Chasing Trillion-Dollar Companies, Founder Ambition, Token Budgets, & Regulatory Capture

2026-08-06 - 39 min - source - Read full transcript

Sarah Guo (host)Elad Gil (host)

Key insights

Only a handful of new trillion-dollar companies will likely emerge in the next three to five years, despite widespread investor belief that many more are coming.

Anthropic, OpenAI, and SpaceX went from near-zero to a trillion dollars in market cap in about five years, compared to the usual 15-20 year arc (Google, older tech giants). That inflection is being read as a new normal, but reaching a trillion dollars requires 50-100 billion dollars of revenue at good margin, which only a very small number of markets can plausibly support in that time frame - most 'huge TAM' stories will produce 20-100 billion dollar outcomes, not multi-trillion ones.

trillion-dollar-companies

Trillion-dollar company formation happens in punctuated bursts tied to technology waves (social, SaaS, crypto, AI), not as a smooth continuous process.

Each wave produces a handful of category-defining 'consolidator' companies and then settles into a steady state until the next platform breakthrough. The AI wave has already produced some consolidators; the open question is how many more giant companies form in the current burst versus over a full 20-year horizon.

trillion-dollar-companies

Investors intellectually endorse AI outcome-based pricing but still underwrite deals using old per-seat SaaS logic.

Firms will say they believe an AI company like Harvey could charge for outcomes rather than per-lawyer seats, but their models and diligence still default to linear, per-seat TAM math. This gap between stated belief and actual underwriting behavior is called out as a key mispricing in the market, especially in coding, where consumption and value captured look like they could be 100x larger than the old per-seat framing suggested.

trillion-dollar-companies

A growing share of otherwise strong founders are retreating into niche markets out of fear of the frontier labs, rather than competing head-on.

Where founders three to five years ago (Harvey, Open Evidence, Decagon, Sierra, Cognition) went after big markets a lab could also enter, a newer trend line shows good founders choosing smaller, more derivative niches - work a lab would never bother with - instead of building a better product and out-competing on distribution. This is framed as founders being less ambitious than they could be, not as a rational response for most of them.

founder-ambition

Whether to sell a company should be a scheduled, unemotional board-level question, not a one-off crisis decision.

A handful of companies (Anthropic, OpenAI) should never sell in the near term, but most companies have a 12-18 month window where they're worth the most they'll ever be worth. The suggested practice, credited to Ben Horowitz, is to pre-schedule a recurring board discussion (ideally every six months in the current AI cycle, more frequent than a few years ago) asking directly whether the company should consider an exit in the next six months, so the answer isn't driven by founder or investor emotion in the moment.

founder-ambition

The largest hidden cost of staying too long in a struggling, overcapitalized company is a founder's most productive years, not just capital.

Founders from the 2020-2021 vintage are cited as still running companies years later that aren't working, having been effectively locked up through the entire AI transition. Because each year of AI-era progress compresses roughly three to four years of normal-era progress, the opportunity cost of staying in a stalled company is judged to be much higher now than in past cycles - and secondary sales are called an unsatisfying middle option that solves short-term liquidity without resolving the underlying problem.

founder-ambition

Physical compute scarcity, not algorithmic limits, is the binding constraint on AI progress right now, and it effectively enforces an oligopoly among frontier labs.

Because compute is allocated roughly pro rata across the ecosystem, the ceiling on how fast any single lab can progress is set by physical compute access rather than by ideas. This forces closer competitive parity between the major labs than would otherwise exist, and the key open question is when compute constraints ease and what the landscape looks like once they do.

ai-research-talent

A small number of researchers, on the order of a few dozen per lab, drive the large majority of results, and labs are increasingly allocating compute and hiring toward that power-law tail.

This mirrors a broader pattern seen in breast cancer research, subfields of math and physics, and startup founders generally: a handful of people account for most of the progress in a field. As compute becomes the scarcer resource relative to headcount, some labs have raised their hiring bar sharply, since the real cost of a researcher is the compute allocated to them, not their salary - motivating a 'return on invested tokens' framework for deciding who gets outsized compute budgets.

token-economics

Belief that recursive self-improvement (RSI) is roughly 18 months away is driving manic overwork and burnout-adjacent behavior among AI researchers, comparable to how people behave when they think they're dying.

Researchers at major labs have reportedly asked whether they should get married given uncertainty about the world in 18 months, and some feel their contribution is becoming irrelevant. The '18 months away' prediction has recurred every 18 months for roughly five years, making it a weak predictor even though the code-to-training-data extension of AI capability is judged easier to believe than the harder problem of gathering verifiable data in more complex domains.

ai-research-talent

As AI reduces the need for less-productive engineers at top-tier tech companies, that displaced talent is likely to flow into traditional, non-tech enterprises that previously couldn't recruit it.

