All podcasts / All-In Podcast / Summary

Mark Cuban on the AI Bubble: Who Actually Gets Wiped Out?

2026-07-20 - 45 min - source - Read full transcript
Jason Calacanis (host)Mark Cuban

Key insights

This AI bubble is unlike dot-com and will mostly wipe out VCs and PE funds, not the public.
Cuban argues that unlike 1999, there are no thinly-traded public companies with no revenue getting bid up by retail investors. The exposure this time sits with venture and private-equity funds that deployed capital late, at inflated entry prices, chasing outcomes similar to Anthropic and SpaceX.
ai-bubble-risk
Big Tech's AI capex is increasingly debt-funded, stacking risk on top of an already-strained private credit market.
Cuban notes that even cash-flow-rich leaders like Google and Meta are borrowing hundreds of millions of dollars because their capex is outpacing free cash flow, layering new debt onto a private credit market he says already has problems, with some players issuing bonds out as far as 50 years.
ai-bubble-risk
A price-performance breakthrough in AI compute could strand today's data center buildout, echoing the dot-com-era fiber glut.
Cuban compares today's data center race to the late-1990s fiber-optic buildout: once bandwidth jumped from 1 gigabit to 100 gigabit fiber, most of the laid fiber went dark and unused. He expects a similar efficiency curve on AI compute and power, which could leave some data centers unable to get utilized ('turned into pickleball courts').
ai-bubble-risk
The AI boom needs more mid-size IPOs so disruptive companies have stock as acquisition currency.
Cuban argues that because private capital dominates this cycle, too few AI-adjacent companies are going public at the $50-100M range. Without a public stock currency, a company that needs to acquire a competitor, or a company with data/domain expertise it needs, has to raise expensive private capital instead of doing cheap stock-for-stock M&A the way Cuban did with Broadcast.com.
ai-investing-strategy
Employees at hot AI labs should start hedging their concentrated equity now, the way Cuban did after selling Broadcast.com.
Cuban describes building a synthetic collar on his Yahoo stock in 1999 by having Goldman Sachs construct a custom internet-stock index and shorting it, since no packaged hedging product existed yet; he lost tens of millions on the short itself but says it protected his downside. He suggests employees at Anthropic, OpenAI, and SpaceX consider similar downside protection today given how much their paper wealth has already changed their lives.
ai-investing-strategy
AI is much harder to implement inside real enterprises than personal prompting success suggests.
Cuban says most people massively overestimate enterprise AI readiness because they've had easy personal wins (writing code, doing research, '100x' personal productivity). But CEOs largely don't understand the technology, and the fact that Anthropic, OpenAI, and Microsoft are all hiring thousands of forward-deployed engineers to implement AI for customers is itself proof the tech isn't the self-serve, ask-and-it-does-it product AGI narratives imply.
enterprise-ai-limits
Predicted mass white-collar job losses from AI haven't materialized because it still can't reliably run open-ended, recurring tasks.
Cuban notes that two years after predictions of 50% white-collar job loss, employment is still growing and companies are hiring more AI-literate people, not fewer employees. Current tools can answer one-off questions but fail at multi-step, recurring workflows (e.g., a standing weekly research report) without a human with a 'programming mindset' iterating on broken outputs.
enterprise-ai-limits
AI's current failure points are themselves a business opportunity for AI-literate operators.
Because most businesses hit a wall past basic AI use cases (comparing it to how companies once needed a dedicated Excel or PowerPoint expert), Cuban says anyone with a basic technical background can walk into a small, medium, or large company, diagnose where their AI implementation is breaking, and get paid to fix it - unlocking value the business couldn't capture on its own.
enterprise-ai-limits
AI agents 'drift' as the underlying model changes, creating an ongoing maintenance burden most builders don't anticipate.
Cuban describes his portfolio companies' internal tools breaking over time not because the automation logic changed, but because the underlying large language model powering it evolved and no longer matches how the tool was originally programmed, meaning AI-built software needs continuous re-tuning rather than being a one-time build.
enterprise-ai-limits
Today's LLMs lack basic physical and causal intuition, which Cuban expects video-trained 'world models' to eventually fix.
Cuban argues that because current models are built almost entirely on text and images, they lack the causal, physical intuition a toddler has (e.g., predicting that knocking a cup off a high chair gets a reaction). He expects the next real capability leap to come from video-scale training data, citing his investment in Matter.com, which uses satellites and spectrography to generate world-model training data, and says video/world-model demand is his best guess for where AI compute needs will exceed today's estimates.
enterprise-ai-limits
Cuban expects large language models to become a counterweight to social-media-driven political polarization, because their business depends on being trusted as accurate.
He argues that social media's business model rewards engagement and therefore amplifies polarizing content and 'algorithm-savvy' political figures, while LLMs' commercial incentive is the opposite: losing user trust by being dishonest is existential for a product like Claude or ChatGPT. As more people turn to LLMs for straight answers on policy questions, he expects that to reduce information asymmetry over time, even if it won't replace social media.
ai-and-politics
NBA franchise valuations have decoupled from wins and attendance and now track streaming subscriber numbers.
Cuban says team value is driven by subscriptions to services like Peacock and ESPN rather than TV ratings, attendance, or championships, and that the real open question for future valuations is whether those subscribers churn out after a title run (like the Knicks') or stick around.
sports-team-economics

