The AI Data Center Oversupply Crisis Is Coming | Ed Zitron

The episode opens with Zitron unpacking the news that Meta, after five reorganizations of its AI division, over $100 billion in capex, a $14 billion Scale AI acquisition, and a middling Llama-successor model called Behemoth or 'Muse Spark,' is now internally discussing leasing out its own compute to other companies. Zitron calls this a red flag rather than a bullish sign, arguing that Zuckerberg himself hedged at a recent shareholder meeting by saying the company thinks it has a use for the compute, but will rent it out if it doesn't. To Zitron, that's the language of a company that never had a coherent AI strategy to begin with, and it raises uncomfortable questions about what all the remaining capex, still projected at $125 to $145 billion this year, is actually for.

He also flags Meta's messy contractual entanglements, including a $17 billion-plus deal with Nebius and a $30 billion-plus deal with CoreWeave, both of which those companies reportedly used to raise financing to build data centers specifically for Meta. If Meta doesn't need that capacity anymore, it's unclear what happens to those contracts or to massive projects like the $30 billion Hyperion data center in Louisiana, described as being the size of Manhattan. Zitron notes that Meta's AI lead, Alexander Wang, publicly clarified that Zuckerberg's comments about slow agent progress applied to the whole industry, not just Meta, which only deepened investor unease.

Zitron's central thesis is that genuine, diverse demand for AI compute barely exists outside of OpenAI and Anthropic. He estimates the two companies together consume roughly three and a half gigawatts of the available compute, essentially all of it, while outside demand from smaller players like Lambda or other neoclouds amounts to maybe 500 megawatts to a gigawatt, a comparatively tiny figure given the capital involved. He points out that CoreWeave's biggest customers are Microsoft (for OpenAI), OpenAI itself, Google (for Anthropic), Meta, and Nvidia, the same small cast of names recurring across every contract, which he says proves there's no broader market.

This concentration leads to his starkest prediction: if Meta ends up selling its surplus compute directly to OpenAI or Anthropic, rather than building a legitimate inference business for outside customers, that will be definitive proof the industry has no other source of demand. He calls this scenario 'game over,' since it would mean the entire trillion-dollar buildout exists only to keep two unprofitable companies alive through what amounts to compute welfare.

Turning to what happens when hyperscalers eventually pull back, Zitron cites Goldman analyst commentary suggesting the first company to cut capex will be rewarded by markets, which he predicts will trigger a cascade as the others follow, dressed up in language about 'efficiency' rather than admitting AI isn't delivering returns. He initially assumed Meta would be the last holdout but now thinks it could be the first to blink, while Microsoft, Google, or Amazon could also move given signals like Amazon's planned $200 billion capex or Satya Nadella's billion-dollar bet on forward-deployed engineers, which Zitron reads as evidence AI tools aren't working well enough on their own.

He stresses that OpenAI and Anthropic have never built their own infrastructure, relying entirely on hyperscalers, whom he estimates have spent around a quarter of a trillion dollars building it for them. If Microsoft, Google, and Amazon cut back, OpenAI and Anthropic would suddenly need to fund infrastructure themselves, something Zitron doubts they can afford. He frames the entire ecosystem as cowardly and directionless, chasing capex because there's no other growth story, and warns that when the retreat begins, it could unwind quickly.

The conversation turns to a widely circulated report that an OpenAI engineer found a way to halve inference costs. Zitron is skeptical, noting the story reads like a leaked Slack message rather than an official announcement, and that OpenAI would have trumpeted a genuine breakthrough loudly rather than let it leak quietly. He suspects any real savings apply narrowly, perhaps to serving degraded models to free users, rather than representing a fundamental cost breakthrough.

More importantly, he uses this to illustrate a structural bind facing the industry: if OpenAI and Anthropic actually do cut inference costs substantially, they need less compute, which exposes the industry's oversupply. But if they don't cut costs, they need endless venture capital and debt to keep buying compute, which is unsustainable. Either path, he argues, ends badly, and there is no scenario in which the current spending trajectory is validated by underlying economics.

Zitron takes particular aim at Nvidia's new partnership program, which offers to lease or buy back GPUs from cloud providers in exchange for a share of profits. He calls this a desperate maneuver rather than reassurance, arguing it mirrors the round-tripping schemes of the dot-com era involving companies like Lucent and Nortel. He explains the mechanics: neoclouds like CoreWeave or Nebius take a hyperscaler contract to a bank to secure financing, and Nvidia has separately guaranteed to buy back CoreWeave's supply, a practice Zitron says should be illegal and only makes sense if genuine end-user demand doesn't actually exist.

He notes Nvidia has committed $26 billion over five to six years to rent back its own GPUs and up to $80 billion in non-cancelable commitments to TSMC, exposing even a hugely profitable company to real risk if customers cancel contracts. He describes the whole arrangement as 'kayfabe,' industry participants performing confidence in a market they privately doubt, hoping the performance eventually becomes real.

The discussion shifts to Samsung and SK Hynix's announced plan to spend over $500 billion expanding memory fabrication capacity, which Zitron connects directly to soaring RAM and storage prices already hitting consumers, evidenced by used laptops now being sold without RAM installed. He explains that high-bandwidth memory used in GPUs is extremely high margin and is crowding out fab space for ordinary consumer memory, creating an artificial shortage that could persist until at least 2028. He criticizes memory makers for removing price caps on long-term deals and extending minimum contract terms, calling their current 84%-plus margins nakedly exploitative rather than a fair correction for past down cycles.

He warns that when the AI bubble eventually bursts, demand for high-bandwidth RAM will collapse, leaving memory companies overextended just as they were in previous boom-bust cycles, but this time the fallout will be worse because inflated consumer electronics prices won't simply reverse. Zitron insists the AI industry bears direct responsibility for this consumer-level inflation and argues the culpable parties will largely escape consequences while ordinary buyers absorb the cost.

In the closing stretch, Zitron lays out what would need to be true for the trillion-dollar buildout to make economic sense: LLMs would need to become the 'set it and forget it,' hallucination-free, fully autonomous systems the industry has promised, essentially the Star Trek computer, something he calls fantasy given OpenAI's own research shows hallucination is mathematically inherent to the technology. He calculates that even a best-case scenario, where OpenAI and Anthropic together reach Salesforce-level revenue of $40 to $60 billion annually, would still mean spending roughly $540 billion in infrastructure and funding merely to produce 'one more Salesforce or two.' He argues that outcome would still fall vastly short of justifying the investment, since real payoff would require subscription-level pricing and reliability rivaling Netflix or Spotify at many multiples of scale, an outcome incompatible with the expensive, unreliable nature of current LLMs.

He closes by pointing to signs of retrenchment already visible, Tesla capping AI usage, a UBS study showing 60% of enterprises minimizing token usage, and Microsoft deploying armies of consulting engineers just to make AI tools functional inside client businesses, all of which he reads as proof the technology isn't delivering as promised. For Zitron, the entire episode is emblematic of an industry built on unproven assumptions, propped up by circular financing and hype, that is now approaching a reckoning it cannot indefinitely postpone.