Keith Rabois
The month in one sentence
Rabois spent the month arguing that elite judgment remains scarce even when technology becomes abundant: scarce judgment in venture capital, company-building, AI model selection, legal services, geopolitics, and even basketball. The striking thing is how consistently he applies the same worldview everywhere: identify the few people or institutions that actually know what they are doing, ignore consensus narratives, measure people by outcomes, and treat almost every disagreement as something that can eventually be settled by a scoreboard.
AI: the models commoditize; judgment becomes more valuable
The most substantive technology argument of the month concerns AI model routing. Rabois thinks enterprises will increasingly use a neutral layer that decides which model should perform which task rather than blindly sending everything to the frontier model. When people tell him OpenAI, Anthropic, and other model providers will build routing themselves, his answer is essentially: they can't do it honestly. A model vendor has no incentive to tell you that a competitor—or an open-source model—is cheaper or better for a particular task.
That incentive conflict is the core of his thesis. Foundation-model companies can route among their own models, but they will not naturally route customers toward rivals or open source. When challenged that frontier models may already be cheaper after accounting for retries, that applications require stable model behavior, and that model vendors can eliminate obvious arbitrage themselves, he simply rejects the objections: “none of this is accurate.” Elsewhere he makes the more interesting argument explicitly: vendors have “massive incentive misalignment” because conceding that open source works for many tasks would undermine their economics.
This connects to another repeated point: tokens are becoming an actual corporate cost category. He describes tokens as perhaps the fastest-growing business expense not yet properly governed by finance systems, tying the thesis directly to Ramp.
Why this is interesting: Rabois isn't primarily making a technical prediction about which model wins. He's making a market-structure prediction. If models proliferate and performance varies by task, then the valuable layer may move upward from model ownership to model selection, orchestration, cost control, and workflow. His bet is effectively that abundance at one layer creates a new scarcity at the layer above it.
He is still fighting the great “Rabois was wrong about AI/SF” trial
A huge amount of the month is Rabois defending old predictions and his venture record.
Critics resurrect a bet that no company founded in San Francisco during a particular 2020–23 window would achieve a $10 billion-plus exit, along with criticism that he missed Anthropic and underestimated the AI resurgence in San Francisco. His defense has several layers.
First: OpenAI doesn't count as evidence against the original thesis because it was founded earlier, and Khosla had invested in it years before the current boom.
Second: he says he wanted to invest in Anthropic but was prevented by a conflict, repeatedly telling skeptics to “ask the company.”
Third: he still thinks many of the celebrated high-valuation AI startups won't ultimately remain independent multibillion-dollar companies. When someone lists Anthropic, Cursor, Perplexity, Sierra, Cognition, Mercor, Harvey and Supabase as misses, he responds that “several of those will definitely not be worth $ Billions.”
Then reality begins resolving the bet in public. When a purported SpaceX acquisition of Cursor would apparently satisfy the conditions, Rabois first checks whether it is actually a closed acquisition rather than one of the quasi-acquisition/licensing structures that had become common. Once persuaded it is substantially real, he concedes it is “98% real.”
The revealing part isn't whether he ultimately wins or loses that particular bet. It's his conception of venture epistemology: VC isn't a prediction contest where every incorrect statement counts equally. Someone defending him says a venture investor only needs to be spectacularly right a few times. Rabois agrees with the spirit of this and points repeatedly to his investing record and rankings. His implicit metric is not calibration; it is capital-weighted outcomes.
When an economist mocks how someone with his opinions survives in venture, Rabois answers: “by winning every day, loser.”
That is basically his entire theory of knowledge compressed into five words.
Venture capital as a craft of extreme discrimination
Several threads reveal what Rabois thinks venture investors are actually paid to perceive.
Asked how often he encounters a genuinely unique startup strategy, he says roughly one to three times per year—from founders, “definitely not fund managers.”
He also strongly agrees with the proposition that at seed stage, VCs often reject the founder rather than the stated business idea. A founder may hear a polite objection about market, strategy, or product because “I’m not inspired by you as a person” is difficult to say and impossible to retract if the person later turns out to be exceptional.
