Anthropic's IPO: Forget the $42 Billion. Watch the $518 Billion.
September 30, 2026 | AI, Technology, Investment
Mike Rawaan, Founder and Managing Director
The headlines regarding Anthropic's IPO prospectus leak focuses almost entirely on the $42 billion, which tells us almost nothing.
Reuters obtained a draft of Anthropic's S-1, and the coverage followed one script: enormous loss, enormous commitments, enormous valuation. About $34 billion of that net loss is a non-cash charge tied to convertible financing, which the company marks up as its valuation climbs. The operating loss, a little over $8 billion on $4.6 billion of revenue, is the number that measures the business. That is still a large loss, but it is a quarter of the size of the headline. decryptFortune
There's a second number that drew an even louder reaction. Anthropic expects to spend at least $518 billion over the next decade with six infrastructure partners, and about 80% of that total is non-cancelable or payable regardless of usage. Skeptics divide it by 2025 revenue, get roughly 113 times, and declare the matter closed. biggo
That division is an error. It pairs a decade of obligations with a denominator the company outgrew months ago. Preliminary Q2 2026 revenue topped $11.5 billion, more than double the company's revenue for all of 2025. Analysts who attack Anthropic with 2025 math criticize a company that no longer exists. Analysts who defend it by pointing at growth alone ignore the feature of this deal that matters most. Both camps skip the standard exercise, which is to define the proper benchmarks, and then run the scenarios. jfeed
So I will do both.
The benchmarks that fail
Three proxies dominate the commentary, and in my opinion, none of them holds up.
Net loss. It mixes a $34 billion accounting entry with real operating results.
Commitments divided by trailing revenue. It uses a stale denominator and treats a ten-year schedule as a one-year bill.
Commitments against year-end 2025 cash of $20.28 billion. The ratio is about 25 times, but the cash figure predates a $65 billion raise at a $965 billion valuation in May, and the payments arrive over years.
A better benchmark
1. Compute cost per revenue dollar. This is the best single measure of whether the model works. Anthropic spent 71 cents of compute on every revenue dollar in Q1 2026 and projected 56 cents for Q2. The direction matters more than the level. A falling ratio means each generation of capacity earns more than the last. letsdatascience
2. Gross margin. The range is wide. SemiAnalysis estimated that inference gross margin rose from 38% to more than 70%, while blended estimates for 2026 run from roughly 44% to 60% depending on how an analyst treats training cost. The gap between those two figures is the real debate, and the audited S-1 should settle it. rdworldonline
3. Operating margin cushion. Q2 adjusted operating profit of $559 million on $10.9 billion of revenue works out to about 5.1%, and that figure excludes stock-based compensation. A critique from late May also argued that the adjusted figure omits certain equity-funded compute commitments that a GAAP presentation would include. I would treat 5% as the ceiling of the cushion, not the floor. rdworldonlineklover
4. Commitment burden. I define it as annual commitment payments divided by revenue. It replaces the broken 113x ratio with a measure that moves with the business. It also shows where the risk sits, which the scenarios in the next section of this article make concrete.
5. Customer quality. Nearly a quarter of 2025 revenue came from two customers, and the company warned that many large clients have no long-term contracts. Revenue that a customer can cancel next quarter supports a weaker case for a ten-year payment schedule than revenue under multi-year contracts. Zero Hedge
6. Valuation against run rate. At $2 trillion and a run rate near $65 billion, the multiple is about 31 times revenue. Reuters Breakingviews put the same valuation at roughly 325 times operating profit, assuming a 5.1% margin on a $120 billion run rate. The market prices margin expansion, not current earnings. rdworldonline
The arithmetic of a thin cushion
Here is the calculation that I'm nearly 100% certain the analysts performed, but the coverage missed. Take the Q2 figures: compute at 56 cents, operating profit at roughly 5 cents, which leaves about 39 cents for everything else. If all of those costs hold fixed in the short run, a revenue shortfall of about 5% erases the operating profit entirely. A 10% shortfall puts the margin near negative 6%. The real number sits somewhere in between, because although some costs flex, the point stands. Fixed commitments plus a single-digit margin exposes the profit to small misses.
I think of an airline and its order book. An airline that orders 200 aircrafts is acting rationally if load factors hold at 85%. The same order ruins it if load factors slip to 70% while the delivery schedule keeps running. In Anthropic's case, computing plays the role of the aircraft, and utilization plays the role of the load factor. The order is sound or reckless depending on what demand does after the order is signed.
