Try Pro free

The best 12 min of 31

Why Frontier AI Labs Could Go Bankrupt | Ege Erdil

Ege Erdil argues that frontier AI labs face risky valuations because heavy compute spending may not translate into durable margins.

MTS · Theo Jaffee & Sophia Dew

  • Ege Erdil is a research engineer at Google DeepMind; former Mechanize cofounder and Epoch researcher.

The best minutes

  1. 0:59Frontier labs may struggle to sustain high margins5 min
  2. 9:36Long-term compute commitments could put lab equity at risk3 min
  3. 16:20Compute and data constrain how fast AI research can acceleratePro pick 3 min

Also worth watching

  1. 21:05AI timeline forecasts need measurable real-world outcomes5 min
  2. 12:43Profitable memory stocks look safer than frontier labs to Erdil3 min
  3. 6:09Low switching costs weaken AI labs’ pricing power3 min
  4. 25:34Erdil sees the METR report as weak evidence of AI acceleration2 min

Skip the sponsor reads, especially the mid-episode break and post-interview VCX ad.

Full analysis · 6 min read

Real Alpha · Nov 12

Ege Erdil puts lab bankruptcy risk above 10% as an example

  • Erdil called lab bankruptcy a decent chance, giving more than 10% as an example over three years.
  • His concern is long-term compute commitments meeting uncertain demand, alongside weak customer switching costs.
  • Operators should map contract exposure and test whether model portability limits pricing power.

Free with an account

Read the rest of the analysis

6 sections, the 8 other voices it tests Erdil against, and what it means for you. Every episode we cut gets one.

Signed up before? The same buttons sign you in. The episode, its moments and the transcript stay free.

1:27

Technology importance does not guarantee durable margins

The host said OpenAI had just raised at a $1.25 trillion valuation. Those were host-supplied examples, not evidence of lab earnings.

Erdil said the valuations looked too high because he had not heard a credible path to sustainable high margins. He accepted that the technology could matter enormously.

His distinction is between technology value and company economics. Airlines show why heavy customer spending does not guarantee strong returns for providers. For example, Erdil described airlines as capital-intensive businesses with low margins despite air travel’s importance.

He also pointed to Google’s former search economics, where network effects supported lower capital needs. The comparison does not prove labs must resemble airlines. It asks whether a lab can defend its position without spending continuously.

Our read

The valuation question is not whether AI matters. It is whether labs can retain enough of its value after paying to stay competitive.

5:02

Labs must keep spending to defend model leadership

Erdil said Anthropic could charge about 85% to 90% gross margins on inference. He tied that pricing to having the best model, not a permanent customer advantage.

If a lab pauses frontier spending, he expects to lose its lead within one or two years. That makes current margins a weak proxy for returns after future training costs.

He considered scale as a possible defense, since more capital can buy more compute. But he said frontier labs had not clearly pulled far ahead of Chinese open-source labs.

He said Chinese labs might have about 30 times less compute than Anthropic. Yet he did not see a commensurate lead.

“Their position is entirely dependent on them staying in the lead”
5:19

Our read

The margin number is real only for a moment in the model cycle. The underwriting question is whether each new dollar of compute protects enough pricing power.

8:40

Easy switching limits model providers’ pricing power

Erdil used cloud migration as a contrast. Moving a business from AWS to GCP can be burdensome, while changing model providers can be much easier.

That difference matters because customers can use switching to bargain over price. For example, a lab cannot assume a strong model alone lets it keep premium pricing.

On The Real Eisman Playbook in November 2026, Steve Eisman put flagship models at $2 to $10 per million tokens. He put open-weight models around $0.20 per million tokens.

That price gap makes provider pricing vulnerable if quality differences fail to justify premiums. But model quality and customer workflows may still support differentiated services.

Brendan Foody, Mercor co-founder and CEO, discussed model customization on Sourcery in November 2026. He said hard-to-port models can create switching costs.

“switching costs for AI are really low”
8:40

Our read

Erdil’s case is strongest for uncustomized access where alternatives are easy to reach. Product integrations could change the economics, but providers must prove that friction persists.

10:17

Compute commitments can turn weak demand into equity risk

Erdil’s balance-sheet concern is a duration mismatch. Labs promise to buy compute for years while demand and market prices may change.

