ventures returns were estimated at about 4.1x. Investments Maris led personally performed even better. This demonstrated Lesson Three: don't bet against computer science. Applying the right computer science at the right time to the right problem yields the right answers.
In 2017, Maris decided to start his own fund despite advice that he was insane to leave Google's "warm womb" of free lunch and massages. Others advised raising as much money as possible for large management fees. He rejected this. Over time at Section 32, six funds were raised, investing in companies like Crowdstrike, Coher, and Coinbase. All six funds averaged about $400 million in size and performed in the top decile. The only measure that counts in venture is DPI.
Lesson Four: small funds outperform large funds. This is simply the math, not an opinion. Smaller funds allow more focus and attention to founders. Maris previously managed a multi-billion dollar fund with hundreds of employees, which was distracting.
Top decile performance of DPI shows funds smaller than $750 million averaged 4.76x return, while funds larger than $1 billion averaged 2.42x. Funds below $750 million represented 95% of top decile performers, with discontinuous return compression above $750 million.
With a $500 million fund owning on average 10% of companies, $5 billion in exits are needed to get money back. The 75th percentile of venture loses money, with performance persistence in the top quartile. To achieve a 3x goal requires returning $15 billion in exit value. A $7 billion fund would require returning $210 billion, exceeding total venture-backed M&A and IPO exit value in most years.
A large fund strategy focused on later-stage rounds of well-proven companies was contrasted with smaller funds. Maris noted he had not seen data science supporting that latest-stage companies represent an ongoing trend beyond this unusual moment with multi-trillion dollar exits. At a certain point, collecting assets that are not venture becomes a different craft from making concentrated bets of time and capital on entrepreneurs.
Maris raised an objection to companies wrapping themselves in public benefit language while keeping value creation to an elite group of investors through much of the curve, then claiming to benefit humanity. What humanity needs is money, so going public sooner might be better. He questioned how multi-trillion dollar IPOs will perform.
If Google decided to arbitrarily cut the cost of tokens to 80%, what happens to the business models of OpenAI and Anthropic? Companies would pay 80% less for a basically identical product from Google Gemini, causing super-critical compression and pressure on those businesses. Maris stated this is the rational thing for Google to do. Capital as a weapon and tokens as a weapon can grab market share and install base, though cash generation is ultimately required.
Forcing overpriced products on 401k holders who did not get to participate early makes retirement accounts the bag holders. There is significant risk in this approach. Maris's objection is against claiming actions are for the benefit of humanity while doing the opposite.
A handful of funds will have ginormous returns, such as Founders Fund potentially printing a $100 billion return on $200 million of invested capital. However, venture is extremely bimodal: a handful of winners and a ton of losers. 75% of funds lose money. To realize paper gains, stock must be sold to someone else, requiring a public market business case based on discounted future cash flows. It remains to be seen what the public market thinks of companies like SpaceX or Anthropic.
Maris made an analogy to the gaming industry. Early games like Zork and Planetfall from the 1980s were brittle with turn-based responses and strict command requirements. The most sophisticated retail AI system today looks remarkably similar. The transformation that occurred in gaming from the 1980s to today will happen in AI over the next 5 years, compressed in time. Games became photorealistic and inhabitable; AI will achieve ambient computing, solving problems of memory, consistency, and session resets.
Maris does not plan on investing in larger models. Just as better games came from controllers, physics engines, and GPUs rather than better stories, he is interested in the platforms and machinery needed for AI: physics engines, controllers, GPUs, and everything required to make ambient computing reality. AI is currently at the Atari command line stage and will reach the PlayStation 10 stage in the next 5 years.
Maris founded Calico and has been interested in life sciences, investing in Flatiron, Verily, and others. The dual benefit of helping people and doing well financially is attractive. However, therapeutic spaces requiring human clinical trials are specialist areas not currently emphasized. Interest lies in computational biology. While there appears to be a renaissance in human health with cancer vaccines and peptides, current class winners were driven by good science from 10 years ago, not computation. Maris hopes computational approaches will drive massive advances but notes human biology and FDA requirements mean finding a compound is only about 5% of the work; titrating and safety testing remain. Realistic in silico stimulation of human cells would accelerate progress, though this is not yet achieved.
The US indexes on human safety over speed to market, which has costs. Other countries index in the opposite direction, which also costs lives. Research, experiments, and cloning are happening in China and elsewhere that are not occurring in the US. The gutting of the CDC and NIH and an anti-science vibe have driven mind share elsewhere as funding for basic research dries up. China now has its own paperclip model, recruiting top scientists from Europe and India. The US is losing access to this scientific talent pool. Pushing out H1B holders makes it easier for talent to go elsewhere, which is not good for science.
Deep tech areas with long investment cycles, capital intensity, and high risk are becoming more tractable due to AI enablement and physics engines. Elon Musk has successfully tackled deep tech with SpaceX and Tesla. Areas of focus include human biology and healthcare as the largest TAM, plus all underlying components of the AI revolution: physics engines, controllers, GPUs, and enabling infrastructure.
Fund size determines fund strategy. Dividing fund size by 20 to 25 names achieves portfolio diversification and determines check size, which in turn determines where a fund plays in the market.
After Thomas's presentation, the question arises whether it is better to focus on late venture or early growth, writing $50 million checks and waiting for breakouts, rather than playing in the noisy super early stage game.
The incentive structure of venture capital creates problems for the late-stage approach. A $5 billion venture fund that returns 1.01x can claim to be in the 75th percentile and raise its next fund. Stanford endowment LPs face no repercussions for writing such a check, as they need to deploy two or 500 million into a fund multiple times.
On the GP side, a manager of a $5 billion fund returning 1.01x will make more money than a manager of a $500 million fund returning 3x.
On the entrepreneur side, a researcher from OpenAI starting a company faces distorted offers. A $500 million fund might offer $20 million at a $100 million valuation for 20% ownership. A giant fund needing to deploy $250 million will instead push valuation to $4 billion to put that capital to work, and most entrepreneurs will accept unless they are seasoned and understand the pitfalls.
The incentives are broken across LPs, GPs, and entrepreneurs. The pendulum will eventually swing back. Staying late stage and attempting to snipe larger companies will not work long term. The data suggests this approach will not succeed over time.
Keep All-In Podcast in your library
Save the videos and channels worth coming back to, and find them again in one place.
