Big Technology Podcast · · 1h 1m
Why The Anthropic IPO Will Fail — With Gary Marcus
With Gary Marcus
Summary
Gary Marcus argues on the Big Technology Podcast that Anthropic's IPO will fail to deliver returns, because the company is expensive to run, lacks a durable moat, and carries roughly $900B in future liabilities against about $100B a year in costs. He points to slowing revenue growth after mid-2026, with Q3 growth around 47% versus 147% in Q2, which he attributes to the end of the token-maxing fad, competition from cheap Chinese and open-weight models, and enterprise defections such as Goldman Sachs. He expects Anthropic to be acquired for far less than the IPO valuation, perhaps around $400B, rather than fail outright, while noting that Claude Code is a genuine neurosymbolic success. The discussion also covers OpenAI's weaker fundraising at a lower valuation, the diminishing returns of scaling compute, export-control politics involving Nvidia and Jensen Huang, and his view that AGI is likely next decade.
Key takeaways
- Gary Marcus (bullish on Anthropic): Anthropic's Claude Code is neurosymbolic AI, vindicating Marcus's hybrid-architecture thesis over pure scaling.
- Gary Marcus (bearish on Anthropic): Anthropic is a bad business: unprofitable, possibly never profitable, facing bankruptcy or a deep-discount sale.
- Gary Marcus (bearish on Anthropic): Anthropic's models face price wars and commoditization as open and Chinese models catch up.
- Gary Marcus (bearish on Anthropic): Anthropic's 147% quarterly growth projections are clearly wrong; growth has slowed since July as token-maxing fades.
- Gary Marcus (bearish on Anthropic): Costs of each capability gain are exploding, so Anthropic's frontier-model spending is unsustainable.
What was said, by company and theme
20 views from 1 speaker. Every quote is checked against the episode's transcript (how).
Anthropic
Across the shows: 125 speakers in other episodes, 189 bullish and 92 bearish calls, past 30 days
Where everyone stands on Anthropic
Gary Marcus · New York University (professor emeritus)Bullish
Anthropic's Claude Code is neurosymbolic AI, vindicating Marcus's hybrid-architecture thesis over pure scaling.
Marcus has long argued that pure scaling of large language models is not enough and that symbolic components are needed. He says Claude Code is as neurosymbolic as you can get: it is a large language model integrated with conditionals, loops, tools and abstract representations drawn from classical AI. Scaling was part of it, but the symbolic additions make it a victory for neurosymbolic AI rather than pure scaling. He wrote that this was great when Claude Code came out. This is a positive view of the technology, separate from his view of the business.
“Nobody talks about it, but Claude code is as neurosymbolic as you can get. It is a big, big, large language model. Scaling was part of it.”at 1:15
Gary Marcus · New York University (professor emeritus)Bearish
Anthropic is a bad business: unprofitable, possibly never profitable, facing bankruptcy or a deep-discount sale.
Anthropic is expensive to run and has a good product but not a good business. Its only defensible niche is the highest-end models, since it gets undercut on cheaper tiers. Marcus does not expect profitability anytime soon and says it is not clear it will ever be profitable, given about $900B of future liabilities and roughly $100B a year in bills. If it cannot pay those bills it could fail, though he thinks it more likely it will be shopped around and acquired by a company such as Nvidia, Amazon, SpaceX or Apple for far less than the IPO valuation, perhaps around $400B after back-and-forth.
“I don't think that they're gonna be profitable anytime soon. It's not clear that they will ever be profitable, and I think there's a real chance they will go bankrupt.”at 3:54
Gary Marcus · New York University (professor emeritus)Bearish
Anthropic's models face price wars and commoditization as open and Chinese models catch up.
Marcus argued in 2023 that since everyone was building the same thing, large language models would become a commodity with price wars and hard-to-make profits. He says this has happened faster than he expected, with Chinese models arriving quickly and by mid-2026 Anthropic struggling to find a market niche because people used Chinese models. Anthropic's only sustainable niche is a small set of hard problems where its top model is needed, and open-weight models keep improving and eating that slice from below. Cheap alternatives are available on OpenRouter.