Media referenced

Companies

Techniques and frameworks

Summary

Jason Calacanis sits down with Mark Cuban for a wide-ranging 45-minute conversation that opens on whether AI is in a bubble. Cuban's answer is a qualified no: this isn't 1999, because there's no wave of thinly-traded public companies with no revenue getting bid up by retail buyers. The risk instead sits with venture and private-equity funds that deployed capital late and at inflated entry prices, chasing the same handful of hot private names (Anthropic, SpaceX) that everyone else is chasing. He flags a second-order risk in how that capital is being financed: even cash-generative giants like Google and Meta are borrowing heavily to fund capex, stacking new debt on top of an already-stressed private credit market, with some issuance stretching to 50-year bonds. Drawing a direct line to the late-1990s fiber-optic buildout, where bandwidth breakthroughs left most laid fiber dark and worthless, Cuban warns that a similar price-performance curve on AI compute could strand today's data center investment.

His prescription is structural rather than doom-laden: the market needs more mid-size IPOs so disruptive AI companies have public stock as cheap acquisition currency instead of needing to raise expensive private capital to buy up competitors or acquire domain expertise, the way he used Broadcast.com stock to buy companies before selling to Yahoo. He also floats a more personal hedge, describing how he built a synthetic collar on his own Yahoo stock in 1999 (shorting a custom Goldman Sachs internet-stock index because no packaged product existed) and suggesting employees at today's hottest AI labs think about protecting their now-life-changing paper wealth the same way.

The conversation's middle section is the most pointed: Cuban argues AI is far harder to actually deploy inside real businesses than personal prompting success suggests, and that the CEOs making enterprise AI decisions largely don't understand the technology. He treats the fact that Anthropic, OpenAI, and Microsoft are all hiring thousands of forward-deployed engineers as proof the technology isn't yet the self-serve product the AGI narrative implies - if it were, you wouldn't need armies of humans implementing it. That gap is why predictions of near-term mass white-collar job loss haven't played out; instead, it has created an opportunity for anyone with basic AI literacy to walk into a company, diagnose exactly where its implementation is breaking, and get paid to fix it. He also flags a less-discussed cost: AI agents "drift" as the underlying models keep changing, meaning tools built on top of them need ongoing maintenance rather than being a one-time build.

Cuban closes the technology thread by predicting the next real capability leap will come from video-trained "world models" rather than further scaling of text-and-image LLMs, since current models lack the basic physical and causal intuition a toddler has. He cites his investment in Matter.com, which uses satellites and spectrography to generate world-model training data, as a bet on that shift, and says video/world-model demand is his best guess for where AI compute needs will most exceed today's estimates.

The final third pivots through politics and sports. On the socialist-versus-capitalist narrative playing out in U.S. local politics, Cuban argues it's less an ideological shift than a demonstration that whoever is best at driving social media algorithms wins elections - and he's cautiously optimistic that LLMs, whose business model depends on being trusted as accurate, will become a counterweight to that dynamic over time. He also dismisses wealth-tax proposals as "showmanship," citing a conversation with the Berkeley economist behind Elizabeth Warren's model who admitted no behavioral response was modeled. The episode ends on Cuban and Calacanis trading NBA talk, where Cuban argues franchise valuations have fully decoupled from wins and attendance and now track streaming subscriber counts, and that the league's new "second apron" salary rule has ended dynasties by forcing title teams to break up their rosters.

Notable Quotes

"It's not a bubble that's going to impact most people in the room, right? Or most people across the U.S., but it could just destroy a lot of VCs and a lot of funds and a lot of P.E., right? Because they're going all in." - Mark Cuban

"If there's a price performance curve on AI that minimizes the power requirements, there's going to be a lot of data centers that are going to be turned into pickleball courts." - Mark Cuban

"AI is a lot harder to implement than anybody expected." - Mark Cuban

"Every single mother fucking business plan ever written in the history of business plans is wrong... So what if the AI is wrong, if the model was wrong? You know, because you're going to learn and you're going to iterate." - Mark Cuban

"Their currency is getting you the correct answer and the correct knowledge, social media's currency is keeping you engaged... you're not going to get rid of social media, but I think people... are going to go more and more to large language models." - Mark Cuban