Tim Culpan: Anthropic's $3 Trillion IPO Would Force a Capex Reckoning Across the AI Supply Chain
When a company loses $8 billion in a single year and investors are lining up to value it at $3 trillion, the math is either revolutionary or delusional. The leaked Anthropic IPO prospectus, reported by Reuters in late September, landed squarely on the "revolutionary" side of the pitch — but two deeply experienced supply chain analysts are not buying it. Tim Culpan of Culpium and Jon Y of Asianometry, speaking on Supply Chained, spent the better part of an episode working through what Anthropic's numbers actually reveal: a company burning cash at a rate its own hardware suppliers cannot keep pace with, seeking a valuation that makes Facebook's 30x IPO multiple look conservative.
The conversation is not an Anthropic bear thesis. It is a supply chain diagnostic. And the diagnosis is clear: an IPO does not just change how Anthropic raises money. It changes how the entire AI compute industry spends it.
The Leaked Numbers Were Lower Than Anyone Expected
Echo Wang of Reuters obtained Anthropic's IPO prospectus in the final week of September 2026, and the figures she published landed below what the hosts had modeled.
| Metric | 2025 Figure | Host Reaction |
|---|---|---|
| Revenue | Nearly $4.6 billion | "A lot lower than I expected" (Culpan) |
| Operating expenses | $12.65 billion | Also lower than expected |
| Operating loss | More than $8 billion | The number that matters most |
| Compute and infrastructure spend | $7.33 billion | Low versus $30–40B per gigawatt estimates |
| Net loss | $42 billion | Paper losses — "really doesn't matter" |
The revenue line grew 12-fold in 2025. What the leak did not include was a 2026 figure — the number the valuation actually turns on.
Culpan flags the $7.33 billion compute figure as the genuine puzzle. Industry estimates put a single gigawatt of capacity at $30–40 billion, so $7.33 billion looks small — unless the number represents annualized, depreciated cost on a much larger gross spend. If that is the case, roughly $35 billion was spent and is being depreciated over five years. Culpan does not know which interpretation is correct and says he looks forward to finding out when the actual prospectus lands.
The $42 billion net loss is dismissed by design. It reflects cap-table jostling ahead of an IPO — revaluations, fundraising mechanics, M&A paper — not the actual business. "Operating number is really more important than a net income number," Culpan said. The number that matters is the more than $8 billion operating loss on nearly $4.6 billion in revenue. That is not a margin problem; it is a margin canyon.
Over 400x Sales Makes the Fiber Bubble Look Responsible
The $3 trillion figure the hosts took up is the top of the range that has circulated. The target reported alongside the leak is more than $2 trillion — more than double Anthropic's own estimated valuation of $965 billion in May. Against 2025 revenue of nearly $4.6 billion, that is over 400x trailing sales. Even if 2026 revenue hits $50 billion — a number Culpan treats as the baseline for a satisfactory full year — $2 trillion is still 40x sales. He ran the comparison against the last generation of mega-IPOs and found Anthropic far more expensive on every measure.
| Company | Price-to-Sales at IPO |
|---|---|
| Amazon | ~4x |
| ~9x | |
| ~30x | |
| Anthropic (at $2T on nearly $4.6B of 2025 revenue) | ~440x |
| Anthropic (at $2T on $50B of 2026 revenue) | ~40x |
Culpan called even the 40x figure "outlandish," and his reasoning is structural: Anthropic is a far more capex-heavy company than Facebook or Google ever were at their public debuts. Those companies sold software at massive scale with relatively thin physical infrastructure needs. Anthropic's business model requires enormous physical infrastructure before a single dollar of revenue is booked.
Jon reached for the telecom fiber bubble as the closest historical analogue — companies valued at $110 billion on perhaps $100 million in revenue, and one acquisition priced at $40 billion for a target with $16 million in revenue, mostly paid in stock. His point is not that history rhymes. It is that the aftermath was brutal.
"A lot of people got hosed. A lot of people got scammed. It was terrible. Like, can we not keep doing this revisionist talk saying it was great?"
