There are two revenue figures inside the Anthropic filing, and they cannot both be true.
The first regime reads clean: $4.8 billion in Q1, $11.5 billion in Q2, $16.8 billion in Q3. Extrapolate the curve and you land on an annualized run-rate north of $330 billion. The second regime is stated flatly for full-year 2025: $4.6 billion. The distance between the two is roughly sevenfold.
A sevenfold gap is not a rounding error. It is a fingerprint. I have read enough of these documents to know that when a company publishes two incompatible versions of its own top line, the discrepancy is not sloppiness — it is architecture. One number is built for the underwriters. The other is built for the auditors. Neither was built for you.
Every market has two ledgers. One is shown to the people who set the price. The other is the one the price eventually has to answer to.
I audited my first token distribution in 2017, scraping block explorers by hand for three weeks to map the concentration of a 25-million-token pre-sale. That exercise taught me the only rule that has ever survived contact with this industry: the ledger remembers what the analysts forget. So when a two-trillion-dollar IPO arrives with a sevenfold accounting ghost inside it, I do not read the press release. I read the footnotes.
Context: How to Read an IPO Like an On-Chain Audit
Anthropic is not a blockchain protocol. But the structure of its filing is one I recognize, because I have watched the same shape emerge in DeFi hundreds of times.
A pre-mine concentrates supply in a small set of wallets, then markets the float as if it were distributed. An IPO concentrates information in a small set of institutions, then markets the disclosure as if it were transparent. The mechanics rhyme. Goldman Sachs and Morgan Stanley lead the book, which tells you the deal is large and the demand is manufactured. What it does not tell you is whether the price is honest.
The core claim is a valuation above $2 trillion — a number that would place this company above the vast majority of listed corporations on earth. To justify it, the bankers need one specific story: a profitability inflection married to a trillion-scale infrastructure commitment. Both appear in the filing. Both are softer than they look.
Here is the methodology I applied. I stripped the narrative, isolated every numeric claim, and cross-checked them against each other for internal consistency — the same reconciliation I run on any protocol's reported TVL before I trust a single metric. The result is a company with a real business, a genuine franchise, and a disclosure package engineered to survive exactly one reading.
One more piece of structure matters. The filing references private data shared with select investors ahead of the public offering. That is standard practice, and it is also the largest information asymmetry in the deal. Institutions will price this asset against a disclosure set the public will never see in full. I have watched the same asymmetry in token sales where insiders saw the vesting schedule and retail saw the marketing site. The outcome was never in doubt.
Core: The Evidence Chain
Start with growth, because growth is the denominator of every valuation.
The filing shows Q1 to Q2 revenue doubling — roughly +140%. Then Q2 to Q3 decelerates to +46%. That is not a plateau; that is a cliff wearing a plateau's clothing. High-multiple software valuations are priced for sustained hypergrowth. The moment the sequential rate halves, the multiple has to compress, and no amount of underwriting can outrun arithmetic.
Then look at how the company defines profitability. Q3 is described as the second consecutive quarter of "adjusted profitability," with positive cash flow. Simultaneously, the filing discloses a full-year 2025 operating loss exceeding $8 billion. These two statements coexist only because of aggressive non-GAAP adjustments: stock-based compensation is excluded from "profit," and training cost is excluded from the claimed 80% gross margin. Excluding training cost from an AI company's margin is like excluding gas fees from a DeFi protocol's yield — you have removed the one variable that determines whether the machine actually works.
The most useful data in the entire document is the unit economics, because it is the hardest to dress up. In full-year 2025, every $1 of compute cost generated $0.63 of revenue — negative contribution. By Q3, that ratio improved to $1.70 of revenue per $1 of compute. That is a genuine turnaround in efficiency, and I will not pretend otherwise. But note the base: it started underwater. And note the structure: compute remains the single largest cost line, consuming roughly 59% of Q3 revenue.
Now the commitment that anchors the bull case: at least $518 billion of cloud and infrastructure spending promised over ten years — an average of roughly $51.8 billion per year, against 2025 compute spend of $7.33 billion. That is a planned sevenfold escalation in the cost base, underwritten by a revenue line that is simultaneously decelerating.
