Better Offline

Monologue: Meta Admits It Bought Too Many GPUs

6 min
Jul 3, 202615 days ago
Listen to Episode
Summary

Host Ed Zitron argues that Meta's reported plans to sell cloud AI compute access is an admission that it massively overbuilt GPU capacity with no clear business case. He contends there is no viable business model in generative AI, with real demand concentrated almost entirely in OpenAI and Anthropic, both of which are deeply unprofitable. Zitron also highlights Nvidia's reported financial backstops to cloud GPU renters as further evidence that organic demand for AI compute does not exist at scale.

Insights
  • Meta's pivot to selling cloud AI compute is best interpreted as an admission of GPU overcapacity, not a strategic expansion into cloud services.
  • Nvidia offering financial backstops to GPU cloud renters signals that natural market demand for AI compute is insufficient to sustain the current infrastructure build-out.
  • The generative AI compute market is dangerously concentrated, with OpenAI and Anthropic estimated to represent 70–85% of all global AI compute demand.
  • The AI infrastructure investment cycle may be a debt-fueled bubble, with trillions in projected revenues that analysts argue will never materialise.
  • Sam Altman's reported proposal to give 5% of OpenAI to the US government is framed as a potential precursor to a government bailout, though congressional approval would be required.
Trends
Hyperscalers and AI-first companies are beginning to monetise excess GPU capacity by entering the cloud infrastructure market, signalling overcapacity.Nvidia is subsidising its own demand chain by offering revenue-share deals to GPU cloud providers, a potential indicator of weakening organic AI compute demand.Mainstream financial media is beginning to cover AI bubble narratives, suggesting a shift in consensus sentiment.AI compute demand is consolidating around a small number of loss-making frontier model companies, creating systemic financial risk.Corporate AI strategies are increasingly reactive and directionless, driven by competitive pressure rather than clear product-market fit.Government entanglement with AI companies (e.g., OpenAI's reported equity offer to the US government) may foreshadow regulatory or financial intervention.The AI hype cycle is showing signs of entering a correction phase, with infrastructure overcapacity and lack of profitability becoming harder to ignore.Neo-cloud GPU rental businesses are structurally unviable without vendor subsidies, pointing to a shakeout in the cloud infrastructure tier.
Topics
Meta GPU overcapacity and cloud infrastructure monetisationNvidia financial backstops to GPU cloud providersGenerative AI business model viabilityAI compute demand concentration (OpenAI and Anthropic)AI infrastructure investment bubbleMeta AI strategy and internal reorganisationSam Altman's proposed OpenAI equity offer to US governmentScale AI acquisition by MetaInference cost and profitability at scaleMasayoshi Son and SoftBank AI investment strategyAI hype cycle and potential market correctionCorporate governance failures in Big Tech AI spendingNeo-cloud GPU rental market sustainabilityMainstream media coverage of AI scepticismAI consumer product failures (Meta smart glasses, AI chatbots)
Companies
Meta
Reported to be developing a cloud infrastructure business to sell excess AI compute capacity after overbuilding GPUs.
Nvidia
Reportedly offering financial backstops to GPU cloud renters in exchange for revenue share, signalling weak organic d...
OpenAI
Cited as one of only two companies with meaningful AI compute demand, yet loses tens of billions annually with no pro...
Anthropic
Named alongside OpenAI as one of the only real customers for AI compute, estimated to share 70–85% of global demand.
Scale AI
Acquired by Meta for $14 billion, described by host as an inexplicable strategic decision.
Bloomberg
Reported Meta's plans to develop a cloud infrastructure business selling AI compute access.
Reuters
Published a piece by Jeff Horwitz linked in the episode regarding a death connected to Meta's AI chatbot.
The Information
Reported on Nvidia's financial backstop arrangements with GPU cloud rental companies.
SoftBank
Referenced in the context of Masayoshi Son's shareholder speech and AI hype cycle, subject of an upcoming episode.
Amazon Web Services
Cited as an example of a tech mythology used to justify AI infrastructure investment that host argues misled investors.
CNBC
Host appeared on CNBC to discuss AI bubble concerns, noting mainstream media is now covering these stories.
People
Ed Zitron
Host delivers a monologue arguing Meta's GPU overcapacity and Nvidia's subsidies prove the AI bubble is collapsing.
Mark Zuckerberg
Quoted saying Meta could sell its compute if it didn't find a use for it, cited as evidence of strategic aimlessness.
Sam Altman
Reported to be proposing giving 5% of OpenAI to the US government, framed as a potential bailout precursor.
Jeff Horwitz
Authored a Reuters piece about a death linked to Meta's AI chatbot, referenced and linked in the episode.
Andrew Bosworth
Named as one of Zuckerberg's inner circle of yes-men enabling poor strategic decisions at Meta.
Javier Olivan
Named alongside Bosworth as a senior Meta executive who enables Zuckerberg's strategic misdirection.
Masayoshi Son
Referenced in relation to an upcoming episode about his shareholder speech and the AI hype cycle collapse.
Kal Penn
Appears in a post-episode ad read promoting the Hearsay audiobook club podcast on iHeart.
Quotes
"I think I spent $100 billion for a good reason. Dunno yet though."
Ed Zitron
"It's blatantly obvious that demand doesn't exist at scale for AI Compute, that the only real customers are two bulbous fail sons, Anthropic and OpenAI, that lose tens of billions of dollars a year with no path to profitability."
Ed Zitron
"There is no business model in generative AI. There is no way of serving inference at scale in a way that is consistently profitable."
Ed Zitron
"The debt fueled hyperscaler Capex cash dump has pumped a global stock market and left the world invested in an impossible payoff of trillions of dollars of revenue that will never exist."
Ed Zitron
"Nvidia is offering financial backstops to young cloud providers that rent its GPUs in exchange for a share of their revenues — a euphemism for literally paying their customers to buy their stuff."
Ed Zitron
Full Transcript

