Three Trillion-Dollar Companies in Five Years. Elad Gil Would Not Bet on a Fleet of Them Next

Trillion-dollar companies used to arrive like buses on a 15-to-20-year timetable. Google’s arc started in the 1990s. SpaceX’s started in the early 2000s. Then the last five years produced three of them at once. Anthropic barely existed. OpenAI was still early; GPT-3 had just come out. SpaceX was trading around $80 to $100 billion. All three now sit near a trillion dollars of market cap. On this episode of No Priors there is no outside guest—just the two hosts, Conviction founder Sarah Guo and investor Elad Gil, talking through how many more of those buses the industry should expect.

A lot of people, Gil says, are now assuming a bunch of other trillion-dollar companies will form in three to five years. That would be unprecedented. He can think of one name that might get there. He will not say which. He is sure it will not be multiple. Robotics, materials, and other exciting fields may produce giants over a decade. They are unlikely, in his view, to produce a fleet of trillion-dollar outcomes on the compressed clock the last vintage just ran.

The conversation that follows is really four arguments at once: how many multi-trillion outcomes are still available, when founders should sell, a manic 18-month mood inside the labs as recursive self-improvement is talked up, and the damage done when regulation scores only risk. Guo’s aside for new founders is that Conviction’s Embed grant is open again—$250,000 on an uncapped note, plus compute from partners including OpenAI and Anthropic, for ten frontier startups. Then they get into the math.

Gil reaches for punctuated equilibrium. In evolutionary theory you get long stretches of stability, then a burst—he analogizes it to a Cambrian explosion—then consolidation. Technology looks the same: a social wave, a SaaS wave, each settling down. The internet itself had several inner waves, from 1990s portals through social around 2010–2012, then cloud, security, crypto. AI is the latest burst. There will be another step-change in model capability and another crop of startups. Some consolidators have already emerged. What comes after that, over two or three years, is a lot of $100 billion companies. Multi-trillion is a different species.

His threshold is blunt. A trillion-dollar market cap wants something like $50 to $100 billion of revenue with decent margins. There are only a handful of markets in the world that can support a single company at that scale, and a dozen-plus companies already sit near that club. The live question is not what happens in 20 years. It is how many more firms can actually get to $50 to $100 billion of revenue in the next five. Physical goods make the speed problem worse. Energy, robotics, and anything with factories, channels, and deployment have a footprint question. Gil can imagine a short list—supply chain, energy-type technologies, five or six promising areas. He does not doubt the size of those markets. He doubts how fast you can get there.

Guo’s investors fail in the other direction: a failure of imagination. Many have intellectually accepted that AI companies sell services value, then still underwrite Harvey or Abridge on seats, lawyers, or doctors, and never ask what happens if the firm charges for outcomes. Coding has already shown consumption and value with another 100x of room; you no longer have to be a genius to carry that logic into other domains, and a $10 billion market cap that Gil once wrote was hard is now a seed round for some neo-labs. People are investing, she says, as if velocity to a trillion were there for many companies, when some will get to $20 billion, some to $100 billion, and some to zero. The two questions—is the market big enough, and can you get there in three years—are being priced as one.

The other distortion, Guo says, is almost the opposite of over-ambition. She is watching really excellent founders go after niche markets because they are not scared of the neo-labs there. Harvey, OpenEvidence, Decagon, Sierra, even Cognition two-ish years ago, walked into large markets that could have been on a lab roadmap and still built companies. The new trend line is a flight to hardware—“the labs will never do this”—plus American dynamism, narrow applications of AI, and products that an inference cloud should probably just provide. Some of that work will be derivative. Not every startup. Enough of a subset, and it is a subset of the good founders.

Gil’s version is more direct. There are markets the labs will eat naturally, and you should stay out of those. There are a bunch of markets they will not eat, and you can still compete on product, distribution, and experience. People, he thinks, are staying away from both. Guo is not talking about the median founder, who can lie down if they want. She is talking about the best. “I am more often disappointed right now that founders are being less ambitious than they could be.” That is the trend line she means.

The ambition talk turns into the question every founder dodges: when to sell. Gil’s framework has three layers. A handful of companies should never sell any time soon—Anthropic, OpenAI, that class. Most companies in any era should at least consider it. There is usually a 12-to-18-month window when the company is worth the most it will ever be worth; he thinks a recent major exit probably hit that window. For hygiene, he cites Ben Horowitz: pre-schedule a board meeting once a year whose only topic is, without emotion, whether to consider an exit in the next six months. Nobody is pushing. The answer can be keep going. The conversation still needs to happen.