Engineers who are only middling by Google or Meta standards can still be exceptional relative to an 'old school' enterprise like GE, PG&E, or Hershey's that never had the brand or comp structure to hire top-tier engineering talent. This is framed as a likely multi-year trend rather than an imminent mass layoff event.

token-economics

California's proposed billionaire tax and talk of a companion exit tax are already accelerating founder and company migration out of the state, echoing prior regulation-driven shifts in tech hub geography.

The billionaire tax, backed by the state's Democratic Party, would force asset sales for founders of companies worth over ten billion dollars if it passes; a parallel exit tax under discussion would penalize people who try to leave. Migration is already happening and is compared to how Boston lost its 1980s position as Silicon Valley's main counterweight, and to the newer hardware and energy corridor emerging in Texas (originally centered on El Segundo, now including Austin) as a direct reaction to regulatory environment rather than lifestyle factors.

regulatory-capture

Historical regulatory capture in pharma and energy shows the cost of weighing safety without weighing benefit, and AI risks repeating the pattern.

Paul Janssen's account of FDA history argues regulatory capture and a risk-only (not risk-reward) posture made drug development far slower and more expensive. France gets 70 percent of its power from nuclear with essentially no major accidents, while the US sits at 18 percent and hasn't built a new reactor in 40 years because a 1970s safety lobby effectively killed abundant clean energy. The parallel drawn for AI: focusing only on downside risk (e.g., email getting hacked) while ignoring upside (e.g., faster healthcare breakthroughs) risks the same multi-decade stagnation seen in pharma and nuclear energy.

regulatory-capture

Media referenced

Paul Janssen YouTube interviews on regulatory capture in pharma - other - Cited as evidence that FDA safety-only focus (risk without weighing benefit) slowed and raised the cost of drug development; used as an analogy for how AI regulation could go the same way.

Companies

Anthropic - Cited as one of three companies (with OpenAI and SpaceX) that went from near-zero to a trillion dollars in market cap in roughly five years.
OpenAI - Second of the three trillion-dollar-in-five-years companies; also named as a company that should never sell.
SpaceX - Third of the trillion-dollar-in-five-years companies; contrasted with the ~15-20 year arcs of prior giants like Google.
Google - Reference point for how long (15-20 years) it used to take to build a trillion-dollar company; also the subject of an old joke about how few employees actually drive the company's results.
Harvey - Legal AI company used as an example of investors still pricing AI value on a per-seat/per-lawyer basis instead of outcome-based value.
Open Evidence - Named among founders who went after markets the frontier labs could also target, illustrating aggressive rather than niche positioning.
Decagon - Named alongside Sierra and Open Evidence as founders competing in markets that could be in a frontier lab's roadmap.
Sierra - Same point as above, cited as an example of a company entering a market a lab could also pursue.
Cognition - Cited as a portfolio company that entered a big, contestable market roughly two years ago.
Cursor - Referenced in the discussion of financing structure and matching capital access to competitive position.
Minecraft - Cited as an early real example of a tiny team (five to ten people) producing a multi-billion-dollar outcome, predating today's AI-driven one-person-company narrative.
Microsoft - The acquirer of Minecraft, referenced in the same point about small teams and outsized outcomes.
Tesla - Mentioned as part of the hardware/energy corridor shift toward Texas.
Anduril - Cited as part of the growing hardware corridor originally centered around El Segundo before shifting toward Texas.
Conviction - Sarah Guo's venture fund; its grant program (embed.conviction.com) is promoted mid-episode, offering $250,000 on an uncapped note plus compute and partner services.
Chai Discovery - Named as one of Conviction's grant program cohort companies.
Listen Labs - Named as one of Conviction's grant program cohort companies.
Physical Intelligence - Named as one of Conviction's grant program cohort companies.
Base10 - Named as one of the partners providing compute and services to Conviction's grant program.
GE - Cited as a likely destination for engineering talent displaced from Google/Meta-tier companies as AI reshapes staffing.
PG&E - Same point, cited as an 'old school' enterprise that could absorb displaced big-tech engineering talent.
Hershey's - Same point, another example of a traditional enterprise that could newly access top engineering talent via AI-driven displacement.

Techniques and frameworks

Punctuated equilibrium (evolution analogy applied to tech waves) - Used to explain why trillion-dollar-company formation happens in bursts (social, SaaS, crypto, AI) followed by consolidation rather than as a smooth continuous process.
Return on invested tokens (ROIT) - Proposed framework for deciding which projects and people should receive outsized shares of a compute/token budget, treating tokens as a scarce resource to be allocated for maximum return rather than distributed evenly.
Pre-scheduled, non-emotional exit-consideration board meeting - A Ben Horowitz-derived practice: put 'should we consider exiting in the next six months' on the board calendar in advance so the question gets asked rationally rather than only in a crisis or under investor/founder pressure.