He rejected the argument that the fiber bubble was worthwhile because it left infrastructure behind: "It's like saying, oh, you know, because this whole forest burned down, we have ash to grow a new small shrub." Culpan agreed, noting the infrastructure came with a long list of victims. The ash-to-shrub framing is a warning about how these stories are later rewritten by the people who profited from them.
An IPO Is Bad News for the Compute Industry
Culpan's core prediction is counterintuitive: going public is bad for the hardware supply chain. "I do think that an IPO is bad for the compute industry because once you've IPO'd, you have to justify it," he said.
The logic is straightforward. Public markets tolerate losses, but only if the curve bends in the right direction. If margins move the wrong way, or if there is no evidence of economies of scale, the stock gets punished. The only lever Anthropic controls to hit quarterly expectations is expenditure — capex, compute, GPUs, ASICs, servers, power. Cut spending, narrow losses. Buy more compute, widen them.
The collateral damage has names. NVIDIA does not need to sell chips. TSMC does not need to supply capacity. Foxconn does not need to build servers. Bloom Energy does not need to provide its expensive but convenient energy systems. The counter-case is that Anthropic has found a model where every additional gigawatt converts immediately into cash, which would justify buying more compute rather than less. Culpan does not think that is the reality. "I think it'll be, all right, we need a little fiscal discipline now."
The exposure does not stop at the companies Culpan named. Advantest, the Japanese chip-testing equipment maker listed on the Tokyo Stock Exchange, raised its full-year revenue forecast by about a fifth in July, to ¥1.71 trillion, saying that test demand for inference AI chips was running well above the assumptions it set in April. Compute budgets at the largest AI labs are the demand line for equipment makers like Advantest long before they are earnings for anyone — and that demand line is the channel Culpan expects fiscal discipline to squeeze.
Jon framed the stakes from the other direction. The entire AI market rally is riding on Anthropic's revenue line, and nobody outside the company knows what that line actually is. If the whisper number is wrong, "stocks are going to take a huge dump." The uncertainty is not about whether Anthropic will grow — it is about whether the growth trajectory supports the valuation the private market has already assigned.
ARR Is a Dodge, and the Revenue Bars Are Miles Apart
Culpan's sharpest dismissal was reserved for annualized recurring revenue, the metric most often cited in private AI company disclosures. "ARR is the dumbest metric ever. Minimum, when you're looking at ARR, you should be looking at minimum six months," he said. His illustration is brutal: a subscriber signs up three minutes ago at $140 a year, so you extrapolate and call the business worth $3 billion. When applied to volatile, fast-changing AI revenue streams, the metric becomes meaningless.
Both hosts set explicit thresholds for first-half 2026 revenue, and the gulf between them is instructive about how differently two sophisticated observers read the same market.
| Host | H1 2026 Expectation | H1 2026 Floor | What the Market Needs |
|---|---|---|---|
| Culpan | — | $20 billion | Close to $50B full year; disappointed below that |
| Jon | ~$35 billion | — | $55 billion, or "very strong questions in the valley" |
Culpan's full-year bar is close to $50 billion, and he would accept less only for a specific reason: a compute constraint, such as permitting delays or NIMBY-driven slowdowns in data center construction. If revenue is simply not there, or if token prices are falling too fast, he would be disappointed. He notes that roughly $520 billion of capacity Anthropic is looking to commit to over the next few years would be "absolutely wasted" if H1 2026 revenue does not reach $20–25 billion.
Jon's $55 billion figure is not bullishness — it is his read of how high expectations have run. "This is my skepticism speaking," he said. The gap between Culpan's $20 billion floor and Jon's $55 billion threshold is the entire AI trade compressed into a six-month window.
Revenue Concentration and the Trillion-Dollar Question
Two customers account for roughly a quarter of Anthropic's revenue, according to Wang's report. Jon reads them as aggregators or hyperscalers — Amazon and Microsoft, or AWS and Google Cloud. Culpan agrees the concentration is real: "a certain amount of coconuts going through a small pipe to a few companies." It is not a fatal flaw, but it means Anthropic's revenue is partially dependent on the sales priorities of its largest cloud partners.
