Mark Cuban: OpenAI Will Never Return The $1 Trillion It's Investing
TL;DR
Mark Cuban thinks trillion-dollar AI capex won’t pay back — he says companies like OpenAI are “throwing away the money at scale” because compute gets faster and cheaper quicker than the market hype assumes.
The real risk is foundational models may become a winner-take-most business — Cuban compares it to search or streaming and says if there are only one to three durable winners, everyone else may have spent enormous sums just to become “a website” or “an app.”
Vertical AI advantages may depend more on proprietary IP than model size — he points to OpenEvidence buying healthcare IP and warns hospitals and universities not to publish valuable research because once it’s public, every model can train on it.
Cuban sees Anthropic’s coding focus as a smarter niche than broad ‘AI for healthcare’ claims — in his framing, programming is a cross-industry workflow advantage, while healthcare inside a general model is “a feature, not a product.”
He argues today’s text-and-image models still don’t understand the real world — using the example of a 2-year-old pushing a sippy cup off a high chair, he says the next leap is a ‘worldview’ approach grounded in physics, video, and real-world perception.
Cuban draws a contrast between Dario Amodei and Sam Altman — he says Dario’s scary rhetoric is at least legible as fundraising strategy, while Sam is “all over the map,” which Cuban thinks could backfire on trust and business relationships.
The Breakdown
The opening shot: trillion-dollar AI spending doesn’t pencil out
Asked what kind of return would justify the current AI buildout, Cuban doesn’t hedge: “They’ll never get it.” His core point is that this isn’t a referendum on whether AI works — it’s that the projected data center spend assumes demand and infrastructure economics that may not hold, especially as processing gets faster and cheaper sooner than people expect.
Why everyone is still forced to spend like crazy
Cuban acknowledges the trap: if you’re a foundation model company, you almost have to go all in because nobody knows whether this market ends up like streaming, with several viable players, or like search, with one dominant winner. That uncertainty is what drives the ring-kissing, nonstop fundraising, and giant capex bets — even if some companies are burning more cash than they actually have.
Vertical AI may belong to whoever owns the IP
When Alex Kantrowitz argues that specialized model training could create multiple winners, Cuban counters with healthcare. He says companies like OpenEvidence are doing something stronger than training — they’re buying intellectual property — and he tells hospitals, research campuses, and schools via Cost Plus Drugs: don’t publish, sell, because once you publish, every model gets your edge for free.
Why Claude’s coding lead feels more durable than ‘AI for healthcare’
Cuban says Anthropic’s programming focus makes sense because coding is not really a vertical; it’s a process layer that can apply everywhere. By contrast, he calls healthcare for OpenAI “a feature, not a product,” arguing that if someone else can spend for the same IP, the moat is thinner than people think.
Gemini has distribution; others still need a real niche
He gives Google a clear edge because “Google search is Gemini” — the AI answer is already embedded in the product and surrounded by ad inventory. But when he looks at Meta and OpenAI, he says he still doesn’t know what their durable niche is, and if too many companies end up in the same lane, then after all that spending, “you’re just an app.”
The current models still don’t understand the world
Cuban’s bigger critique is that today’s foundation models are built mostly on text and images, not on actual real-world understanding. He uses the image of a 2-year-old knocking a sippy cup off a high chair to show the gap: models can absorb lots of information, but that doesn’t mean they understand physics well enough to generate something like E=MC^2.
From LLMs to a ‘worldview’ model of AI
He says the next wave won’t just be larger language models but systems grounded in video, physics, and sensory understanding of reality. He mentions investing in a satellite company that can identify material composition — wood, rock, and more — via spectrograph-like analysis, and says that sort of worldview capability could supersede what Claude, Grok, and similar systems do today.
Dario as fundraiser, Sam as chaos agent
When asked whether he sees himself in Sam Altman or Dario Amodei, Cuban says “maybe a little bit in Dario, but not in Sam.” He frames Dario’s apocalyptic messaging as classic money-raising theater — comparing it to how he pitched Broadcast.com as the thing that would replace cable decades before it really happened — but says Sam feels “all over the map,” citing stories like OpenAI backing out after moving to buy 40% of a memory-chip company’s output and warning that eventually people stop trusting you.