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

Mark Cuban argues the current AI situation differs from the dot-com bubble because no companies are going public with inflated valuations and no revenue. However, he warns the bubble will destroy many VCs, funds, and PE firms that are going all in, while ordinary people will largely be insulated from the fallout.

Cuban identifies serious structural risks in AI infrastructure financing. Google and Meta are borrowing hundreds of millions to billions on top of spending all their cash flow on capex, layering onto an already stressed private credit market. Committing tens of billions over ten to twenty years for data centers is pricing to perfection, and for investments like OpenAI's 100 billion dollars to work, returns must come back not just as revenue but as profitable margin dollars. Early-stage angel deals that once priced at 5 to 10 million dollar valuations began coming in at 40 to 60 million for products not yet launched. Cuban draws an analogy to the fiber buildout, where bandwidth went from scarce to abundant and dark fiber sold for pennies on the dollar, predicting the same price-performance dynamic will hit AI infrastructure. He acknowledges he could be wrong if video, world models, and robotics drive far greater compute demand than expected.

Cuban is skeptical that AI will eliminate 50 percent of white-collar jobs within two years, noting employment is still growing and companies are hiring more AI-literate people. He cites Microsoft hiring 6,000 people for forward deployment of AI as evidence that enterprise implementation remains far harder than anticipated, and says CEOs broadly have no clear understanding of what is happening with AI. On the productivity side he is more optimistic, noting Lovable is generating 770,000 applications per week, with only 30 percent of its business in the US and only 20 percent of its users being engineers. Small teams using tools like Lovable are building software that would have cost two to three million dollars per year through outsourced development. Cuban also describes generating a full business plan including a patent and licensing outline in 12 minutes using AI, compared to six months for a prototype and twelve months to launch historically.

Cuban identifies meaningful technical limitations in current AI. AI does not learn from repeated user errors within a session and does not proactively surface fixes that worked for others. AI agents drift over time as changes to the underlying large language model diverge from how the agent was originally programmed. He also argues AI cannot understand physical causality the way a two-year-old does, and that world models require far more than YouTube video, needing physical data collection such as people wearing gloves performing tasks. He notes matter.com is launching satellites to capture spectrographic video of everything beneath them to feed world models.

Cuban invested in Open Evidence, a medical AI company, and argues AI will not replace doctors but will make them smarter by handling the volume of new medical information no human can memorize. He estimates that 95 percent of medicine is guessing because doctors cannot keep up with new research, and believes combining wearable data including sleep, heart rate, EKG, and steps with blood panels and diet data will produce significantly better health insights.

On AI and politics, Cuban argues that algorithms drive how people vote in the United States more than anything else, and that whoever controls the algorithm controls elections in many cases. He contrasts social media algorithms, which reward engagement, with large language models, which he says are incentivized to seek truth because their currency is getting users correct answers rather than keeping them engaged. He predicts that as political uncertainty grows, more people will turn to large language models for guidance, and that LLMs will reduce information asymmetry in politics more than any prior technology.

Cuban disclosed that when he sold broadcast.com he created a custom index of internet stocks he considered overvalued with Goldman Sachs and shorted it as a hedge before a real collar on his Yahoo stock was in place, losing tens of millions of dollars on that short position in the interim. He advises employees at companies like Anthropic or OpenAI to consider using a collar to protect downside while preserving some upside, given that current private valuations may not hold.

Cuban argues NBA team valuations are now driven primarily by streaming service subscriber counts rather than attendance figures, wins, or losses, with the caveat that significant subscriber churn could eventually threaten those valuations. He describes the second apron rule as a complete game changer that forces teams to break up rosters and makes it structurally impossible to maintain three max players simultaneously, framing OKC's strategy of accumulating draft assets as a direct response to the reality that the second apron will eventually force them to dismantle any expensive core.