They shorten the cadence on the air. One year of AI time, Gil says, is like three to four years of a normal cycle, so three years is a decade. Model capability, vertical apps, roll-ups, infrastructure—radically different world three years ago. Double-check your thinking more often because the underlying facts are changing faster than they ever have. Guo adds a question that is new for this moment: as costs fall and capabilities rise, are you capturing the value? If you are on the wrong side of the secular change and cannot see a path to the other side, you should sell. She puts Cursor on the table: if the way you want to compete is capital and compute access, is this perhaps a maximally valuable point in time? Pride about never considering a sale is, she says, nonsense. The *Situational Awareness* reminder still applies: everyone has to stay alive to profit.

Financing, Gil thinks, is the part people have backwards. Three-trillion-plus outcomes in a short window are returning an enormous amount of venture capital. Bigger funds need a place to put it, and they will put it against the next trillion-dollar narrative. Valuations will likely keep rising over the next year or two. Guo’s correction is that private markets can stay irrational for a long time, and the ability to finance is the equivalent of avoiding a margin call. Money will skew toward scale of opportunity—or, more precisely, perceived scale. So sit down and run the math: expected outcome, a later 10x or 15x, future dilution, years of work, the moment compounding slows. Do not take the number from investors, the press, or Twitter.

The scarce input is time. Your most productive years are on the line. You can walk away with money and go do the next giant thing, and people will work with you because you have done it—or you can roll the dice, which for some companies is the right answer, but only after the arithmetic. Secondaries, Gil thinks, are a weak intermediate. They solve a short-term need and do not create a solution. Plenty of 2020 and 2021 companies are still running five years later, not working, their founders locked up for the entire AI change. That lifetime cost does not get discussed enough. You only live once, and it is a short life.

Inside the frontier labs, Guo reports a manic energy: code is six months, or year-end, from being solved, and light recursive self-improvement arrives by the end of next year, with models training big chunks of themselves—post-training first, pre-training later. If you have a year or a year and a half of productive work left, the logic goes, you should work 16-hour days, because every week is 2% of the time before you are displaced. Gil’s first cut is that if the leading scientists believe models can improve their own training, and it is an extension of code and math, of course you should believe it. Guo’s data is colder: very smart, even self-aware researchers have felt that RSI or ASI was 18 months away, every 18 months, for the last five years. Extension from code to training code is easier to believe than gathering data in unverifiable domains or beating physical compute, which she thinks are the likelier limiters.

Physical compute, Gil says, actually reinforces an oligopoly. It puts a ceiling on how fast any single lab can go if you assume compute is roughly pro rata among the big players, and it forces closer competition than would otherwise exist. He wonders when that ceiling lifts. The second-order effect of the 18-month belief is a possible burnout cycle. People at a major lab have asked him whether they should get married, because they do not know what the world looks like in 18 months. His advice was to get married; it will be okay. Guo finds the mood tragic. It is psychologically close to thinking you are going to die: how would you spend the last two years? Work, travel, only work, skip the wedding, because your contribution is irrelevant given RSI? She would rather people act as if they have a life, and treat the timeline as an unverified forecast.

Gil then names the concept he wants remembered: return on invested tokens, ROIT. At some labs, a few dozen researchers drive something like 80% of the results—the same human power law he sees in breast-cancer research, subfields of math and physics, and founding. Compute is increasingly given to those people. Hiring has slowed unless the bar is very high, because the cost is not the researcher. It is the compute attached to the person. Who gets the token budget, and why? That is the question. It is the old internal-tools problem in a new unit: tech companies starved the tools team because they would rather put engineers on product. “The death of SaaS is a little bit overstated,” he says, if you would rather not spend tokens on a cheap annual SaaS line when the same tokens could lift the core product or a massive margin. Enterprises have moved from “everybody try AI” to measuring spend and shifting some of it to open source. The next wave is which projects and which people get an outsized piece of the token budget.

Guo asks whether a company or a person’s token budget three to five years from now should look like rent. It depends what the budget is for, Gil says: Minecraft was five or ten people when Microsoft bought it for billions, so the one-person unicorn already happened about 15 years ago, and AI just radicalizes that leverage. Give tokens to people who can take outsized advantage; mass layoffs are not imminent, but even before AI some big-tech engineers were unproductive, and GE, PG&E, and Hershey’s will covet people who look mediocre at Google or Meta. Guo’s advice to a researcher without an outsized token allocation is to take comparative advantage into bio, supply chain, and other fields that should accelerate, rather than worrying about the downfall of mathematics and vacationing until the world ends. An old Google joke asked how many people it took to run a 50,000-person company: inside search and ads the answer was 20% of that org, outside it was everyone, which is why concentration of contribution depends on where you stand.

