Meta's Compute Confession
Zitron opens by cataloguing Meta's dysfunction: five reorganizations of its AI division, an open-source Llama model nobody would pay for, a $14 billion acquisition of Scale AI, and a Behemoth-class model, Muse Spark, that landed as merely middle tier despite over $100 billion in capex. Against that backdrop, reports that Meta is now considering leasing out its compute to outside customers reads to him not as clever monetization but as a confession. Just weeks earlier at Meta's shareholder meeting, Mark Zuckerberg had said the company believes it has a use for the compute but will rent it out if it doesn't, a hedge Zitron treats as an admission that Meta never had a real plan.
He pushes the implication further: most of Meta's fleet is aging H100 and H200 silicon, meaning the company risks becoming, in his words, the largest provider of outdated GPUs rather than a serious AI player. The deeper question, he says, is why Amazon, Google, and Microsoft aren't scrambling to offload their own surplus the same way, and his answer is that Anthropic and OpenAI are currently absorbing nearly all the excess demand in the industry. That leads to his starkest prediction of the episode: if Meta ends up selling capacity directly to Anthropic or OpenAI, rather than finding independent customers, it signals there is no diverse demand left anywhere, and the industry has effectively reached its endpoint.
"This is a sign they don't need the compute. And it's also a sign that perhaps they overbuilt."
A Market Built on Two Unprofitable Customers
Zitron traces contract after contract, Nebius, CoreWeave, Lambda, cipher mining, and shows they all lead back to the same handful of names: Microsoft for OpenAI, Google for Anthropic, Meta, and Nvidia buying its own chips back. He estimates Anthropic and OpenAI together consume around three and a half gigawatts of compute, the overwhelming majority of what's actually available, while independent demand beyond them amounts to perhaps 500 megawatts to a gigawatt, a fraction of what's been built or financed. He notes that both Anthropic and Google's compute commitments, when lined up, appear to exceed what could plausibly be delivered simultaneously, suggesting even the headline deals are shakier than they look.
He also raises Nebius's roughly $17 billion Meta deal and CoreWeave's roughly $30 billion one, asking pointedly whether those companies used Meta's contracts to raise financing to build data centers Meta may no longer need. Hyperion, Meta's $30 billion, Manhattan-sized data center project in Louisiana, becomes a symbol of this uncertainty, since nobody, including Meta's own AI leadership, seems to articulate what it's actually for. Zitron mentions that Meta's AI chief, Alexander Wang, publicly clarified that Zuckerberg's comments about slower-than-expected agent progress applied to the whole industry, not just Meta, a clarification Zitron says only soured investor mood further.
"Are we we near over a trillion dollars now. Where are the outcomes?"
The Efficiency Trap
The conversation turns to a report from The Information claiming OpenAI engineers found a way to more than halve inference costs. Zitron is skeptical, arguing that if such a breakthrough were real, OpenAI would trumpet it publicly rather than let it leak as an offhand Slack exchange. But he uses the story to expose a deeper bind facing the entire industry: if AI companies actually cut inference costs, they need less compute, which exposes the overbuilt supply; if they don't cut costs, they need endless venture capital and debt to keep buying compute, which will eventually run dry. Either path, he says, ends badly.
He extends this logic to Nvidia's new partnership program, which offers to lease back or repurchase GPUs from cloud providers in exchange for a cut of their profits. Rather than reading it as reassurance for jittery customers, Zitron calls it a near-admission that real demand doesn't exist, comparing it to round-tripping, a practice he says would draw regulatory scrutiny if the SEC were functioning properly. He details how CoreWeave and Nebius use hyperscaler contracts as collateral to secure bank loans for data center construction, meaning the entire financing structure depends on circular commitments rather than end-user revenue, a dynamic he repeatedly likens to the dot-com era collapses of Lucent and Nortel.
"We're in this weird situation where everyone wants them to reduce their cost. But if they reduce their cost, they reduce the need for AI compute, which means we've overbuilt the supply."