He said he thought Anthropic’s commitments were close to $1 trillion over three years. He then said he did not know the exact amount, and the host asked whether the horizon was 15 years.

Erdil separately used a hypothetical $800 billion commitment over ten years. He compared it with Google’s roughly $500 billion in annual revenue, not profit.

If demand disappoints and compute prices fall, a lab locked into higher prices could lose margin. Erdil called that a leveraged bet, not an inevitable bankruptcy.

The compute assets need not vanish in bankruptcy, but shareholders could still be wiped out. He put the risk below 50% over three years, while giving a higher figure only as an example.

Erik Bernhardsson said on Latent Space in October 2026 that renting GPUs can avoid hard-asset exposure but leave long-term commitments. David Sacks said on All-In in October 2026 that CoreWeave’s 2.5-year payback sat against typical five-year deals.

Those are adjacent infrastructure examples, not direct evidence about lab contracts. Sundar Pichai told Big Technology Podcast in October 2026 that Google plans compute against demand and committed revenue. He also acknowledged overbuild can happen.

“more than 10% chance, for example”
12:45

Our read

The risk is not simply high spending. It is a mismatch between long commitments and uncertain demand, prices, and future capacity needs.

13:35

Current profits offer a firmer starting point

Asked where he would invest, Erdil said memory stocks looked better than labs because they already had high margins. He still tied those margins to continued demand.

This is an earnings-versus-expectations distinction, not a guarantee that memory profits persist. It is a firmer starting point than an unproven profit story.

He also remained bullish on Nvidia, saying he expected revenue to keep growing. The host noted Erdil was a shareholder and had put Nvidia’s fair value at $15 trillion.

Erdil agreed with that fair-value view and called a recent selloff an overreaction. He said short-term market moves may say little about long-term profitability.

His existing AI-correlated positions limit the neutrality of his portfolio perspective. Readers should treat the picks as his opinion, not independent valuation evidence.

“memory stocks look way better to me than the labs”
13:35

Our read

Erdil compares present earnings with a speculative route to future profits. His Nvidia view is also a forecast from a disclosed shareholder, not independent proof of value.

21:30

Forecasts need observable outcomes and disclosed evidence

Erdil asked for tangible outcomes that could show whether intelligence-explosion forecasts were wrong. He objected to forecasts resting on abstract metrics without specified physical consequences.

He argued that forecasters should define their own tests rather than leave skeptics to invent them. That makes forecast tracking possible without accepting his timeline.

His own case for compute constraints is suggestive, not decisive. He linked progress in chess, SAT solving, and computer vision with faster compute scaling. He said the pattern was observational, not causal.

Jeffrey Ladish said on MTS in November 2026 that METR’s estimated 2x acceleration deserved attention. He also asked Anthropic to release the underlying data.

Erdil gave the result little weight because the metric was hard to interpret and important evidence remained redacted. Both men point to withheld evidence, but assign the available result different weight.

Herbie Bradley said on MTS in November 2026 that fast takeoff remains low probability. He also said progress may depend on data.

Richard Ngo said on MTS in November 2026 that forecasts can outrun evidence. He warned against inferring rapid self-improvement from partial automation.

Finally, Erdil put full automation of white-collar work maybe 20 years away. That estimate is his forecast, not a deadline established by the episode.

“most of them are not making any concrete tangible prediction”
22:34

Our read

The practical standard is a named outcome with a date and a failure condition. Erdil’s maybe 20-year estimate reinforces his long horizon, but does not settle the debate.

What it means for you

  1. Map compute obligations. List contract duration, minimum commitments, and exit terms beside your demand forecast. Stress-test the economics against lower compute prices and weaker usage.
  2. Test model portability. Measure how long it takes to switch models without changing workflows. If that test is easy, treat provider pricing power as uncertain.
  3. Separate earnings from stories. When comparing AI businesses, distinguish current margins from promised future margins. Record what demand must do for each to persist.
  4. Set forecast failure conditions. Ask teams to name observable customer or operational outcomes that would disconfirm their AI timeline. Track the result against the forecast.

What to watch

  • Whether labs disclose more detail about the duration and scale of compute commitments.
  • Whether model customization makes provider switching materially harder for customers.
  • Whether METR’s underlying evidence becomes available for independent assessment.