This is an iHeart podcast. Guaranteed Human this July 4th at Lowe's get up to 45% off select major appliances, plus save $80 on the select Char Broil Performance Series gas Grill now $299. Our best lineup is here at Lowe's Lowe's we help you Save valid through 7 8, while supplies last selection varies by location. See Lowes.com for more details. Visit your nearby Lowe's. Media Greetings, Greetings, hello and salutations. I'm Ed Zitron and this is your Better Offline monologue for the week. Better Offline earlier this week, Bloomberg reported, and I quote, that Meta is developing plans for a cloud infrastructure business that will sell access to its AI computing power and models. A few weeks after CEO Mark Zuckerberg said at its annual shareholders meeting that he thought that Meta had a use for its compute but it could sell it if it didn't. I genuinely love that he said that. It's like I think I spent $100 billion for a good reason. Dunno yet though, and there's no other sensible way at least to interpret this news other than that Meta is admitting that it overbuilt its GPU capacity after five different reorganizations of its AI team, a $14 billion acquisition of Scale AI for no imaginable reason, multiple versions of their loathsome pervert glasses and celebrity faced AI chatbots that have led an old man with dementia to his death. I'm not kidding. I'll link you to that great Jeff Horwitz piece from Reuters. In reality, Meta never had any reason to get into AI other than the fact that Zuckerberg is a directionless oaf that can never be fired from the company, surrounded by loathsome yes men like Boz, Adam Mirsky and Javier Olivan who cater to his everywhere and say yes to any idea, no matter how goddamn stupid they are. Every time I hear from somebody at Meta, they vibrate with sadness and despair, lost in a company that exists as a large social experiment to try and conjure up more ad revenue for a platform designed to engage far more than build any meaningful connections. AI bulls and the recently concussed argue that this is actually a smart move because now Meta can monetize its AI capex, the kind of ridiculous thing you say when you like big tech, are fundamentally out of ideas and horribly desperate. Meta is doing this because Meta has no idea what to do other than to try something, anything to make some money out of these God damn GPUs. The real question is who Meta will sell to and how much it will sell. Epoch estimates that it has around 1.61 gigawatts of total capacity as of the fourth quarter of 2025. And if that's the case, it's invariable that it gets snapped up by either Anthropic or OpenAI. And by the way, when that happens, we're at the top. We're done. Wrap this bad boy up. And if it doesn't do that, I'm not sure what the demand looks like outside of those two companies, who I estimate take up anywhere from 70% to 85% of all global AI compute. Meanwhile, the information reports that Nvidia is offering, and I quote, financial backstops to young cloud providers that rent its GPUs in exchange for a share of their revenues, a euphemism for literally paying their customers to buy their stuff. This is a direct admission that Nvidia doesn't see there's a diverse or rigorous demand for AI Compute and that the only way that these horrible little Neo clouds can ever run a business renting out GPUs is by letting them latch onto the side of Nvidia like a parasite. It's blatantly obvious that demand doesn't exist at scale for AI Compute, that the only real customers are two bulbous fail sons, Anthropic and OpenAI, that lose tens of billions of dollars a year with no path to profitability. This is not a real industry, but a reflection of the desperation and directionlessness of a tech industry that's run out of hypergrowth ideas. After giving up on investing at the seed stage and chasing out those who actually build real things, this industry deserves the punishment that's coming for it. The debt fueled hyperscaler Capex no, it loads refused cash dump has pumped a global stock market and left the world invested in an impossible payoff of trillions of dollars of revenue. Well, that will never exist. Everybody involved in inflating this bubble should be ashamed of themselves. Everybody who refused to heed my warnings in favor of the mythologies of Amazon Web