Two companies are currently assumed to be worth a trillion dollars each before they have listed. Keith Rabois says every valuation below them is priced off that assumption.
Most of the froth argument is about revenue multiples. Rabois's is not: he says the revenue traction is genuine and unprecedented, and that the open question is which of these businesses still exists in 2030.
"That would be catastrophic and it would immediately reset virtually all valuations in the AI world and probably deprive half the companies that are raising money from access to capital."
Rabois is a Managing Director at Khosla Ventures, was an early investor in Airbnb at a valuation he puts at $1.7 million, and practiced as a securities litigator before any of it.
I listened to the full segment so you can skip it. 26 minutes of audio, 16 minutes of reading.
Here are the 12 takeaways that matter.
👤 Guest: Keith Rabois, Managing Director at Khosla Ventures, an early investor in Airbnb and a former securities litigator, making his fifth or sixth appearance on the show
🎙️ Hosts: John Coogan and Jordi Hays, who present TBPN live on X and YouTube every weekday
📰 Published: 9 September 2026 on YouTube (TBPN)
🔴 YouTube | 🟣 Apple Podcasts | 🔗 Show notes | ⏱️ 26 min | ✅ Time saved: 10 min
Key Takeaways
The revenue is real; the question is which of these companies survives to 2030
Reaching $100M of revenue in two years is unprecedented by any historical standard
A weak first quarter from Anthropic or OpenAI as public companies would reset the whole market
Rabois thinks a miss would come on margin rather than on revenue
Nvidia is turning from a platform company into a product company, and will start competing with its own customers
One failure among the companies marked in the tens of billions would read publicly like WeWork or Theranos
Which is part of why he thinks nobody in a position to prevent it wants one right now
Venture math changed because two orders of magnitude were added to the upside
He invested in Airbnb at a $1.7M valuation; today a $50M check at $500M can still be rational
He will not fund a company growing slower than the frontier labs, with thinner talent, at a higher multiple
There is no evidence AI is destroying jobs, and he says the burden is on the doom argument to say how the world ends
He turns down founders whose businesses he thinks are bad for society, and says most investors do not
He puts prediction markets in that category and cites the 99% who lose money
1. Frothy? Traction Says No
The hosts opened by putting the frothiness charge to him directly.
His first move was to refuse the single label. Some of these companies produce meaningful margin on their revenue and some do not, so a claim that they are all worth X or all not worth X does not survive contact with the fundamentals
He does not price private companies on multiples. "I don't think multiples are the right way to evaluate private companies typically anyway."
The traction, he said, is not in question. "The ability to go to 100 million in revenue from, you know, two years or so or hundreds of millions in four years, basically unprecedented."
The question he substitutes is durability. "But if you work backwards from 2030 and say which of these companies are going to be sustained, which ones have durable advantages, as I like to talk about it or communicate it through accumulating advantages, that's less clear."
What makes it hard is not customer churn but the shape of the industry. He said the way the ecosystem around AI evolves is not remotely easy to forecast, and that his partnership debates it constantly
His summary of the tension is the strongest line in the section. "The founders are impressive. The teams are impressive. The revenue and the adoption is impressive and even the usage often is impressive." Whether any given one is a company that lasts twenty or thirty years is, he said, not obvious
Being acquired, he added, is a perfectly good outcome if the cards are played correctly
2. Nvidia's Product Turn
The hosts raised an unnamed figure in AI who they said appears unwilling to let any AI company fail right now. Rabois picked up the thread with the economics rather than the name
The strategy is coherent from where that company sits. If revenue depends on AI companies buying what you sell, keeping them alive is rational
The change he flagged is a shift in what Nvidia is. It is moving from a platform company to a product company, which he said is visible in its recent acquisition
The consequence is competition with customers. "Right now, any AI growth is great for Nvidia, but as they start having their own products, that may change how people perceive them and how they perceive others."
