Why Gary Marcus Won't Buy the Anthropic IPO
A $1.4 trillion listing, $900 billion in liabilities, and growth down from 147% to 47%. Marcus's bear case — and the bull case he concedes.
Guide7 min read
Anthropic is about to go public at up to $1.4 trillion. It has never made a profit — with one heavily subsidized exception — and carries nearly $900 billion in future liabilities. Gary Marcus, NYU professor emeritus and longtime scaling skeptic, wrote the post the market is arguing about: just say no to the IPO. Big Technology's Alex Kantrowitz had him on to walk through the bear case, rebuttal by rebuttal, and Marcus barely budged. This is that argument, with the numbers as stated on the show.
A good product can be a bad business
Marcus's distinction is the frame for everything else: "It's a good business in the sense that they make a good product. It's a bad business in the sense that it's really expensive to run." High revenue, narrow niche — "basically the highest end models and nothing else" — brutal competition, and costs that rise faster than prices.
His lemonade analogy: "You could have a good product and be a bad business. You can make the world's best lemonade with hand-squeezed lemons... but if you sell it for a quarter a cup instead of two bucks a cup, your lemonade might be good, but your business is bad." His tech analogy is blunter: "It's like Juicero."
The endpoint he predicts is not a soft landing. "I think there's a real chance they will go bankrupt," he says; more likely, "they will eventually be acquired, but they'll be acquired for a lot less." His sketch of the deal: someone offers $200 billion, Anthropic says no way, and "in the end, it'll be like $400 billion." Bad outcome for a $1.4 trillion listing, survivable outcome for the technology.
Token maxing is over, and so is the growth story it bought
A large part of the valuation rested on growth charts from the "token maxing" era, when companies told employees to "use as many tokens... and we won't even look at the numbers." Marcus's tweet from six months ago: "token maxing is stupid, prove me wrong." Nobody did. People were "running 40 agents overnight because they got paid more money because they were running more agents" — and then, he says, "all that stopped."
The numbers on the show: Q2 revenue up 147% over Q1; Q3 up "a mere 47%" over Q2. Marcus's name for extrapolating the hot streak is the "trillion pound baby" fallacy, after a joke tweet about a baby doubling its weight in a month and hitting a trillion pounds by college: "you can have exponential growth for a while, but usually it tails off." His verdict on the 147%-forever projections: "clearly wrong. And if they're clearly wrong, and they have $900 billion in liabilities — the math doesn't math."
The professor problem: $100 billion for four benchmark points
Even the growth that happened may not be repeatable at any price. "If you have to spend $1 billion to impress your users this year, $10 billion next year and $100 billion the year after... at some point, that is just not sustainable." His estimate of the treadmill: wanting "four points better on the benchmark" will suddenly cost "$100 billion... and then ultimately a trillion."
Kantrowitz's framing sharpens it: a year ago the choice was a sixth-grader versus a college freshman — frontier wins by default. Now it is professor versus associate professor. "Is there a hundred billion dollars worth of revenue for having the slightly smarter professor?" And the expensive model still hallucinates, still fails on problems no system can solve: "You're going to spend premium prices hoping that you can solve this super hard problem and then you don't." Customers, Marcus predicts, will increasingly "stick with the associate professor and hope for the best."
The bull case, as Marcus himself grades it
To his credit, Marcus concedes ground where the numbers force him to. On demand: Bloomberg reported the IPO "multiple times oversubscribed," possibly raising as much as $90.4 billion against the feared $64.4 billion shortfall. "Maybe I'm wrong," he says — while noting IPO-day enthusiasm is "almost like a meme stock thing" and the test is what happens after.
On efficiency: Anthropic's S-1 shows revenue per dollar of compute rising from 40 cents to $1.72. "That is their best hope," Marcus allows — then points out the catch: "because the competitors are doing the same thing, it's going to be hard to sustain the prices." On customer breadth, Kantrowitz cites the S-1: accounts spending $100K+ a year grew from under a thousand to nearly twelve thousand in twelve months — enterprise checks that require approvals, not experiments. (My read: the transcript renders the two counts with stray dollar signs; the meaning from context is customer counts, not dollars.) Marcus's answer: more than a dozen studies show no ROI, big customers are already cutting back, and "can you sustain that growth?" is still unanswered. Last year you sold lemonade for $3; now it costs you 40 cents to make, "but now you can only charge a dollar... you can charge a dollar, but it's costing you $1.10."
On products: Claude Code, Cowork, and the new Sheets put Anthropic in direct competition with Microsoft and Alphabet as a native product company, not just a model vendor. "If they can do that in a distinctive way, then that helps a lot." His condition, twice repeated: it has to actually work — spreadsheets that hallucinate will disgust users the way Copilot did — and rivals must fail to copy it in three months.
On geopolitics, he grants Kantrowitz the strongest point outright: "It is the best argument that you've made." Export controls and the Huang fallout could hobble Chinese open-weight models, shrinking the field that drives price wars. "In that way, I think Jensen's blow up is helpful to Anthropic."
Why he still says no
Three things keep him bearish. First, the moat is one slice wide — top-end models only — and open-weight models keep eating the bottom of it. Second, the local-models shift: as capable models run on users' own hardware, "people aren't going to need all these giant data centers," and any single growth pillar can break the rocket. "Any of them basically kills them."
Third, and most revealing about Marcus, is what he concedes about his own thesis. Claude Code, he insists, is "as neuro-symbolic as you can get" — conditionals, loops, tools grafted onto the model — which is exactly what he has demanded since graduate school. "I give Czerny a lot of credit for getting rid of the dogma." But the vindication stops at the technology: being right about neuro-symbolic AI, he argues, never implied the business works. "Nobody will admit that, but people who watch your show will know it."
His timeline puts the rest in perspective: real AGI, in his view, needs world-model and continuous-learning breakthroughs nobody has, so "I still don't see it happening this decade." And the industry, he notes, has made "no progress on AI safety in any real substantive way." My read: that is why the episode works as an IPO story rather than a technology story — both men agree the product is formidable; the disagreement is whether formidable is fundable.
Would he buy the shares? "I mean, not at this price. I just don't see the route to profitability. It feels to me like it's going to be a long, slow decline."
More podcast deep dives from Podwise.


