Supply Chained: A $3 Trillion Company and One Number Nobody Has Published

Contents

1. What the episode is about
2. The leaked figures, and two things that don’t add up
3. 500 times sales: three old IPOs as his yardstick
4. The fibre bubble: $40bn for $16m of revenue
5. The numbers they each stake out, and those 10 percentage points
6. Two customers for a quarter of revenue
7. After the IPO, the pressure travels down the hardware chain
8. The part he most wants to read is the appendices
9. How does a company losing money get valued at $3 trillion?
10. Should I buy on the day it lists?
11. Further reading
12. One thing to take away

Great love brings great expense; hoarding much brings heavy loss. Know what is enough and you suffer no disgrace; know where to stop and you come to no harm — and so you last.

—— Laozi, Tao Te Ching, chapter 44 (Spring and Autumn period; translation mine)

In the 13 October 2026 episode of Supply Chained, Tim Culpan (who writes the newsletter Culpium) and Jon of Asianometry go through Reuters reporter Echo Wang’s story on the Anthropic IPO prospectus she obtained in the last week of September: $4bn of 2025 revenue, $12bn of operating expenses, an $8bn operating loss, and roughly $7bn spent on compute and infrastructure. The valuation being talked about is $3 trillion, which Tim converts to 500 times sales, and he says that even if 2026 revenue reaches $50bn and the multiple falls to 40 times, that is still too expensive for a company this capital-intensive. Both hosts end up stuck at the same gap: first-half 2026 revenue has not been published. Tim’s floor is $20bn for the half; Jon expects $35bn and thinks the market needs to see $55bn to avoid a sell-off.

What the episode is about

Supply Chained is hosted by Tim Culpan and Jon. Tim writes Culpium, a newsletter on the technology industry, and has covered Asian supply chains for years; Jon makes Asianometry, which covers semiconductors and industrial history. Both of them watch the same chain from upstream to downstream, so the angle here is not whether the models are good. It is where the money comes from, where it goes, and who has to absorb it.

The material is a document that has not been formally published. Echo Wang obtained Anthropic’s IPO prospectus and wrote a story with numbers in it. Tim says he would love to see more of them, but she put in what she could. The episode is two people turning those few figures over and over.

The leaked figures, and two things that don’t add up

Four numbers are on the table: $4bn of 2025 revenue, $12bn of operating expenses, an $8bn operating loss, and a $42bn net loss. Tim says revenue and operating expenses both came in lower than he expected, and that the $42bn can be set aside — it is a paper figure from revaluations, fundraising and M&A, the sort of thing you see for a few quarters while the cap table gets jostled ahead of an IPO. It looks alarming and tells you nothing about the business. The number he actually wants is one he does not have: cash flow from operations.

The other mismatch is compute. Anthropic spent around $7bn on compute and infrastructure in 2025, while the figure being quoted in the industry is $30bn to $40bn for a gigawatt of capacity. Tim’s guess is that $7bn may be an annualised, depreciated amount — spend $35bn, write it down over five years, call it $7bn a year — and then he says he does not know what that number is and looks forward to seeing it. Jon adds one line: that year was before Claude Code.

What I wanted to note in this stretch is how he handles not knowing: offer a possible reading, then stop at “I don’t know.” That matters more than any of the inferences, because every multiple that follows rests on how these figures were drawn up.

500 times sales: three old IPOs as his yardstick

With no earnings to look at, all you have is a sales multiple. Tim’s yardstick is three old IPOs: Amazon at 4 times, Google at 9 times, Facebook at 30 times — and he says Facebook was the expensive one of the three at the time. Against that, a $3 trillion valuation is 500 times sales, which he calls outlandish. Even if 2026 revenue reaches $50bn and the multiple drops to 40 times, he says 40 times is outlandish too.

His reason is capital intensity. Facebook and Google did not have to bury tens of billions of dollars a year in machine rooms, while Anthropic’s growth is stacked on compute, so the same sales multiple sits on a completely different cash requirement. Having said that, he adds a line of his own: I’m not the capital markets, people decide for themselves.

Worth recomputing: $3 trillion divided by the $4bn he quoted is 750 times, and divided by $50bn is 60 times. The 500 and the 40 he uses are more generous than his own figures imply — and he has already called both of them outlandish.