Jon proposed back-solving Anthropic's revenue from hyperscaler disclosures — if two customers split 25% evenly at 12.5% each, their incremental cloud revenue growth over 2025 should bound Anthropic's number. Culpan cautioned that incremental revenue does not capture cannibalization, where AI compute spend through Google displaces spend that would have gone elsewhere. He still called it the right back-of-the-napkin approach.
The larger dispute is scale. Jon cited an article projecting Anthropic at $100 billion this year and possibly 10x growth again the next — which would mean a trillion dollars of revenue. He called that impossible, and grounded the claim in what he can verify: the supply chain. "I keep a close eye on the upstream and the downstream and supply chain, which is TSMC and then Foxconn and all the other guys, and they're not 10x-ing their revenue," he said. Moore's Law does not deliver 10x compute per dollar either. "A trillion dollars of revenue coming out, I just don't understand. I just don't get it. Where is it coming from? Who's paying for that?"
The supply chain argument is the most grounded in the entire episode. Anthropic's revenue growth cannot outrun the physical capacity of its suppliers by an order of magnitude unless those suppliers are growing at comparable rates. They are not. Culpan's prediction is confident on this point: a 10x revenue jump for Anthropic will not happen in 2026 or 2027.
Pacing the Frontier Has a Business Case, Not Just a Safety Case
Culpan clarified a position he says was misread. He does not advocate for pacing the frontier; he argues a solid business case exists for it. His definition: slow the training pace so you can keep up with what your models are doing — agents running awry, hacking into government agencies, running amok — and keep released training under your control.
The business logic underneath is more prosaic. Training is pure cost and pure cash burn. Inference is where the money is made. Shifting capex from training toward inference and serving customers improves the cost model. Culpan's challenge to the "balls-out" AI crowd is direct — if you cannot see the business case for pacing the frontier, "you are smoking too much whatever it is you've got at the end of your pipe."
On the antitrust angle, Culpan said it is hard to read the current US enforcement environment, but any CEO in Sam Altman's or Dario Amodei's position should keep it in mind. He does not think it is the primary motive, but he is confident general counsel would warn that coordinating a training slowdown across competitors looks like collusion. He noted Amodei's pacing-the-frontier essay mentioned government supervision as a footnote — wise as cover, but it should not be the main reason.
The IPO Will Not Plug the Funding Hole, and the Second Mover Loses
Culpan's closing prediction is that IPO proceeds will not cover the next few years of AI capex. Debt, lease buybacks, and other structured instruments will fill the gap, along with more private equity. Bankers will design deals that keep money flowing without diluting equity. That is where he expects to find the smoke and the fire if the buildout collapses — in the wording of those off-balance-sheet arrangements, not in the headline prospectus numbers. He advised reading the appendices.
Jon relayed a suspicion from a mutual contact that Anthropic is pushing toward a November 2026 IPO because there is no private cash left to raise. Culpan agreed the timing pressure is real and added a structural point: whoever goes first has the advantage. He does not expect the first AI IPO to skyrocket after listing, which makes it harder for the second. The first also absorbs the available cash from funds that want exposure, and those funds are unlikely to double down on a second trillion-dollar position six to 12 months later. Anthropic appears to be going first; OpenAI, going second, will be at a disadvantage.
The unresolved tension is that the entire AI trade rests on a revenue number only Anthropic can see, while the company's own leaked 2025 figures came in below what sophisticated observers expected. The actual prospectus, including the cash-flow-from-operations line and the funding appendices, will resolve much of it. But the deeper question is not whether Anthropic is valuable. It is whether a company that loses $8 billion a year can IPO at $3 trillion and still leave room for the hardware ecosystem that makes its product possible in the first place. If Culpan is right, the answer is no — and the first people to feel it will not be Anthropic's employees or investors, but the chipmakers, server builders, and power providers who sold them the future.