Bill Maris: How Google Could Crush AI Competitors, Why Small Funds Win, and AI's Atari Stage

Why does it matter? Because every assumption holding up today's AI market has a timer on it.

Bill Maris built Google Ventures, then spent the years since running six funds averaging $400M each — all in the top decile. He shows up here not to celebrate but to systematically dismantle the math behind how the AI era's biggest bets are being constructed.

• Google can erase OpenAI and Anthropic's business models overnight by cutting token prices 80% — and Maris considers it rational, not hypothetical
• A $7B VC fund needs $210B in exits just to return 3x, a number that exceeds total venture-backed M&A and IPO exit value in most years
• AI today is Zork — brittle, stateless, turn-by-turn — and the money is in physics engines and controllers, not larger models
• Late-stage AI IPOs at trillion-dollar valuations will force overpriced shares onto passive index funds and 401k holders who missed the entire private upside

Google can kill OpenAI and Anthropic's business models with one pricing decision — and Maris says it's inevitable

Google has the weapon loaded. "If I were Google, that's what I'd do," Maris says, after walking through the logic: cut token prices 80%, offer what is effectively an identical product through Gemini, and watch the pressure hit "super critical" on OpenAI and Anthropic. Every enterprise customer reprices. Every startup built on the OpenAI API recalculates. The business model collapses.

When the hosts push on whether this is speculative, Maris barely hesitates. "It's clear they're going to do it." The only open question is the mechanism — whether Google uses its own margin or burns investor cash Uber-style to grab install base and market share before monetizing. Either path lands in the same place: tokens as a weapon, competitors as the target.

The entire private AI ecosystem is implicitly priced on the assumption that frontier model providers can defend their margins. That assumption has a single actor who can shatter it unilaterally, owns the distribution, and has every rational incentive to pull the trigger. The question isn't if. It's whether the trillion-dollar valuation holds before or after.

A $7B fund needs $210B in exits to return 3x — that number exceeds total venture-backed M&A and IPO value in most years

Run the arithmetic. A $500M fund, assuming 10% average ownership, needs $5B in exits to return capital. Set a 3x target: $15B. That's a real number. Scale to a $7B fund and the same math produces $210B — "which exceeds the total venture-backed M&A and IPO exit value in most years," Maris says. Not a bad year. Most years.

The empirical data is equally stark. Funds below $750M averaged 4.76x returns. Funds above $1B averaged 2.42x. Sub-$750M funds represented 95% of top decile performers, with "discontinuous return compression above $750 million." Not a gradual slope — a cliff.

Maris built Section 32 on exactly this logic, rejecting the advice he got when leaving Google, which was to raise as much as possible. Six funds, averaging about $400M each, all top decile. The model was built to clear the math that megafunds structurally cannot. LPs allocating to billion-dollar-plus VC vehicles are buying into a structure that can't produce venture-class returns — the brand names just make it easier to write the check anyway.

VC incentives are broken at all three levels simultaneously — and the distortion compounds

Start with the LP. A $5 billion venture fund that returns 1.01x gets to claim it's in the 75th percentile of the industry. "No one at the Stanford endowment is going to get in trouble for writing that check." The LP is structurally protected from embarrassment by the mediocrity of the benchmark itself.

Then the GP. "If I have a $5 billion fund, I return 1.01x, I'm going to make more money than Bill with his $500 million fund that returns 3x." Two percent of five billion beats two percent of five hundred million regardless of performance. Growing the fund isn't about returns — it's about the fee structure.

Then the founder. A researcher leaves OpenAI. Maris offers $20M at a $100M valuation for 20%. Giant Fund Y counters: $250M, and your company is now worth $4 billion. "They're going to take that deal every day," Maris says flatly, "unless you're a seasoned entrepreneur who has been down the road and knows the pitfalls." The founder gets an inflated headline number. The megafund gets the check size it needs to justify its own existence.

All three distortions point in the same direction: capital deployment over returns. The machine is working exactly as designed — just not for the people who think they're the beneficiaries.

AI is at the Zork stage — the real value is in physics engines and controllers, not bigger models

Turn response. Turn response. Grab the lamp. "Oh, I didn't — it's a lantern. I should have said lantern." Maris uses Zork, the brittle 1980s text adventure, to locate exactly where AI sits today: stateless, session-resetting, turn-by-turn, completely unforgiving of imprecision. Put the most sophisticated retail AI system next to it, he says — tell him how different it looks.