The fibre bubble: $40bn for $16m of revenue

Tim says there is one group of companies for which these multiples were normal: optical fibre companies at the peak of the telecom fibre bubble. Jon then pulls up the work he had done on it — companies worth $10bn on maybe $100m of revenue, and one acquisition of about $40bn for a company with $16m of annual revenue, most of it paid in stock. Here he stops and delivers the hottest stretch of the episode.

That bubble destroyed hundreds of thousands of jobs and hollowed out an entire industry. A lot of people got hosed, a lot of people got scammed. He asks, please, that we stop the revisionist talk that it was a good thing, that the overhang it left behind is what made Google possible. His words: “It’s like saying, oh, because this whole forest burned down, we have ash to grow a new small shrub.”

Tim agrees with the first half of the argument — the world did inherit a lot of cheap infrastructure — and then says there were a great many victims along the way. The comparison is worth keeping because it separates two things: the infrastructure survived, and the people who funded it did not.

The numbers they each stake out, and those 10 percentage points

Then they press each other. Tim’s expectation is close to $50bn for full-year 2026. He would accept less, on one condition: that the reason is a compute constraint — NIMBYs, permitting, data centres that could not get built, which only pushes the installation of capacity back. If it is not capacity but demand that is not there, or token prices that fell too far, he would be disappointed. His floor is $20bn for the first half of 2026; below that, he asks what the point was of the $520bn of capacity they are looking to commit.

Jon expects $35bn for the half and wants to see $55bn: below $50bn, there will be some very strong questions in the valley. He says this is not bullishness, this is his scepticism talking — expectations right now are propped up that high. He also brings up Dwarkesh’s piece, which has Anthropic doing $100bn this year and possibly growing tenfold again next year. Jon’s reaction: brother, that’s a trillion dollars of revenue, who is paying for that? Tim says it cannot happen, because he watches the upstream and the downstream, and TSMC, Foxconn and Wiwynn are not tenfolding their revenue — nor does Moore’s law let the same dollar buy ten times the compute.

In between, Tim dismisses ARR. He calls it the dumbest metric there is and says you want six months of data at a minimum: by ARR logic, I got a subscriber three minutes ago at $130 a year, extrapolate that out and Culpium is now worth $3m. Jon’s reply: at fibre bubble multiples you’re worth $150bn in my eyes.

Underneath the joke is the core of the episode: the thing to model is the long-term operating margin. Whether it settles at 20%, 30% or 40% three to five years out is unknowable now, and 10 percentage points decides whether this is a $1 trillion, $2 trillion or $5 trillion company. Only as more recent data arrives, quarter by quarter, does that curve start to show.

Two customers for a quarter of revenue

One more detail from Echo Wang’s story: two companies account for a quarter of Anthropic’s revenue. Jon’s first reaction is that this sounds like aggregators, the cloud resellers. Tim does not disagree, but says it still tells you something — a certain amount of concentration, a lot of coconuts going through a small pipe to a few companies.

Then he offers a back-of-the-napkin method: assume it is AWS and Google Cloud at 12.5% each, look at the incremental revenue increase those two posted over 2025, and back it out from there. He flags the flaw himself — incremental revenue does not strip out cannibalisation, because some of the money spent on AI compute would otherwise have been spent elsewhere. Imprecise, but it gets you into the right ballpark.

After the IPO, the pressure travels down the hardware chain

The most counterintuitive stretch is his argument that an IPO is bad for the compute industry. His chain is clear: once you have listed, you have to justify it; capital markets will accept losses, but quarter after quarter they want those losses to narrow, and as long as the curve is moving in the right direction they will keep pricing it high; the moment a quarter is fumbled, margins move the other way, or there are signs of no economies of scale or operating leverage, the punishment is severe.

At that point the only lever that saves a quarter is cutting spending, and the line to cut is capex — which is compute, which is GPUs, ASICs and massive servers. He runs through four apologies in a row: sorry Nvidia, we don’t need your chips; sorry TSMC, we don’t need your capacity; sorry Foxconn, we don’t need you to build these servers; sorry Bloom Energy, we don’t need your expensive but convenient energy systems.

The other side of the story is that Anthropic has landed on a business model where every extra gigawatt of compute turns straight into a cash cow — they bought printing presses, so they will want more and more compute. Tim says he does not think that will be the reality; the reality will be, all right, we need a little fiscal discipline now.