Guo asks what could disrupt the whole stack. Investors could lose their nerve on capex; markets could freak out about debt and the returns profile—she sees minor signs, not real pressure yet—or a technical alternative to transformers could matter. Gil picks two worries: technology unknowns, and attempts to restrict usage of models we already have, or of open source, in order to slow progress. On architecture he is short. The industry will consume all available compute and power whatever the underlying design, so more memory- and power-efficient architectures are more interesting than ever. Catching transformers on scale and hardware match remains tough. The high-probability outcome is that whoever invents the next thing gets copied, the labs do it, and they have all the compute. The low-probability outcome is a neo-lab that keeps a secret, scales, and pulls away—until one person leaves for Anthropic or OpenAI and the knowledge spreads, which is what has been happening.

The regulatory line lands on California. The state has passed a billionaire tax; the state Democratic Party came out in favor. Guo’s question for a founder of a $10 billion-plus company Gil backed: forced asset sale next year? Dozens of founders, he says, have to sell big chunks of their companies. It is not clear regulators thought through execution and compliance. The immediate effect is that a huge number of people trying to create value choose to leave. That is already happening. “This is the fastest way I can think of to chase the entire ecosystem out.” The statute, as he reads it, is broad enough to re-implement and lower the bar in later years. By 2028 there is talk of an exit tax, a penalty for leaving. He thinks the people who wrote the bill want the flight. The other bad sign, for him, is ballot-harvesting initiatives. He is hopeful California figures it out. If an easy alternative existed, even more people would already be gone. Quite a few he knows are leaving this fall.

Renaissance-style bursts, both agree, happen where enough smart people agglomerate on common work, not where the weather is nicest. Boston was Silicon Valley’s counterweight in the 1980s and lost in the early 1990s; the weather is awful, and biotech is what it still holds. Texas is assembling an energy-tech and hardware corridor. People are moving from the Valley because you can experiment. Hardware started around El Segundo, where SpaceX and then Anduril sat; parts of SpaceX and Tesla have moved to Texas, and a new ecosystem is forming beyond Austin. The driver, Gil says, is regulation, not whether Texas is a better place to live.

The sharpest analogy is pharmaceutical history. Paul Janssen of Janssen Pharmaceuticals, among the best drug developers of all time, sat for interviews 30 or 40 years ago on why new drugs got expensive and slow: regulatory capture, and an FDA that, in his mind, looked at safety and risk and not enough at benefit. Only one side of the equation. You can imagine a lab version of that, where the safety burden is so high that even a dramatically larger positive outcome cannot clear it. Constrain the outside while you push progress internally on an exponent, and a one-year internal lead—worth three or four normal years—is a massive advantage. Nuclear is the existing exhibit. France still generates 70% of its power from nuclear, with none of the accidents and kerfuffles the lobby promised. The U.S. is at 18% and has not built a reactor in 40 years. Japan is at 25%, very safe, very abundant. A 1970s safety lobby, Gil says, basically killed abundant clean energy in America. Guo notes plants are being built now. “We’re not making much,” he replies. Safety has hurt energy and medicine. Society gets to place the needle on AI. There are many worlds.

Before they go, Guo asks what Gil is excited about on the positive end of that wheel. Productivity, education, healthcare, daily life, self-driving, care for the elderly: there is so much good that can come of this, he says, which is exactly why he is cautious about where the spectrum should sit. Proper societal safeguards are worth having. Historically, for big industries, the United States has gone too far. Tech has been so successful so quickly, and has had so much human impact, because it has been lightly regulated. He would rather keep it that way. Go the other direction and you lose optimism, momentum, and progress. That is what happened in biotech. That is what happened in energy for a long time. Guo files the episode as a call to arms against regulatory capture.

The operational residue of a conversation this large is smaller than the narrative. Put a rational review on the calendar every six months and ask one question: given the speed of falling costs and rising capability, are you still on the side that captures the value? Founders ask it of the company. Everyone else can ask it of their skill stack. The answer is not a mood and not a Twitter thread. Start keeping a token ledger. Rank projects and people by return, not by department habit; Gil thinks that allocation will be the next efficiency divide. And split “is the market big enough?” from “can we get there in three years?” before you let a valuation, or a sudden meekness around the labs, smuggle one answer into the other.