His read on the motive is growth. For whatever set of reasons, he said, the company believes it needs its own products to sustain the rate it is growing at
The hosts' framing throughout was that the company mints money, grows quickly and carries large profits
3. A $30B Failure Headline
Rabois named a category of company that is large in valuation and small in importance. A great many businesses are marked in the tens of billions of dollars, he said, while being insignificant to the AI trade itself
The problem is the headline, not the loss. He said a failure would run across the Wall Street Journal, the Financial Times and every other paper as a "$30 billion company" — the scale of a WeWork or a Theranos, for entirely different reasons
His view is that the market could absorb it. The processing problem is public rather than financial
But the incentive runs the other way. While the cash is flowing, he said, there is an incentive to buy a struggling company another two years — even where no other investor is willing to back it
The self-interest is explicit. The longer the AI wave runs, the better it is for Nvidia, and the higher the whole industry goes the better still — until somebody builds a better version of Nvidia or removes its advantages, which he said always happens eventually in technology
4. He Buys Through ChatGPT
The example he reached for was Google. Asking what the future of Google is felt silly five years ago and feels realistic now, with a chat interface potentially displacing search
He expects the financials to lag the behavior. Advertising revenue will trail, so the business shows signs of life for a while after the usage moves
His own behavior changed inside a year, and he disclosed his bias while saying so. "I went from last year not making any real like commercial decisions with AI to suddenly everything I buy all that research is happening within ChatGPT and I feel like I'm buying better products and I'm having a better experience"
The specific frustration it solves for him is Amazon's long tail. He does not want the drop-shipped product from a company that will not exist in a year: "I want to buy the thing that I'm willing to pay 30% more to buy the thing where the company has been around for 50 years."
A host's aside was that the running example is paper towels, and the paper towel holder
5. The IPO Miss Scenario
One host asked a long question about a data-center freeze — powered shells becoming more valuable, millions of unplugged graphics processors being shuffled around the world, and political pressure building on both sides of the aisle. Rabois answered a different question first.
He named the single event that would do the most damage. A stumble by Anthropic or OpenAI in their public listings — a valuation that comes in below the assumption that both trade in the trillions
The shape of a miss. Metrics over the first month or the first few quarters that make investors value either company less than they do now: a revenue miss, a margin miss, or something similar
The consequence he described is immediate and system-wide. "That would be catastrophic and it would immediately reset virtually all valuations in the AI world and probably deprive half the companies that are raising money from access to capital."
He hedged the forecast twice. "I'm not saying that's going to happen or is likely and I think if they miss it would be more on the margin side than on the revenue side right now."
Why one company's print resets everybody else's price. Every current valuation is predicated on being able to build a trillion-dollar company inside the life of a venture fund. Two examples exist; several more are believed to be candidates. Remove the belief and the amounts being invested stop making sense
6. Two Zeros on the Upside
His history lesson on what a home run used to be. A very successful venture outcome was $10 billion to $50 billion, perhaps $100 billion at the maximum
The change is arithmetic, not sentiment. "It is possible now that at least one order of magnitude, possibly two orders of magnitude have been added to the upside potential of a startup"
That is what makes today's entry prices defensible. It adjusts the prudence of putting $50 million in at $500 million, in a way that it would not have twenty years ago
The comparison he used is his own. Rabois noted that "when I invested in Airbnb, it was at $1.7 million pre- or post." A host's reaction was that it sounds crazy now
His conclusion is conditional rather than bullish. That world may or may not return, but with a zero or two added to the top end, an investor can look themselves in the mirror and take the shot
7. Talent Density Is the Tell
The hosts described the rounds they cannot make sense of. Series B and C companies carrying a richer revenue multiple than the leading labs, growing no faster, with thinner teams — and all of them getting funded. One host called it "that's like the face of frothiness"
Rabois agreed and said none of them are his. He offered, as a joke, to introduce the hosts to competitors who are doing those rounds
He then named the variable he thinks the hosts had identified correctly. Critical density of talent matters, and a lot of people lose sight of it
At low scale, growth is not the signal. A company growing slowly early is defensible if the right team is assembled against the right problem — and at that stage, he said, it does not much matter whether you value it off Meta's multiple, OpenAI's implied multiple or Nvidia's
His bar for the team is not that they would have got a job at a frontier lab. They have to be world class at their own craft, and assembled densely
The arbitrage he described is finding people the labs undervalue. There is an archetype a company can be scaled on
But the standard for the trillion-dollar claim is absolute. "you still have to believe you're scaling with incredible talent like top one basis point, 10 basis point on some dimensions to really believe that you're going to create a trillion dollar business from scratch in less than a decade"
The pitch he says he would turn down, in the hosts' words as much as his: growing more slowly than the frontier labs, less talent-dense, at a lower scale and a higher multiple. His answer to that would be no
8. Cults and Their Tenets
The hosts asked whether differences in philosophy — how likely someone thinks AI is to end humanity, how convinced they are that machine intelligence is imminent — sort talent into separate pools. They cited Scott Wu of Cognition as an example of an optimistic, application-layer team.