Send this to the CFO and infrastructure lead reviewing compute commitments.

9:41

Watch the compute-risk discussion, with Erdil’s odds near the end

Written by Real Alpha from the episode's transcript and the other moments it links, then checked against the transcript. Quotes are the speaker's exact words.

Minute by minute

  1. 0:59Frontier labs may lack durable margins

    Erdil says the labs’ valuations assume they can turn important technology into durable high-margin businesses. He compares them to airlines: the technology can matter enormously while the businesses still require high capital spending and produce low margins. He says durable high-margin businesses usually have a moat and low ongoing costs.

  2. 6:09AI may not be winner-takes-all

    Erdil says labs must keep investing to maintain the lead, and he questions whether scale gives them a decisive advantage. He notes that Anthropic has not pulled ahead as much as its resource gap over Chinese open-source labs—perhaps 30 times more compute—would suggest. He also says customers can switch AI models more easily than cloud providers.

  3. 9:36Compute contracts create failure risk

    Erdil says labs could fail or restructure while their compute assets remain, wiping out equity holders. He estimates Anthropic’s commitments could approach $1 trillion, but cautions that he is unsure of the exact number and timing. He puts the chance of lab bankruptcies over the next three years above 10% but below 50%, citing the risk that falling compute prices leave long-term contracts underwater.

  4. 12:43Erdil favors memory stocks and Nvidia

    Erdil says memory stocks look safer to him because they are already high-margin businesses, provided demand stays strong. He calls OpenAI riskier than Anthropic. He also agrees with a $15 trillion fair-value estimate for Nvidia, arguing that its revenue can keep growing and that recent market sell-offs may overreact.

  5. 15:53Compute and data constrain AI timelines

    Erdil rejects the idea of a near-term intelligence explosion, arguing that automating AI development still depends on data and compute. He says cognitive labor has diminishing returns once those inputs become bottlenecks. He points to correlations between compute scaling and software progress across fields including chess, SAT solvers, natural language processing, and computer vision.

  6. 21:05AI forecasts need testable outcomes

    Erdil says people predicting an intelligence explosion should specify observable events that could prove them wrong, rather than relying on abstract metrics. He cites a prior forecast that AI would execute machine-learning experiments from high-level instructions within about three years as the kind of testable claim he values. He says predictions about robots also need clear definitions.

  7. 25:34Erdil calls METR evidence weak

    Erdil says the METR report offers weak evidence for accelerating AI research: he questions what its progress metric measures and notes that some supporting evidence is redacted. He also argues that social pressure can shape how people update their views, and says he has not seen decisive evidence in either direction.

  8. 27:51White-collar automation may take 20 years

    Erdil revises his earlier estimate of 30 years to roughly 20 years for full automation of white-collar work. Asked what to put to Buck, he suggests asking for a concrete forecast; he says a tenfold rise in global primary consumption over the next three years would make him admit he was wrong.

Published Nov 12, 2026.

Transcript · 24 min read

MTS, hosted by Theo Jaffee and Sophia Dew, with Ege Erdil.

Transcribed by machine (Whisper) from the episode’s audio, so a word may be off here and there.

  1. For Anthropica, I think it's close to a trillion dollars that they have committed over the next three years. And that's a very large amount. Like Google's annual revenue is only in order of like $500 billion. That's revenue, that's not even like profit. So if you are making a spending comment of like $800 billion over 10 years, like that's basically like you're saying I'm going to spend Google's entire income for 10 years. And I think that that is a real danger because labs have to enter into these long-term, multi-year contracts for compute. But it means the equity holders could get wiped out.

  2. All right, we are back. We are live with Ege Erdil, who's a research engineer at Google D-Mine, previously co-founder of Mechanize and a researcher at Epoch. We love Epoch. Ege, welcome to MTS. Yeah, thanks for having me. so you have a lot a lot of takes a lot of things to talk about um one thing i'm definitely

  3. Ege Erdil · 5 min

    “Their position is entirely dependent on them staying in the lead, which is a very expensive thing to do.”

Read the whole transcript
  1. There's definitely evidence against it because if you were seeing the opposite, it would have been evidence for it. So conservation of expected evidence.

  2. Ege Erdil · 3 min

    “So you can very easily just switch to a different model. It's completely trivial.”