Services and Uber should know that they are partly to blame for the chaos that will follow. This was blatantly obvious years ago. It was obvious before I started this show. It's been so obvious for so long. I've got better at this. I've got more nuanced in my arguments, but my arguments have stayed the same. There is no business model in generative AI. There is no way of serving inference at scale in a way that is consistently profitable and People will argue about this and people will have the little natterings on their various blogs. I won't read them. I've got stuff to do. Sorry. But the truth is that everything I'm saying has been coming true. I've been trying to warn people. All of you, I know I'm speaking to an audience that kind of gets it already. All of you have been sticking with me and I love it. I was just on cnbc, did a great hit there. I'll link to it in the notes. Hate to link to Twitter, but it's one of the only places I can actually upload it without going dmca. But I think you're really going to like that interview. And the fact that the mainstream is finally opening up to these stories is a sign that we're wrapping this up. And they're trying to kind of layer on the news and layer on the stories that say, hey, we didn't fall behind this. We were on this. And good on them for doing it, albeit a little too late. In fact, I think all of the kind of bubble discussion is too little too late. And now, Glammy Scammy Klammy Klamath, we love glamour. We don't like Clemantha. He wants to give 5% of OpenAI to the US government, I assume in exchange for money. And I know some of you are going to say that's a lineup to a bailout. No, it isn't. It'll require congressional approval. And this is at a time when AI is one of the least popular things in the world with voters. And also no one wants to give $40 billion to goddamn Sam Altman. All signs to this. All signs point to this, I mean, blowing up sometime, I think in the next six months. And I feel like I've been saying that for about six months, but we're seeing the desperation start and we're hearing these conversations happen. I think things are going to get weird. Things are going to get very weird. And next week we've got a great joint episode with the Trash Future crew talking about Masayoshi Son's goose filled shareholder speech and the collapse of the AI hype cycle. You're going to love it, I promise. Thanks for listening, Zitron. Out foreign, It's Kel Penn. I'm inviting you to join the best sounding book club you've ever heard with my podcast, Hearsay, The Audible and iHeart Audiobook Club. Every episode I nerd out with amazing guests and dive into the best new audiobooks available on Audible. It's the book club for your ears. Listen to Irsay the Audible and I Heart Audiobook Club on the iHeartRadio app or wherever you get your podcasts. It's called soccer. It's called Football. Soccer. Football. Domino's best deal ever. Lets you get any pizza, including stuffed crust with any toppings for $9.99. Okay, we can agree on that. Yeah, fully. So pineapple. Don't ruin it. Get any pizza, including stuffed crust, with any toppings for 9.99. Finally, something everyone can get behind. And if the rest disagree, that's between them and Domino's. Which means the only thing left to fight over is who's ordering. Dom Minnows. Price is higher for some locations. Excludes XL and specialty pizzas. Select this offer from $6.15 to $7.26 online only. Size availability varies by crust. Tank max 7 toppings, 6 for pan and New York style crust. Minimum purchase required for delivery. Prices, participation, delivery area and charges may vary. M&M's popped caramel do sound different. Oh, no. People are gonna be obsessed. What do you mean? People hate the sound of chewing. Maybe they won't like the crunch. Maybe we're saved. Wait a minute. Yellow. Have you been eating them this whole time? Mmm. So tasty. Hands off us. M&M's popped caramel. It's more fun together. 250 years ago, America made a promise of life, liberty and the pursuit of happiness. Since 1903, Harley Davidson has been living it out on the open road at the next exit ramp with old friends, new ones, and the next generation of riders. Because the best part of any promise is keeping it. Harley Davidson Motorcycles ride. This is an iHeart podcast. Guaranteed Human.

0:00