Steve Eisman Picks Nvidia Over Hyperscalers, Pans SpaceX IPO

Big Short legend Steve Eisman has a pointed structural argument against where AI investment dollars are flowing. In an interview with Fortune, he argued that hyperscalers selling AI services have no durable moats because users move freely between ChatGPT, Gemini, and Claude, forcing providers into price competition with no obvious floor. "Even if AI is the greatest thing since the invention of the printing press," he said, "there are no moats to shield the providers."

The capital cost problem compounds the moat problem. Alphabet reportedly spent $80 billion on AI last year and is on pace to spend between $180 and $190 billion this year, a rate that required an $85 billion stock offering to fund because cash flow cannot keep up. Eisman's point is not that AI is a bad bet generally. It is that the companies spending the most on it may not capture the returns.

His preferred alternative is the infrastructure layer. Nvidia, Arista, and Cisco sell the specialized chips and networking gear that hyperscalers must buy regardless of which AI product wins. That, Eisman argues, is the picks-and-shovels position, preferable to being Meta, Oracle, Microsoft, or Alphabet competing on price in a market where switching costs are near zero.

On SpaceX, Eisman was blunter. He compared the company's revenues to Kellogg's, saying: "SpaceX has the revenues of Kellogg's, which makes Froot Loops, which I love, but no one is going out of their way to buy Froot Loops." SpaceX is reportedly targeting a valuation of close to 100 times sales, while Nvidia trades at around 14 times revenue despite 85% growth last quarter. According to Benzinga's coverage, Eisman called the expected IPO valuation "kind of crazy" and said the only plausible justification would be for SpaceX to become a "retail cult stock." SpaceX's AI division reportedly generated $818 million in revenue against a $2.47 billion operating loss, and Eisman called Grok a "third-tier product" that engineers in Musk's own space division reportedly will not use.

The caveat worth naming is that Eisman is making a relative-value argument at a moment when retail momentum has sustained high multiples well beyond what fundamentals alone would suggest. What the reporting does not offer is a price target or a time frame for when hyperscalers' capex burden becomes visible in investor returns. And if AI agent ecosystems generate switching costs that do not yet exist, the moat thesis weakens. The picks-and-shovels bet is compelling as a structural view, but infrastructure suppliers are not immune to a build-out plateau either.