“everybody's building the same thing, and that means you're gonna have a commodity where there are price wars and profits are gonna be hard.”at 7:42
Gary Marcus · New York University (professor emeritus)Bearish
Anthropic's 147% quarterly growth projections are clearly wrong; growth has slowed since July as token-maxing fades.
Q2 revenue grew 147% over Q1, but Q3 grew only about 47% over Q2 (per the host's figures). Marcus attributes the slowdown to the end of the token-maxing fad, where companies told employees to spend as much on AI tokens as possible, and to competition from Chinese models. He says Anthropic took off for about a year and then slowed around July. He compares the projections to a 'trillion pound baby' fallacy: exponential growth usually tails off. Because the company must keep growing fast to cover its commitments, the slowdown is a serious problem.
“these projections of them making, you know, 147% growth every quarter are just clearly wrong. And if they're clearly wrong and they have $900,000,000,000 in liabilities, the math does the math.”at 8:33
Gary Marcus · New York University (professor emeritus)Bearish
Costs of each capability gain are exploding, so Anthropic's frontier-model spending is unsustainable.
Each new model improvement costs far more than the last, and the host cited a chart showing Anthropic gaining exponentially less capability per billion dollars spent. Marcus illustrates the cost curve: $1B to impress users this year, $10B next year, $100B after that, and eventually $1T for a few more benchmark points. That is not sustainable. Even the most expensive models still hallucinate and have not achieved AGI, so spending keeps rising while some hard problems remain unsolved. Users will eventually decide the premium is not worth paying for.
“if you have to spend $1,000,000,000 to impress your users this year, [$]10,000,000,000 to impress your users next year, and $100,000,000,000 the year after... At some point, that is just not sustainable.”at 11:32
Gary Marcus · New York University (professor emeritus)Bearish
Large customers moving off Anthropic's tools makes its needed hypergrowth much harder to sustain.
Microsoft and Meta have competing tools, and Goldman Sachs told its bankers and traders to stop using Anthropic's tools, keeping Claude Code only for coders. Losing customers does not make the business implode, but it makes continued growth much harder. Anthropic needs to keep growing like a rocket because it is losing money at current levels and carries large liabilities. Each customer that pulls back or limits internal use reduces its chance of reaching escape velocity.
“they are moving off Anthropic. I think Goldman Sachs was also... Goldman, not everybody left, but they basically told their bankers and traders don't use this.”at 18:52
Gary Marcus · New York University (professor emeritus)Bearish
Anthropic IPO buyers may be hurt; day-one hype is meme-like and the valuation is not sustainable.
Marcus says the IPO-demand argument is his weakest, since it depends on public sentiment, and he may be wrong about oversubscription. Still, a shaky round would look bad and could start a death spiral, and OpenAI has already shown weaker terms. The IPO day looks like a meme-stock event, as SpaceX did. The real question is long-term return on the investment. If buyers pay a valuation that climbs back toward $1.5T, he doubts they will earn their money back, because the company is not sustainable.
“If everybody buys the IPO at a valuation that ultimately gets back up to $1,500,000,000,000 or whatever, I have serious doubts about whether that investment's going to return in the long run.”at 30:58
Gary Marcus · New York University (professor emeritus)Bearish
Anthropic has overcommitted to compute, and cheaper local models could shrink demand for its cloud infrastructure.
Anthropic has roughly $900B of future liabilities and committed to too much compute, and Marcus expects it will try to renegotiate those commitments. He thinks the company will not need all of this compute. Local models are improving, so people will rely less on giant data centers and on providers. Rising data center costs, from local resistance and the need for bigger models to gain benchmark points, add pressure. Any of these trends could slow growth enough to damage the business.
“now he has gone too much in on the compute. There's too much commitments... I think in the end, they're not gonna need all this compute.”at 33:44
Gary Marcus · New York University (professor emeritus)Bearish
Falling model-serving costs won't sustain Anthropic's prices because rivals are pursuing the same efficiencies.
Serving costs keep falling as efficiency gains accumulate, and Anthropic's best hope is eking out more of them. But competitors, including DeepSeek, are chasing the same efficiencies, so the prices labs can charge keep falling too. Prices have dropped by a factor of 100 to 1000, so Anthropic can sell for less but may still lose money per unit (his lemonade analogy: charging $1 for something costing $1.10). Because everyone has the same idea, sustaining prices and reaching profit is hard.