Rabois said yes, without qualification, and traced it back to a line he attributed to Peter Thiel about building a cult
Part of what a company sells its employees is a philosophy of life. The example worked through on air was Ramp: a team described as optimistic about the future, using technology to improve society, and confident the world will be better than it was ten or twenty years ago
One host's reply was that cults have tenets
9. No Evidence of Job Loss
Rabois placed himself before he answered. "If I wasn't a tech optimist, I wouldn't be investing in tech companies." He said people tend to orient toward optimism or doom early in life
His challenge to the people leaving is about their own equity. If they genuinely believe they are doing something bad or evil, there is a question of whether they should be building at all, or forfeiting what they hold
His evidence argument is that the last prediction failed. "You can't find any evidence that AI is causing creating job loss. If anything, it's creating jobs."
A host summarized the scoreboard as nought for one, and Rabois conceded the difference: the jobs claim was falsifiable and this one is harder to falsify
What he wants from the doom argument is a mechanism. "But I think the burden is on the people arguing for regulation or slowing progress down of articulating well what goes wrong, why and how and so that people can probe and push back and try to evaluate and quantify that risk." He said he has not seen a compelling answer beyond an appeal to having worked on it
His historical comparison was nuclear weapons. He said the case that "the nuclear bomb in many ways probably made the world more peaceful" is strong rather than proven, measured by deaths in conflict as a share of world population before and after
He then conceded the limits of it. Nobody knew it would play out that way in 1945, and he cited the scene in "Oppenheimer" where Einstein is consulted about whether the atmosphere would ignite and gives no definite answer
His conclusion is that risk is not eliminable. People drive cars and fly helicopters; a zero-defect approach to life does not work for humans
On whether the state should handle it. "Even if that was 100% true, the idea that politicians are going to make it better is also subject to historical debate." Asked whether a genuinely serious problem should be entrusted to the federal government or to private citizens, a host's answer was that it is not obvious
10. p(doom) in the S-1
The hosts asked whether extinction probabilities end up in a listing prospectus, and expected Anthropic to include one as a risk factor
Rabois said it could, and gave the lawyer's argument on both sides. "I used to be a securities litigator."
The argument against disclosure. "You could say, hey, like if the real risk is the world's going to blow up, buying Anthropic shares is no more dangerous or risky buying anything else." On that reading it is not a risk specific to the company
The argument for it. If the political environment turns, a runaway incident could create liability large enough to threaten the investment, and he expects language along those lines — toned down far enough that most readers will not notice it
The drafting tactic he described came from Peter Thiel. Include as many risk factors as possible so that no reader can work out which one matters, producing a document with hundreds of pages of risks
The risk he thinks is actually acute is political. Frightening enough politicians into policy that constrains growth — which becomes a real exposure if a company cannot control what its employees publish
A host's own suggestion was that including it may simply be proper disclosure to the securities regulator
11. The Data-Center Water Myth
Rabois said some opposition to data centers is externally funded. He claimed there has been significant exposure of anti-data-center campaigns being funded intentionally by China, which he called an adversary
On the substance of the water claim he was blunt. "That doesn't mean that the political cost doesn't have to rebut it, but it is kind of a walking IQ test for someone if they really believe that data centers are going to like make the world run out of water somehow."