Summary

Sarah Guo and Elad Gil, the two hosts of No Priors, use this episode as a two-hander conversation (no outside guest) to work through how many more trillion-dollar companies AI will actually produce, and conclude the number is small. Three companies - Anthropic, OpenAI, and SpaceX - went from near-zero to a trillion dollars in roughly five years, compared to the usual 15-20 year climb of prior giants like Google. They frame this less as a permanent new normal and more as a punctuated-equilibrium moment: technology waves (social, SaaS, crypto, AI) produce a burst of category-defining "consolidator" companies and then settle into a steady state. Getting to a trillion dollars requires 50-100 billion dollars of annual revenue at good margin, which only a handful of markets can plausibly deliver in the next three to five years - most "huge TAM" AI companies will land in the 20-100 billion dollar range, still enormous, but not trillion-dollar territory. A related mispricing they flag: investors say they believe AI companies can charge for outcomes rather than seats, but still underwrite deals with old per-seat SaaS math, missing how much bigger markets like coding actually are once consumption-based pricing is taken seriously.

The conversation's second thread is founder ambition and exit timing. Both hosts observe a trend line, concentrated among otherwise strong founders, of retreating into niche markets out of fear of the frontier labs rather than competing head-on and out-executing on product and distribution. On exits, they converge on a shared framework: a small set of companies (Anthropic, OpenAI) should never sell, but most companies pass through a 12-18 month window where they're worth the most they will ever be worth, and the decision to consider selling should be a pre-scheduled, unemotional board conversation (a practice credited to Ben Horowitz) rather than a reactive one. Because a year of AI-era progress compresses roughly three to four years of normal-cycle progress, that check-in cadence has tightened from roughly once a year to roughly every six months. The real cost of staying too long in an underperforming, overcapitalized company isn't just capital - it's a founder's most productive years, illustrated by 2020-2021-vintage founders still running companies years later that aren't working, having been locked up through the entire AI transition.

A third thread digs into compute and research talent economics. Physical compute scarcity, not algorithmic limits, is described as the real binding constraint on AI progress, and because compute is allocated roughly pro rata across the ecosystem, it effectively enforces an oligopoly and closer competitive parity among frontier labs than would otherwise exist. Within labs, a power-law pattern applies: a few dozen researchers drive the large majority of results (a pattern the hosts compare to breast cancer research, subfields of math and physics, and startup founders generally), pushing some labs to raise hiring bars sharply since the real cost of a researcher is the compute allocated to them. This motivates a "return on invested tokens" (ROIT) framing for deciding who gets outsized compute budgets, and a related, more speculative prediction: as AI reduces need for less-productive engineers at top-tier tech companies, that talent will likely flow toward traditional non-tech enterprises (GE, PG&E, Hershey's are named) that never had the brand to recruit top engineers before.

The hosts also discuss the psychological toll of believing recursive self-improvement (RSI) is roughly 18 months away - a prediction that has recurred on an 18-month cycle for about five years, making it a weak predictor even as it drives real burnout-adjacent behavior, including researchers reportedly asking whether they should get married given uncertainty about the world in 18 months. The episode closes on regulatory capture, tying together a California case study (a proposed billionaire tax that would force asset sales for founders of 10-billion-dollar-plus companies, plus talk of a companion exit tax) with historical analogies to pharma and nuclear energy. Paul Janssen's account of FDA history is cited as an example of regulation that weighed safety without weighing benefit, slowing drug development for decades; France's 70 percent nuclear power generation with essentially no major accidents, versus the US's 18 percent and 40 years without a new reactor, is offered as evidence that a 1970s safety lobby - not the technology itself - killed abundant clean energy in America. The hosts argue AI regulation risks the same pattern if downside risk (e.g., an email getting hacked) is weighed without also weighing upside (e.g., faster healthcare breakthroughs), and close with a case for keeping AI lightly regulated relative to how heavily other industries have been regulated historically.

Notable Quotes

Note: the source transcript (podscripts ASR) carries no speaker diarization. Attribution below is inferred from context (the Conviction ad-read, greeting order, and known investing focus of each host) and should be treated as best-effort, not verified against audio.

"We had three companies roughly go from close to zero to a trillion dollars in market cap, right? ... that's unprecedented in human history. Usually it takes 20 years." - Elad Gil

"There's a handful of companies that should never ever sell, at least any time in the near term. If you're Anthropic, you shouldn't sell. If you're OpenAI, you shouldn't sell." - Elad Gil

"Your most productive years of your life are on the line right now. And you can either walk away with a good amount of money... or you can roll the dice." - Elad Gil

"The physical compute basically reinforces an oligopoly market because... it creates a ceiling on the rate of progress any single lab can get." - Sarah Guo

"We had a safety lobby in the 70s basically kill abundant clean energy for us." - Sarah Guo