“But because the competitors are doing the same thing, it's gonna be hard to sustain the prices.”at 35:53
Gary Marcus · New York University (professor emeritus)Bullish
Tighter China chip controls and hawkish backlash help Anthropic by thinning out cheap Chinese rivals.
If the hawks win, Chinese open-weight models may not be served in the US, which Anthropic can be happy about. It is complicated, since prices might rise as a result. But fewer competitors reduce pressure on price wars, so the fallout from Jensen's lobbying is helpful to Anthropic. On that logic, it is the best argument Alex has made for Anthropic.
“in that way, I think Jensen's blowup is helpful to Anthropic.”at 40:03
Gary Marcus · New York University (professor emeritus)Bearish
Enterprise AI demand is shaky: studies show weak ROI and big customers are cutting back.
More than a dozen studies show a lack of ROI from enterprise AI, and some large customers are cutting back. Many companies adopt AI because they feel they must, often without the results they want, so much of current usage is experimentation. Models still hallucinate and have comprehension problems, so whether they work depends case by case. Growth has slowed (he saw a slowdown in Q3), and the key question is whether it can sustain a massive acceleration.
“I think one thing you didn't mention is more than a dozen studies have shown a lack of ROI.”at 42:05
Gary Marcus · New York University (professor emeritus)Bearish
Anthropic has little moat; its lead at the high end flips with OpenAI and rivals can copy products quickly.
So far Anthropic has little moat. Its only moat is its very high-end models, and even there the lead switches back and forth with OpenAI. Product advantages could help if they are distinctive, but Claude Code was a great product that was then undermined by a cybersecurity failure and a code leak, and OpenAI's Codex quickly followed. If Anthropic's products are better and rivals like OpenAI and Google can copy them within about three months, that is not a sustainable moat.
“a lot of the argument that I have made is that so far they don't have a lot of moat. The only moat they have is around these very high end models, and even there, the lead goes back and forth with OpenAI.”at 44:44
Gary Marcus · New York University (professor emeritus)Bearish
Anthropic's IPO valuation is absurd; no visible route to profitability justifies buying shares.
At the proposed price he sees no route to profitability for Anthropic, so he would not buy its shares. He expects a long, slow decline. He grants that the company has smart people and many reasons to like it, but argues the valuation is the real problem.
“And even still, you're not going to buy any Anthropic shares. I mean, not at this price. Like, this price, I just don't see the route to profitability.”at 55:55
OpenAI
Across the shows: 42 speakers in other episodes, 18 bullish and 36 bearish calls, past 30 days
Where everyone stands on OpenAI
Gary Marcus · New York University (professor emeritus)Bearish
OpenAI's revenue rose, but its costs rose faster, so it keeps losing money.
Marcus cites new numbers showing OpenAI made more revenue than before, but its costs went up faster, so it continues to lose money. He ties this to the broader cost pressures on frontier labs, including data center resistance and the need for larger models to get benchmark improvements. The point is that revenue growth alone does not fix economics when spending rises in step.
“There was also just new numbers from OpenAI, and they made more revenue than before, but their costs went up, and so they continue to lose money.”at 23:25
NVIDIA
Across the shows: 156 speakers in other episodes, 151 bullish and 127 bearish calls, past 30 days
Where everyone stands on NVIDIA
Gary Marcus · New York University (professor emeritus)Bearish
Jensen Huang's lobbying against export controls will backfire, with China hawks gaining power.
Accepting Alex's premise, Gary argues that China hawks will exploit Jensen Huang's lobbying against export controls. The fallout is that hawks gain more voice. When the Democratic Congress takes over in January, expect investigations into what happened and hawks asking why the US gives so much to China. He is uncertain about the ultimate outcome.
“I think it is right that China hawks are going to capitalize on Jensen's mistakes.”at 38:32
AI: overall outlook
Where everyone stands on AI: overall outlook
Gary Marcus · New York University (professor emeritus)
Adding symbolic reasoning to LLMs has improved models, though hallucinations persist; he is more optimistic.
Models are better now because they are neurosymbolic. Classical symbolic AI ideas such as loops and conditionals are in widespread use and have improved results, but LLMs are still at the core and wrapped in harnesses. This has not made them foolproof, since hallucinations and comprehension problems remain. He is more optimistic because there is less dogma and more openness to hybrid approaches, which he sees as a great idea.