And the demand side is leaking. Meta and Microsoft — previously significant customers — are now building in-house models. That is the most damning competitive signal available, because it is not a rumor; it is revealed preference. When your largest clients conclude they can reproduce your capability at lower cost, your moat is not a moat. It is a toll booth that customers are quietly routing around.
Add the distillation line: 5% to 7% of revenue, now being cut. Anthropic was valuable enough to serve as a "teacher" model for competitors. Reducing that business signals either tightened licensing or a refusal to keep subsidizing rivals. Either way, it removes a revenue stream and confirms that the frontier is being compressed from below.
The forward expectations tell their own story. Over the summer, investors were told to expect a $100 billion annualized run-rate by year-end. The current estimate is closer to $74 billion. A 26% haircut before the deal even prices is not a forecast revision — it is a demand signal. And when the CEO and CFO choose to lead with margins rather than capability, read the emphasis, not the words. Executives foreground the metric that is holding. They go quiet on the metric that is slipping.
Finally, the pricing. The filing states plainly that pressure from low-cost Chinese labs forced both Anthropic and its closest rival to cut API prices — with the new generation priced roughly 75% below its predecessor. Anthropic is not the price leader here. It is the price follower. The company with pricing power does not drop its price by three-quarters under external pressure.
Contrarian: Correlation Is Not Causation
The consensus reading of the $518 billion commitment is bullish: a mega-cap cloud customer locking in a decade of demand, pulling the entire chip and data-center supply chain upward. I understand the trade. I do not trust the tell.
In my DeFi work, I learned to distinguish two things that look identical on a dashboard: organic flow and circular flow. A protocol that borrows against its own token to inflate TVL shows the same headline number as one with real users. The difference only surfaces when the incentive stops.
The $518 billion commitment has the same texture. AI labs frequently sign compute commitments with cloud providers who are simultaneously equity investors in the lab. The cash can travel in a circle: investor funds the lab, lab commits to spend with the investor's cloud, cloud books revenue, lab books capacity. This structure inflates both the revenue line and the capex line at once, which is precisely why the unit economics look cleaner than the business does. When the same dollar is counted as demand on one ledger and supply on another, the "efficiency improvement" from $0.63 to $1.70 is partly a measurement artifact, not purely an operating win.
There is a second blind spot. The market is treating the profitability inflection as durable. But the "adjusted" profit excludes stock comp and training cost — the two expenses most likely to grow, not shrink, as the compute commitment scales. This is the maturity-mismatch pattern I have flagged in every yield-bearing stablecoin product I have reviewed: short-dated, headline-friendly returns stacked on top of long-dated, illiquid obligations. It performs beautifully in the up-cycle. It reprices first when the cycle turns.
None of this requires bad faith. It requires only the ordinary incentives of a capital-intensive business racing toward a public listing. The same structure shows up in every yield aggregator I have audited: the headline APY is real until the subsidy stops, and the subsidy is always funded by someone who expects to be repaid. Here, the subsidy is compute capacity, and the repayment is a decade of revenue that has not yet arrived.
The filing even concedes the macro risk — rising rates could slow the plan. For a growth equity priced on a $2 trillion assumption, rate sensitivity is not a footnote. It is the load-bearing wall.
Takeaway: The Signal to Watch
Ignore the valuation headline. The number that matters is the reconciliation gap — whether the run-rate regime ($330 billion) and the actual regime ($4.6 billion) ever resolve into a single, auditable figure. Volatility is the noise; liquidity is the signal, and right now the signal is that the disclosure itself has two liquidity layers, one for institutions and one for everyone else.
Watch three things next: the net revenue retention figure once Meta and Microsoft complete their in-house migration; the take-or-pay structure of the $518 billion commitment — how much is irrevocable versus aspirational; and whether the $100 billion year-end run-rate expectation, already marked down toward $74 billion, gets marked down again before pricing.
Every structure has a fingerprint. I have just read this one. The question is whether the market reads it before, or after, the lock-up expires.