What transformed gaming from Atari command lines to photorealistic, inhabitable worlds wasn't better stories. It was controllers, physics engines, and GPUs. The infrastructure layer. That leap took four decades in gaming. Maris expects AI to compress the same arc into five years.

"I think we're at the Atari command line stage of AI and we're going to get to the PlayStation 10 stage in the next 5 years."

So he's not investing in larger models. The bet is on the machinery underneath: memory systems, consistency engines, ambient computing platforms — whatever solves what Zork couldn't. Persistent memory. Stable context. No session resets. Those are engineering problems, not model-size problems. The companies building that stack are where he sees value accumulating — and where frontier model incumbents have no moat.

A trillion in commitments on $60 billion of revenue — and the public's 401k is the designed exit

"We're going to force overpriced products on the 401k holders of America who didn't get to participate early." Maris lays the mechanics out plainly: companies stay private through most of their value curve, insiders and a small group of elite investors capture the compounding upside, then the IPO hits at trillion-dollar valuations that passive index funds and ETFs are structurally required to absorb.

The ratio is concrete: a trillion in spend commitments sitting on $60 billion of revenue, heading into public markets. And for any late-stage fund to convert paper gains into real ones — "in order for Founders Fund or pick any fund to get that $100 billion out, they have to sell that stock to someone else" — there has to be a buyer. Who is it? Retail investors. Pension holders. Index fund buyers. People who had no access to the private rounds and no ability to opt out of the IPO once it joins the S&P.

Maris's objection isn't to the structure per se. It's to the framing layered on top of it. "My objection is don't say you're doing this for the benefit of humanity and do the other thing." The public benefit language is a rhetorical shield. The structure it covers is extractive. When valuations meet public markets, scrutinize the gap between those two things.

The US is actively exporting its scientific brain trust to China — and the policy pipeline is accelerating it

China is running a recruitment campaign right now. "They're recruiting some of the best scientists from Europe and India, and they're all immigrating to China." The talent pool that once flowed toward US universities and national labs is redirecting — not only because China is pulling, but because American policy is doing the pushing.

The NIH and CDC are being gutted. Basic research funding is drying up. An "anti-science vibe" now "pervades this country," Maris says. Layer on H1B restrictions, and the calculation for a researcher who might once have come to the US becomes simple arithmetic. "It's just easier to go elsewhere. That's not good for science."

Scientific talent doesn't restock in a budget cycle. Researchers who leave take institutional networks, research directions, and decades of embedded knowledge with them. The biotech and deep tech investment thesis has always been built on US talent dominance as a baseline assumption. That assumption is now being actively dismantled through a combination of funding cuts, political hostility to research, and visa policy — and the compounding effect is a generation-level shift, not a quarterly variance.

Discovering a drug compound is 5% of the work — the FDA bottleneck means AI won't make biotech exponential yet

Faster target identification is real. It doesn't move the bottleneck. "If you find a compound and you think you've got something, that's like 5% of the work." The remaining 95% — titrating, safety testing, human clinical trials — runs on FDA timelines that no model can compress.

Maris isn't dismissing AI in biotech. He's calibrating. "I don't think it's going to go quite as exponential as we would all like it to." The hype treats discovery as the constraint. It isn't. Human safety regulation is the constraint, and it's there for reasons that don't dissolve because a model found a candidate faster.

The unlock he's actually watching for is specific: "If we can achieve a realistic simulation of a human cell in silicone, then you will see that accelerate as well. We're not quite there yet." In-silico cell simulation that makes early-stage safety testing computational rather than biological — that would actually change the timeline. Not faster screening. Fewer wet-lab steps required before clinical trials begin. That's the precise breakthrough to track, and it's still a research problem, not a deployment one.

The next reckoning is between the story and the arithmetic

Every major claim Maris makes traces to the same observation: the prevailing narrative about AI value, venture returns, and scientific dominance hasn't run the numbers. $210B exit requirements. 80% token price cuts. Trillion-dollar IPOs landing on passive investors. Brain drain measured in immigration statistics. None of these are predictions. They're math.

The story keeps running until the arithmetic forces a correction — and when it does, the people who built their positions on the narrative rather than the numbers find out all at once.

---

Topics: venture capital, AI infrastructure, fund size, Google, OpenAI, Anthropic, token pricing, biotech, brain drain, IPO, VC incentives, Section 32, Google Ventures, deep tech