Who Blinks First on Capex
Zitron predicts that once one hyperscaler pulls back capital spending, the market will reward it, prompting the others to follow within a similar timeframe, all while dressing the retreat in language about an era of efficiency rather than admitting AI isn't delivering returns. He initially expected Meta to be the last company standing but now believes it could be the first to cut, given its compute-leasing discussions, while flagging Amazon's planned $200 billion in spending as another likely candidate for trimming. He also points to Microsoft CEO Satya Nadella's announcement of a billion-dollar investment in forward-deployed engineers as evidence of what he calls full AI psychosis, arguing that needing armies of consultants to make AI functional for businesses is itself proof the technology doesn't work as advertised.
He stresses that neither OpenAI nor Anthropic has ever built its own infrastructure, estimating that hyperscalers have spent roughly a quarter of a trillion dollars building it for them, including around 80 billion dollars just from Microsoft, a figure surfaced in the Musk v. Altman trial. If Microsoft, Google, and Amazon cut back, he argues, OpenAI and Anthropic would suddenly need to fund their own infrastructure, something he doubts they can afford, setting up what he calls a mess waiting to happen.
"They're an industry of cowards. They don't have any ideas. They just they're doing this because they have no other hypergrowth ideas."
The Memory Squeeze That Will Hit Everyone
Zitron shifts to Samsung and SK Hynix's announced plan to spend over $500 billion expanding memory fabrication capacity, connecting it directly to the AI buildout's demand for high bandwidth RAM used in GPUs. Because that premium RAM eats up fab space, ordinary DDR memory used in consumer electronics is getting squeezed, and SK Hynix is reportedly removing price caps on long-term deals and extending minimum agreements from one year to three, giving suppliers leverage to raise prices further. Zitron warns this will keep RAM and NAND storage prices elevated at least until 2028, already visible in the growing number of laptops sold without RAM or with reduced storage to keep sticker prices down.
He frames this as a boom-bust industry making an unsustainable bet: if the AI bubble deflates and demand for high bandwidth RAM collapses, Samsung, SK Hynix, and Micron will be left holding overbuilt capacity, much as they were only a few years ago after the pandemic RAM glut. Unlike that episode, though, he expects the fallout to hit consumers harder, since these companies are unlikely to lower prices even as demand falls, given margins at Micron already reportedly near 84 percent. His verdict is unambiguous: the inflation in electronics prices is a direct, foreseeable consequence of AI infrastructure demand, and the people who inflated the bubble should be held responsible for it, even though he expects they largely will not be.
"The inflation caused by the AI industry is their fault and they should suffer for it. They won't. The rich people will be fine at the end of this."
What LLMs Would Have to Become
Asked what leap in capability could justify the spending, Zitron paints a picture of AI as pop culture imagines it, a Star Trek computer or Commander Data, fully autonomous, error-free, and invisible in its operation, contrasted against what large language models actually deliver: expensive, unreliable, dependent on careful prompting and expert oversight to catch hallucinations. He notes that OpenAI's own research treats hallucination as mathematically inevitable, meaning the fantasy of a self-correcting, always-right assistant is not on the table. For the economics to work, he argues, AI would need to become cheaper than metered billing allows, stickier than Spotify or Netflix, and profitable at a scale ten times larger than Netflix's current revenue, spread across just two companies.
He runs the numbers on a best-case outcome where OpenAI and Anthropic reach Salesforce-level revenue of 40 to 60 billion dollars annually, only to conclude that even success on that scale would represent roughly 540 billion dollars spent to produce the equivalent of one or two more mid-sized software companies. He closes by citing signs of retrenchment already underway, Tesla capping AI usage at 200 dollars a week, a UBS study finding 60 percent of enterprises minimizing token usage, and the absence of fresh annualized revenue figures from Anthropic, as evidence that the revenue story is stalling rather than accelerating toward the scale needed to justify the buildout.
"LLMs are a con perpetuated by con artists. And I'm sick of them."