This is closely related to a long argument about cofounders. Rabois says a nontrivial fraction of companies have the wrong founding team, and a competent VC should not suppress that judgment merely because it is uncomfortable. Matthew Prince pushes back with the Cloudflare experience: great cofounders can be radically different, and he dislikes the idea of a VC investing only after forcing out unwanted founders. Rabois's eventual position is subtler than his opening salvo: founder composition really can be fatal, VCs should speak honestly about it, but the CEO ultimately owns “team construction and editing,” and founders are entitled to choose investors whose intervention style they prefer.
That is one of the month's better exchanges because he moves from provocation to an actual governance principle: the investor's job is to diagnose; the CEO's job is to decide.
“Data” is overrated; great firms contain tacit knowledge
The Harvey/legal-AI discussion gets at another recurring Rabois idea.
A Harvey founder argues that elite law firms such as Kirkland are genuinely differentiated: specialized expertise, processes, accountability, longstanding institutional knowledge, and practice-specific economics make them difficult to unbundle.
Rabois agrees—but says their advantage is not their data.
Asked what it is, he initially replies:
art.
Then expands it:
art, IQ density and brand to CYA.
And adds “some cultural mores too.”
This is probably one of the most revealing clusters in the whole month. His claim is that the durable advantage of an elite professional institution isn't merely a proprietary corpus that can be dumped into an AI system. It is tacit judgment: knowing how to structure an argument, what matters, who is unusually smart, what institutional norms produce good decisions, and the reputational value of being able to tell your board that Kirkland or Wachtell signed off.
It is a direct rejection of the naïve AI-era thesis that all organizational advantage is secretly data.
The Sequoia “scam” debate: sophisticated markets are allowed to be complicated
Another venture argument starts when Brendan Foody accuses Sequoia of misleading people by investing in financings with multiple tranches at different valuations while publicity emphasizes the higher valuation.
Rabois rejects the word “scam” almost completely.
His case:
Founders are responsible for accurately communicating financing terms to employees.
Sophisticated investors can inspect the cap table and immediately understand the blended economics.
Different investors paying different prices is normal and rational because investors contribute different value.
This has been common for years, particularly in AI.
Both valuations should be disclosed, with separate Form D filings where appropriate.
He gives Ramp as a personal example: his investment was at roughly a $27 million valuation while others invested above $40 million.
The deeper position is very Rabois: complexity is not deception merely because unsophisticated observers misunderstand it. Markets don't owe everyone identical economics, and trying to impose aesthetic symmetry on deals would destroy useful price discrimination.
Portfolio maximalism: Ramp, Opendoor and Factory
Rabois does not pretend to be dispassionate about his companies.
Ramp receives almost comically maximalist enthusiasm. Someone predicts Ramp will be bigger than Salesforce by 2030; Rabois answers “obv.” When someone says he led Ramp's seed at around a $40 million post-money valuation, he corrects them downward: “oops it was $27.5 m post.” The correction is itself a flex.
He also insists his recent investing judgment hasn't deteriorated, suggesting his strongest vintages were around 2013–14, 2019–20 and perhaps 2024–26, while acknowledging that comparing recent investments to mature cohorts is difficult.
Opendoor produces the month's nastiest extended fight. A critic argues the business destroys homeowner value and says the company is effectively worthless. Rabois predicts Opendoor will surpass Zillow in market capitalization and calls that equivalent to creating the most valuable residential-real-estate company in the Western world. The ensuing thread becomes mostly mutual contempt, although there is one excellent joke. When his opponent begins a sentence with “I am unqualified to speak to warrants,” Rabois replies:
should just inset a period after the word unqualified.
That is probably his cleanest insult of the month.
Factory gets the same enthusiasm in gentler form. Someone says that if he were a VC he'd be desperate to invest in Factory. Rabois replies:
you’re hired.:)
China: economic relationships become national-security relationships
The most persistent political-business theme is China.
Rabois repeatedly attacks Airwallex, asserting that it is controlled or heavily influenced by Chinese interests and arguing that executives physically located in China create serious problems because of Chinese law. When someone asks whether a startup can safely operate a wholly owned Chinese subsidiary servicing large Chinese companies if key personnel must reside there and comply with Chinese data rules, Rabois says any executive located there is problematic.
When Airwallex supporters accuse him of xenophobia and ask for evidence, he says independent media has substantiated his assertions, demands disclosure of the cap table and percentage Chinese ownership, predicts congressional or governmental scrutiny, and says the company isn't really a direct Ramp competitor because CFOs don't use it to run their companies.