He also explains why pacing the frontier has a business case. Training is pure cost, pure cash burn, and the money only comes in when you inference it. Move capex from training into inference and serving customers, and the cost model looks a lot better. He stresses that he is pointing out that the case exists, which is not the same as advocating for it. As for whether pacing is an antitrust dodge, he says that if several labs agreed together to slow their training it would look a lot like collusion and your general counsel would warn you — and that Dario Amodei putting “in conjunction with government supervision” in a footnote was the wise way to write it, though it should not be the primary reason.

The part he most wants to read is the appendices

Closing out, Tim says the money raised in the IPO will not be enough to plug the hole in what the next few years of AI capex needs. After that comes debt, lease buybacks, private equity, and whatever structures the bankers come up with to get money in without diluting anyone’s equity. He says the smoke and the fire will be in the wording of those deals, and that if the whole thing ever collapses, that is where you will trace it — so in the prospectus, read the appendices and the off-balance-sheet funding arrangements.

Jon adds a friend’s suspicion: that Anthropic is pushing for a November listing because there is no private cash left to raise. The two also agree that whoever goes first has the advantage — the first IPO soaks up the cash of everyone who wants in, and the second one, six to twelve months later, runs into fund managers saying I’ve already put money in that, I don’t want this much exposure to trillion-dollar companies.

How does a company losing money get valued at $3 trillion?

That was the first question that surfaced for me, and it is where I get stuck every time this kind of story appears: the books show a loss, the valuation is astronomical, and the two look unrelated.

The episode gives a concrete way to take it apart. With no earnings, what the market hands you is a sales multiple, and behind every sales multiple sits an assumption nobody wrote down: the future operating margin. Tim lays that assumption out — 20%, 30% or 40%, and 10 percentage points is the difference between $1 trillion and $5 trillion. So the $3 trillion price tag is not a verdict on today; it is a vote on a margin that has not been revealed.

Which makes the order of reading clear: find out what the assumed number is, then ask when you will get to see it. For Anthropic that means reading first-half 2026 revenue alongside expenses — how much expenses grow as revenue grows. Tim calls it seeing what they are leveraging on. In my own experience, when a holding makes me anxious, it is usually because I cannot say which single number would settle the question.

Should I buy on the day it lists?

The more practical worry: if it does list, is buying at the open a mistake.

The episode gives no answer, but it gives two layers you can take apart yourself. The first is supply of money: Tim does not think the price will skyrocket after pricing, and the first IPO absorbs the cash available, which makes life hard for the second. That layer has nothing to do with how good the company is. The second is the timetable: once listed, the company has to deliver a loss curve that narrows quarter by quarter, so every report over the next two years is a chance to check your read. There is no need to settle the judgement on day one.

What I do is write down what I am waiting for. For this episode I would write three lines: first-half revenue, expense growth over the same period, and the off-balance-sheet funding terms in the appendices. Those three lines will not tell me to buy or not buy, but next time a headline lands, I can tell which line it touched — or whether it touched none of them.

Further reading

• Supply Chained, episode of 13 October 2026, hosted by Tim Culpan and Jon
• Echo Wang’s Reuters story on the Anthropic IPO prospectus (late September 2026), the original source of the figures
• Culpium, Tim Culpan’s newsletter, and Jon’s Asianometry channel, where both supply-chain views live
• Dario Amodei’s essay on pacing the frontier; the government-supervision point mentioned in the episode is in a footnote
• The SEC’s EDGAR, for reading the appendices and funding arrangements once the prospectus is formally published

One thing to take away

Every price tag hides a number that has not been revealed. $3 trillion is not a statement about what Anthropic is worth today; it is a statement that the future operating margin will land somewhere, and the gap between 20% and 40% is four trillion dollars. Knowing which number is hidden there beats arguing about whether the price tag is too high.

Here is something I have tried. Take an optimistic claim you heard recently — it does not have to involve investing. Your kid says they can support themselves freelancing, a friend says the new company turns a profit in six months, a colleague says the new process halves the headcount. Write two lines: the first says which single number has to reach what level for the claim to hold, and the second says when and where you will get to see that number. If the first line will not come, the claim is still just a mood; if the second will not come, I park the judgement until the day it becomes visible.

Raw Markdown

This article is an educational discussion of investment method. It is not advice to buy or sell any individual security, offers no target prices, and does not analyze any current holding. Investing carries risk; make your own decisions or consult a qualified professional.