Four scenarios
Every scenario below rests on the same commitment base. The average year carries about $52 billion ($518 billion over ten years). The peak year likely runs higher because the schedule probably back-loads. I use $90 billion as an illustrative peak. That figure is my assumption, and the public S-1 will replace it with the real schedule.
Scenario | End-2028 run rate | Burden, average year ($52B) | Burden, peak year ($90B) | $2T valuation as revenue multiple |
Compounding | $250B+ | ~21% | ~36% | ~8x |
Normalizing | $150–200B | ~26–35% | ~45–60% | ~10–13x |
Plateau | ~$100B | ~52% | ~90% | ~20x |
Break | Below $65B | 80%+ | 138%+ | 31x+ |
Note on the model: The figures above simplify a schedule that has not been made public. I divided the $518 billion in total commitments evenly across ten years to produce an average annual figure of roughly $52 billion. Actual payments will almost certainly vary from year to year, likely rising as capacity comes online and contracts run for different lengths (the Google commitment, for example, ends in 2033, while Amazon's runs to 2036). The $90 billion peak-year figure is my own illustrative assumption, not a disclosed number. I built the scenarios this way because the year-by-year schedule isn't public.
Compounding. Demand keeps outrunning supply, pricing holds, and the compute ratio falls toward 40 cents. In this scenario, the commitments turn into a moat, because rivals cannot secure comparable capacity at those terms. Operating margins reach the 25%-plus range, and the equity looks cheap in hindsight. The signposts are a compute ratio below 50 cents by year-end and net revenue retention that stays high.
Normalizing. Growth decelerates sharply but stays strong, and revenue lands between $150 and $200 billion. In this scenario, the commitments would be covered with thin-to-comfortable room. Financing comes from leases and debt at fair prices, with modest dilution. Margins reach the low teens, and the stock delivers a return well below the IPO-day hype. Most of the business press would call this outcome a disappointment, but it's actually a good result for a company with these obligations.
Plateau. This is the scenario that skeptics should write about, but they don't. Revenue stalls near $100 billion, perhaps because enterprise budgets tighten, cheaper models take the low end, or price compression outruns usage growth. The obligations keep ramping, and the peak-year burden approaches the entire revenue base. The company renegotiates, counterparties extend terms to protect their own investment, and the equity dilutes. The valuation resets (hard!), because ~20X revenue doesn't hold for a business with negative margins and no growth. The signposts are token volume that grows faster than revenue, rising customer concentration, and underused capacity.
Break (Doomsday). Demand reverses across the sector. Revenue falls below today's run rate, and the financing channels that normally cushion a miss tighten up. Every major counterparty (Alphabet, Amazon, Microsoft, Broadcom, AMD) holds exposure to the same demand curve, so they restructure together instead of rescuing one another. The fiber buildout of 1999–2001 offers the precedent. Demand for bandwidth grew enormously, but carriers overbuilt against fixed financing, so the lenders and borrowers failed at the same time. Today's counterparties are far stronger than those lenders, so a full repeat looks unlikely, but the circularity I see between supplier, investor, and customer would spread the damage across all parties.
My weights for each scenario
I put Compounding at 25%, Normalizing at 40%, Plateau at 25%, and Break at 10%. These are my judgment, not model output. They imply a 65% chance that the commitments work out and a 35% chance that they bite. That is a bet many investors would take, and it is also a broad set of downside cases for a $2 trillion price tag.
The call
The bears are right about the structure and wrong about the evidence. The $42 billion and the 113x ratio do not measure the risk. The real risk is narrower and more serious: $518 billion of largely fixed obligations converts any prolonged growth stall into a balance-sheet problem. A BIG balance sheet problem. The bulls are right about trajectory and unit economics, but growth alone does not retire fixed costs, at least not in this business.
The market also misses pricing the middle. Plateau does not require a collapse. It only requires that growth stops or slows down significantly before the payment schedule peaks. A company can grow 10x and still land there.
What will settle the ultimate question?
The public S-1's year-by-year commitment schedule
Audited gross margin with training cost treated explicitly
Customer concentration and contract duration
Whether Q3 prints a second profitable quarter or the reversal the company warned about
Cash after the IPO measured against the next three years of payments
Data notes: the S-1 is a leaked draft. Q2 revenue appears as both $10.9 billion (projected) and $11.5 billion (preliminary) in different sources. Anthropic books cloud-partner revenue on a gross basis, so its revenue does not compare directly with peers that report net. The peak-year assumption and the scenario weights are mine.
Comments