The source gives us his claims and the arguments around them; it does not independently establish whether those factual allegations are true.
The broader framework is clear, though: Rabois sees corporate nationality, employee location, ownership and jurisdiction as inseparable from geopolitics. “It's just a software company” is not a meaningful category once software becomes strategically important.
AI export controls: software has crossed into state power
That same logic appears in the export-control discussion.
When people express surprise that restrictions on powerful AI could apply to foreign nationals—including people physically present in the United States—Rabois points to defense startups as precedent. When challenged on whether government actually has legal authority to prevent release or control access, he repeatedly focuses on the statutory question: national-security concern alone does not manufacture legal authority, but export controls can in principle be applied to AI products.
He also rejects the argument that restricting Chinese access to American chips merely teaches China to become self-sufficient:
China is going to build the best chips they can anyway.
That's a compact but important argument. It rejects a common counterfactual: that without American controls China would somehow have had less incentive to develop world-class semiconductors. In his model, that incentive was already maximal, so controls buy time without creating much incremental motivation.
Interestingly, when someone extrapolates American AI controls into an elaborate thesis about allied “AI sovereignty,” Rabois replies with a much more mundane political point: those countries should stop antagonizing American companies such as X. His instinct is frequently to collapse grand abstractions back into incentives and power.
Foreign policy: wars are about breaking the opponent's capacity and will
Rabois's Iran commentary is among his most aggressive material.
His core analytical framework is “correlation of forces”: understand the relative military, economic, political and strategic strength of each side rather than reasoning primarily from rhetoric or intentions. He calls it perhaps the single most important concept for understanding foreign affairs over the last century.
He repeatedly claims that a recent war with Iran was effectively decided in three days, even when people mock an earlier “48 hours” prediction. His response is basically that three versus four days does not invalidate the strategic claim.
The harsher component is his rejection of fear about destroying the IRGC. When Blake Scholl argues America lacks the military technology to destroy it at acceptable cost, Rabois says capability isn't the problem; the problem is fear that doing so would precipitate civil war, which he considers the wrong lesson from Iraq. When challenged on “kill all of them,” he clarifies that he means enemy soldiers in wartime until their leadership surrenders, invoking Germany and Japan as the model.
That same framework appears at the very beginning of the corpus when he says the correct moral treatment of Hamas would resemble the American wars against Germany and Japan. Whatever one thinks of the analogy, his position is internally consistent: war terminates when one side's organized capacity and willingness to continue are broken, not when violence has been calibrated to some symbolic level.
Sports is where his personality becomes easiest to see
A surprisingly large share of the month is Knicks posting.
Rabois treats NBA games almost exactly like venture debates: strong priors, aggressive predictions, endless argument about evidence, and immediate victory laps.
He believes the Knicks are being systematically mistreated by officials. He describes opponents as “mugging” Brunson, calls a flagrant on Brunson absurd, argues the Knicks should have retaliated physically after perceived dangerous play against their stars, invokes Pat Riley as the model of how such situations should be handled, and at one point says Wembanyama is “worse than Draymond.”
After a Knicks win, he posts:
Even the NBA can’t stop the Knicks:)
This is not really analysis. It's fandom with the rhetorical form of analysis.
There is also an extended mini-bet about Trump attending a Knicks game. Someone predicts security delays, empty seats, booing and bad vibes. Rabois predicts the opposite: full seats, cheers and improved atmosphere, comparing Trump's reception to UFC events. Afterward he immediately posts crowd reports and videos as evidence that he had called it correctly.
The comedy is partly that he treats a crowd-reaction prediction with approximately the same seriousness as a venture thesis.
Soccer: incentives and coaching matter more than structural pessimism
His USMNT comments are similarly terse. Someone argues that America remains far behind top countries in youth development; Rabois responds that “we have a competent coach now.” When another person says improvement has nothing to do with coaching and is instead the result of better player development, experience in stronger leagues and team chemistry, Rabois answers:
lol. that is absurd.
Elsewhere, discussing America's reliance on dual nationals developed abroad, he says there are “many good reasons” for the strategy and that it resulted from a “serious diagnosis of what wasn’t working.”