“I think that the models are better now because they are neurosymbolic.”at 42:28
Gary Marcus · New York University (professor emeritus)
AI's strongest gains are limited to formally verifiable domains like math and code; most of the world remains out of reach.
Recent results are strongest in math because an LLM can generate proofs that a formal symbolic system like Lean verifies. Code also works because unit tests can verify it, which is why Claude Code works. For open-ended problems such as planning an invasion of Taiwan, formal verification is impossible, so this technique stops working. AI is stuck where answers can be formally verified, and most of the world is not like that.
“this technique that is super cool in these very verifiable domains doesn't really work for you anymore. And that is really where AI is currently stuck.”at 51:53
AI bubble & tech valuations
Where everyone stands on AI bubble & tech valuations
Gary Marcus · New York University (professor emeritus)
AI investor appetite is souring: OpenAI raised less money at a lower valuation, and SoftBank is pulling back.
OpenAI went out asking for around $30B at a valuation of about $1.4T, but raised only about $11B at $1.25T. That is less money at a lower price, which Marcus says had never happened to them before. He reads this as a sign of ambivalence in the market, and he adds that SoftBank, which had propped up OpenAI, is now saying maybe it should not put so much into AI, a sign the mood has changed. Credit ratings show similar ambivalence. The market is still listening but is no longer all in.
“So they got less money at a lower price. Less money at a lower price. I don't think that had ever happened to them before, ever. That's the first.”at 17:12
AI infrastructure & data centers
Where everyone stands on AI infrastructure & data centers
Gary Marcus · New York University (professor emeritus)
Believing data centers alone will deliver AI breakthroughs is a stupid, dogmatic idea.
The dogma that a pure LLM is enough was a bad idea, and the idea that all you need is data centers is an extension of that dogma. Dogma has held progress back more than anything else. Combining old and new AI is a better idea because it is less dogmatic. Scaling compute cannot replace the core discoveries still needed.
“The idea that you just need a pure LLM was a really stupid idea, and the idea that all you need is data centers is kind of an extension of that, and that's a stupid idea.”at 59:04
AI regulation & safety
Where everyone stands on AI regulation & safety
Gary Marcus · New York University (professor emeritus)
AI safety has made no real substantive progress, so AGI should not be pursued until it does.
He sees AI safety as the one area where progress has been essentially nil, a real pessimism that offsets his new optimism about the technology. He does not think the field should aim for AGI until it makes progress on safety.
“what we haven't made progress on at all in any real substantive way is AI safety. And I don't think we wanna get to AGI if we don't make progress on AI safety”at 59:29
Across the shows
What other episodes Chorus covered said about the same companies in the 30 days up to this one.
Topics
- AGI timeline
- AI investor sentiment
- AI lab profitability
- AI safety progress
- Anthropic IPO demand
- Chinese open-weight models in US market
- IPO valuation
- OpenAI profitability
- chip export controls to China
- competitive moat
- compute commitments
- enterprise AI ROI and retention
- enterprise customer defections
- inference pricing and margins
- limits of AI reasoning
- model commoditization
- neurosymbolic AI
- neurosymbolic architecture
- revenue growth
- scaling compute as path to AGI
- scaling cost vs capability gains
More from Big Technology Podcast
- How Humans Will Work Alongside AI Agents — With Mike Cannon-Brookes · November 17, 2026
- Storm Clouds Ahead For Apple?, Debating Amazon’s Bee Frontier, AI as UI · November 11, 2026
- Will the Democrats’ Big Win Stop the AI Buildout? — With Teddy Schleifer · November 5, 2026
- This AI Expert Argues We're Underestimating AI's Productivity Impact — With Ethan Mollick · October 28, 2026
- Emergency Pod: Anthropic's S-1 IPO Filing — With Dan Primack · October 26, 2026
Ask your AI about this episode
Chorus gives Claude, ChatGPT and other assistants every view from this episode and every other show Chorus follows. Ask who agrees, who changed their mind, and what else was said about the same companies.
Chorus reports what speakers said. It isn't investment advice. How Chorus works · About Chorus