  3. build? The internal tool, side project, or product you'd otherwise lose a weekend to? With Lovable, you get software you can ship today. That includes an editable code base you can inspect and change, plus two-way GitHub sync. Lovable also handles the backend and infrastructure, including managed Postgres, auth, storage, hosting, payments, and more, so you don't have to wire it all up yourself.

  4. And through Lovable's MCP server, you can create and deploy your project directly from the agents you already use. Turn ideas into software people love with Lovable. Lovable.dev. Now,

  5. Ege Erdil · 3 min

    “But if for whatever reason, the demand for AI just doesn't materialize as much as we think, then compute prices fall dramatically, then all of these commitments are underwater.”

  6. Ege Erdil · 3 min

    “memory stocks look way better to me than the labs.”

  7. You had this famous debate with... Daniel Cucatazzo and Ajay Khotra back in 2023, where you had much longer timelines than either of them. Now, Ajay has updated substantially towards very short timelines. Same as Daniel, they put out this blog post yesterday saying it looks like the intelligence explosion will happen in 2027. You disagree with this. So a couple of questions are, why do you disagree? Where are the

  8. Ege Erdil · 3 min

    “I think a lot of the progress is just driven by data and compute.”

  9. We'll continue monitoring right after this message from our sponsors. Startups have already chosen Neon for their back end. Start on the free plan or get up to $100,000 in credits for your startup at neon.com slash MTS. Special thanks to our sponsor Kong, the AI connectivity platform. Connect APIs, LLMs, agents, and systems with serious security and governance. KongHQ.com. Finally prove what your AI investment is worth with Lairdyn. Head to Lairdyn.com today to start measuring your AI's bottom line. MongoDB. Trusted to run critical applications by 75% of the Fortune 100, the intelligent data platform for the era. Apply today for access to up to $15,000 in Atlas credits, dedicated technical support, and exclusive events and go-to-market opportunities. MongoDB.com. Chargebee. Billing and monetization infrastructure for the AI era. Chargebee.com. ClickHouse is the real-time database built to keep up with every query your apps and agents send its way. And it now runs Postgres, observability, and AI tooling on one platform. Learn more at clickhouse.com.

  10. So what do you expect the world to look like in 2029 such that you can say,

  11. Ege Erdil · 5 min

    “it has to be something with like real world sort of physical consequences”

  12. Ege Erdil · 2 min

    “I just see it as very weak evidence.”

  13. People make a very big deal out of something like this. Do you think full automation of white-collar work is still 30 years out?

  14. 30 years out from now. I mean, I said it was 30 years out some time ago. So, you know, it's— Like a couple years ago. It's still roughly 30 years out? I think full automation of white-collar work maybe was like 20 years out. Okay. All right. Well, final question before we go. After you, we're going to talk to Baksh Ligaris, who's the CEO of Redwood Research and is, you know, much more of a doomer and has also updated towards, you know, know much shorter timelines so like what should we ask him um i guess you can just ask him for the for the thing that we were discussing right we're like uh what would you have to see over the next two or three years for you to admit that you're wrong and i think like like one thing that i can say is is that um i don't know like like if we if we have if like like global primary consumption option goes up by 10x over the next three years like it's like a very hard metric where i was like yeah i was wrong you know like and uh i want to understand if they're making that prediction because i think they might say no no we actually don't predict that like we don't think that's going to happen but then i want to understand what is the thing they do predict is going to happen

  15. so i can figure out like do i actually disagree with that because like sometimes it's actually not even clear like what we're talking about yeah right well this is super interesting um Yeah, we love talking to, like, informed short timeline skeptics because there's so few of them left. So, yeah, keep fighting the good fight. Thanks so much, Ege, for coming on MTS. Thanks, Ege. Bye. performance in the real world. Arena.ai. Support for the show comes from VCX, the public ticker for private tech. The U.S. stock market started history's greatest wave of wealth creation. From factory workers in Detroit to farmers in Omaha, anyone could own a piece of the great American companies. But today, our most innovative companies are staying private longer, which means everyday Americans are missing out. Until now. Introducing VCX, a public ticker for private tech.

  16. visit getvcx.com for more info that's getvcx.com carefully consider the investment material before investing including objectives risk charges and expenses this and other information can be found in the innovations fund prospectus at getvcx.com this is a paid sponsorship

More interviews