This fits his broader worldview: diagnose failure, recruit differently, change leadership, don't romanticize the existing system.
The month's recurring jokes and rhetorical habits
Rabois's humor mostly comes from extreme compression plus excessive certainty.
A critic wonders how he survives in VC:
by winning every day, loser.
Someone says they feel sorry for people who interact with him:
yeah only winners enjoy it.
Someone claims a trillion-dollar outcome disproves his startup thesis:
why?
Someone tells him model companies will obviously build routing themselves:
they won’t.
Asked whether frontier labs will build a particular product:
no.
Someone describes a possible four-year bet:
easy bet to win.
A person says “Inshallah. Knicks in 4.” Rabois's objection is not the prediction but that Trump will probably have to watch from a corporate box, “which is kind of lame way to attend.”
His insults are also unusually performance-oriented. People are not merely wrong; they are “losers,” “stupid,” “clueless,” or people who haven't built anything. One critic saying “I feel sorry for anyone that interacts with Keith” receives the answer “only winners enjoy it.”
There is a worldview embedded in the trolling: status should derive from demonstrated competence, and demonstrated competence grants argumentative standing.
What is actually interesting about the month
The surface-level Rabois feed is chaotic: AI infrastructure, VC scorekeeping, China, Iran, basketball officiating, Trump crowd reactions, soccer, lawyers, Opendoor and occasional book recommendations.
Underneath, it is remarkably coherent.
- He believes outcomes are the supreme corrective to narratives
Again and again, someone presents an argument about how things should work. Rabois asks, implicitly or explicitly: who actually won?
That explains his obsession with investment rankings, market caps, exits, closed acquisitions, sports scores, crowd reaction, wars ending and company revenues. He prefers hard-ish outcomes because he believes discourse is cheap.
The flaw is obvious: outcomes frequently underdetermine causal explanations. A correct prediction can be lucky; an investment return does not validate every intermediate belief; winning a war doesn't prove every strategic claim. Rabois seems much less interested in that distinction than a Bayesian or scientific epistemologist would be.
- His deepest business belief is that exceptional humans remain non-commoditized
Despite being extremely bullish on AI, he repeatedly locates real moats in things AI does not automatically flatten:
founder quality;
unusual strategies;
judgment;
IQ density;
culture;
taste or “art”;
reputation;
organizational accountability;
good coaches;
competent executives.
This is much more interesting than generic AI boosterism. His AI worldview is not “intelligence becomes free, therefore everything becomes commodity.” It is closer to: mechanical intelligence becomes abundant, making genuinely scarce judgment easier to see and more leveraged.
- He thinks incentive analysis beats stated intentions
Why won't OpenAI route to open source? Incentives.
Why are export controls potentially rational even if China develops domestic chips? China already has the incentive.
Why can different venture investors rationally pay different prices? They provide different value.
Why are elite law firms valuable even if workflows can be automated? Buyers are purchasing judgment, specialization, brand and accountability, not merely documents.
Across completely different topics, the method is the same: look for the incentive structure, not the press release.
- He has almost no interest in consensus as evidence
People repeatedly confront him with “everyone knows,” social consensus, media interpretation or professional norms. He usually becomes more confident rather than less.
Sometimes this is genuinely useful contrarianism. Sometimes it degenerates into “you are stupid; I win.” The month contains plenty of both.
His most characteristic response may be to a claim that talented founders are partly products of timing and networks:
lol.
That tells you both the strength and weakness of the Rabois operating system. He has a very high prior on individual agency and discriminating talent, and a very low prior on explanations based on luck, structure or social context.
- He treats public argument as an extension of competitive life
There is essentially no boundary between investor Keith, political Keith and Knicks-fan Keith.
He makes predictions. Others challenge them. He records the challenge. Evidence arrives. He posts the evidence. Someone disputes what it means. He escalates. If the result favors him, he takes the victory lap.
Twitter for Rabois is therefore not primarily a publishing platform. It is a continuous prediction market for his own judgment, except the settlement rules are largely determined by Rabois.
That is why the feed remains interesting even when the individual tweets are absurdly terse. You are watching one person apply the same mental machinery to venture capital, AI architecture, organizational design, international conflict and an NBA foul call—and apparently experience all of them as versions of the same game.