The hosts traced the claim to arithmetic. Their account is that it began as a calculation error off by a couple of orders of magnitude and then spread
Rabois's wider point was about the habit rather than the number. People do not do the calculation and do not feel they need to, even though the tools to check it are now free
12. Values Over Returns
The hosts asked how effective altruism and venture capital interact, given a pattern of blowups and the asymmetry that a venture investor's downside is capped while a founder's is not.
His model of the job is matchmaking. Venture practiced well is matching a founder, a team and a vision against an investor's own view of the world
Which means beliefs are part of the fit. A founder with a particular set of convictions might be a good match for one investor and a bad match for another
He gave a live example against his own deal flow. "I personally don't believe prediction markets are good for society." He said he watches the sector and is friendly with people who work in it
His reasoning uses a line he says predates its most famous user. He agreed with the description of prediction markets as another casino for Americans, and said he used the line before the president did
The evidence he cited is the loss rate. "I don't think we need more forms of speculation in the United States for normal people and the evidence is that you know 99% of the people who engage in prediction markets are going to lose money"
He extended it past his professional life. "I don't really love sports betting even though I love sports"
His estimate of how many investors would set their beliefs aside for a return was "80 90 95%". A host said he agreed with the assessment
What he says his own firm optimizes for. "We wake up and say is this going to have a positive impact in society?" The precondition he named is that the constraint is not money: "We have you know more money than we need." He said the same filter operates at Founders Fund
Bonus Insights
The segment opened on the theft of paintings from a French museum, which Rabois had missed despite having been in France the week before. A host asked whether it is bearish for France and whether he was the culprit
Rabois described the venture relationship at its best as alignment between a founder and an investor or board member on how the world works and what is good for society
He closed by saying he had missed the show's first month and had been on almost every day since, and the hosts said the booking came from realizing it had been too long
Rabois's bottom line is that the AI market's prices are all downstream of one belief — that a trillion-dollar company can be built inside a fund's life — and that a single disappointing set of public numbers from Anthropic or OpenAI would remove the belief and the funding at the same time.
Products, Companies & Tools Mentioned
Khosla Ventures (Rabois's firm — he described its filter as whether an investment has a positive impact on society, and said the same is true at Founders Fund)
Anthropic and OpenAI (The two companies whose public listings he says every other AI valuation is priced against; a margin miss from either would reset the market)
Nvidia (Moving from a platform company to a product company, which he says will put it in competition with customers it currently only benefits from)
Google (The case study in a business whose usage can move to a chat interface years before its advertising revenue does)
ChatGPT (Where all of his own purchase research now happens, a change he dates to the last year while disclosing his bias)
Amazon (The frustration he says AI shopping fixes — drop-shipped products from companies that will not exist in a year)
Airbnb (His benchmark for how much the upside has changed: he says he invested at a $1.7 million valuation)
Ramp (The example used on air of a company whose talent density comes from a shared optimism rather than from a shared risk model)
Cognition (Named by the hosts as an application-layer team with high talent density and little interest in doom)
Founders Fund (The other firm he says applies a values filter to what it will back)
Books & Resources Mentioned
Steve Jobs Would Have Loved AI, and Been Hated for It (The show's own notes for the day, listing the running order and the headlines behind the news block)
Oppenheimer (Cited for the scene in which the physicists consult Einstein about whether the test would ignite the atmosphere, and get no definite answer)
If this was worth your time, send it to someone closer to the industry than you are.
Get the latest market chatter as it happens: