Hook
Observe the numbers first, because the numbers are the entire story, and the numbers are unverified.
A reported raise of up to one hundred billion dollars. A reported valuation approaching two trillion dollars. A reported anchor allocation of up to ten billion dollars from Nvidia. Not one of these figures appears in a registration statement. Not one has been confirmed by the issuing company. Not one has been confirmed by the chipmaker. The structure rests entirely on sources described as people familiar with the matter, and the same report that supplies the figures also supplies the sentence that should be pasted above every aggregation of it: the plans remain under discussion and may change.
I have audited this shape before. In 2017 I spent six weeks reverse-engineering the deployment scripts of an Ethereum infrastructure project that had raised on narrative alone. The vesting schedule was the tell. Three cliffs, all favoring the earliest wallets, all disclosed in a footnote that nobody read. The whitepaper said partnership. The code said extraction. I circulated a private note predicting a ninety percent probability of failure inside eighteen months, and the note was correct for reasons nobody wanted to hear at the time.
What follows is not a verdict on Anthropic. It is a teardown of the transaction as reported, using the same instruments I use on token emissions, liquidity depth, and reserve attestations. The ledger does not lie, but it forgets. When a single anonymous source carries a two-trillion-dollar number, the forgetting has already begun.
Context: What Is Actually Being Valued
Anthropic is a frontier model laboratory. Its product line is the Claude family of models. Its technical identity rests on three pillars: Constitutional AI as an alignment methodology, a Responsible Scaling Policy that commits the company to staged capability evaluations before deployment, and a mechanistic interpretability research program that attempts to open the internals of transformer networks rather than merely observe their outputs.
That is a serious program. It is also, from a capital-formation standpoint, a program built inside the standard transformer paradigm. Constitutional AI and RLAIF are training-methodology innovations. Interpretability is a research frontier with genuine academic standing. Neither constitutes an architectural break. Neither produces a product that cannot be replicated by a well-funded competitor with a different alignment philosophy. The differentiation is methodological and reputational, not paradigmatic.
This matters for the valuation question, because a two-trillion-dollar number has to be anchored somewhere. If it is anchored in a hardware moat, we can model it. If it is anchored in a network effect, we can model it. If it is anchored in a research methodology that any of four competitors can approximate within eighteen months, the anchor is a narrative, and narratives do not have depreciation schedules.
Anthropic's distribution runs through two hyperscalers. Amazon has committed roughly eight billion dollars cumulatively and hosts Claude through Bedrock. Google distributes through Vertex AI and holds a separate stake. Both relationships give Anthropic enterprise reach without requiring it to build a sales organization from zero. Both also give the hyperscalers visibility into Anthropic's demand curve, pricing, and customer list. That is a real commercial asset and a real strategic exposure simultaneously. The same structural tension exists in every cloud-dependent software business, but the magnitude here is different because the compute bill is not a variable cost that can be throttled. It is a physical commitment measured in megawatts.
There is a third layer that the reporting barely touches. Claude Code and adjacent agentic tools represent a move down the stack from model provision into workflow ownership. This is the correct strategic direction for any API business facing commoditization at the model layer. It is also a direction that requires capital, headcount, and iteration speed that a research-first organization has historically struggled to sustain. The company that writes the alignment paper is rarely the company that ships the best developer tool. Those are different muscles.
The essential information a reader needs before evaluating any of the numbers is therefore this: Anthropic is a well-funded, technically credible, enterprise-oriented frontier lab with two hyperscaler distribution channels, a strong safety brand, and no published financials. Everything after this sentence is inference, and I will label it as such.
Context: The Deal as Reported
The reported transaction has four components. An initial public offering seeking up to one hundred billion dollars. An implied valuation near two trillion dollars. Nvidia as an anchor investor committing up to ten billion dollars. And an explicit caveat, buried in the body but governing everything, that the discussions are preliminary and the terms may change.
Anchor investors occupy a specific role in IPO mechanics. They commit to purchasing a defined allocation before the roadshow, accept a lockup period, and in exchange receive pricing influence and the certainty of allocation that other institutional buyers do not get. The anchor commitment is a signal to the book. It is also a structural constraint, because a locked anchor allocation reduces free float, which reduces the supply of tradeable shares at listing, which mechanically amplifies both the initial pop and the subsequent volatility when the lockup expires.
The Nvidia allocation, if it materializes at ten billion dollars against a one-hundred-billion-dollar raise, would represent ten percent of the offering. That is a substantial anchor position. It is also, in absolute terms, a rounding error against Nvidia's own market capitalization, which means the chipmaker is not making a bet it cannot absorb. That asymmetry deserves more attention than it receives. A ten-billion-dollar position that is immaterial to the investor but transformative to the investee is not a joint venture between equals. It is a supplier acquiring optionality over a customer.
The benchmark set is instructive. The largest IPO in recorded history remains Saudi Aramco in 2019, which raised approximately twenty-nine point four billion dollars. Alibaba in 2014 raised approximately twenty-five billion. A one-hundred-billion-dollar raise would exceed the Aramco record by a factor of roughly three point four. No transaction of that scale has ever been attempted, which means there is no comparable precedent to calibrate pricing behavior against. When an event has no precedent, the models used to price it are extrapolations, and extrapolations from a sample of one are not models. They are opinions with decimal places.
Core: The Instrument Distinction Nobody Is Making
Start with the most basic question, the one that the coverage has skipped entirely. What is the thing being sold?
An IPO share of Anthropic is not a claim on model capability. It is not a claim on interpretability research. It is not a claim on the Responsible Scaling Policy or on any safety commitment the company has made. It is a residual claim on future cash flows, senior to nothing, junior to everything, issued by a corporation that has never published an income statement.
I made a version of this argument in 2024, when I worked with a quantitative firm to model the effect of spot Bitcoin and Ethereum ETF inflows on price stability. The finding that mattered was not the volatility estimate. It was that seventy percent of retail participants in our survey could not articulate the difference between holding an ETF share and holding the underlying asset. The wrapper had become indistinguishable from the contents. That confusion is not a retail deficiency. It is a structural feature of financial intermediation. Wrappers are designed to be legible. Contents are not.
The same conflation is forming here. A public listing will make Anthropic legible to a class of investor that cannot currently access it. That legibility will be interpreted as validation of the underlying technology. It will not be validation of the underlying technology. It will be validation of the corporation's ability to pass a disclosure regime and clear a book. Those are different tests. Passing one says nothing about passing the other.
This distinction has a practical consequence. Once Anthropic is public, its share price becomes a proxy for the entire frontier-lab sector. Passive flows, index inclusion, and thematic fund construction will attach to the ticker. Investors who want exposure to AI capability will buy the ticker. They will not be buying capability. They will be buying a levered, depreciating, capital-intensive operating business whose principal asset has a useful life measured in quarters and whose principal liability is a set of multi-year purchase obligations that do not appear on any current summary. The instrument being sold is an equity claim on a capital-consuming entity. The asset being narrated is a software capability. The two are not the same object, and the gap between them is where the entire valuation question lives.
Core: The Arithmetic Gap
Now run the numbers, with the caveat that the denominator is missing.
A two-trillion-dollar valuation is a price. Price divided by earnings is a multiple. Price divided by revenue is a sales multiple. Without a revenue figure, the valuation cannot be evaluated โ it can only be compared to other valuations, which is what the reporting did, and which is a category error dressed as analysis.
So let me construct the reverse calculation, which is the only honest approach when the denominator is withheld. What revenue would justify a two-trillion-dollar price under conventional public-market multiples?
At a twenty-five-times-revenue multiple โ generous for a company with no published gross margin, no published customer concentration, and a capital intensity profile closer to a semiconductor fab than to a software vendor โ the required annual revenue is eighty billion dollars. At forty times revenue, the required annual revenue is fifty billion dollars. At one hundred times revenue, the multiple assigned to the most optimistic growth names in the most permissive markets, the required annual revenue is twenty billion dollars.
Frontier model laboratories in the current cycle have disclosed, through various indirect channels and partner statements, annualized revenues in the single-digit billions. Exact figures are contested and I will not assert one. The point is the order of magnitude. The gap between the disclosed range and the requirement implied by a two-trillion-dollar price is somewhere between one and two orders of magnitude. That is not a stretch. That is a different category of company.
The valuation table below is not a forecast. It is an exposure of the information gap.
| Verification Item | Reported Figure | Cross-Check | Risk Flag | |---|---|---|---| | IPO raise | Up to $100B | Would exceed Aramco 2019 (~$29.4B) by ~3.4x | No precedent in market history | | Implied valuation | ~$2T | Would rank top five globally by market cap | Revenue undisclosed; multiple uncomputable | | Nvidia anchor | Up to $10B | 10% of raise; immaterial to Nvidia's own cap | Structurally asymmetric | | Revenue requirement | Not disclosed | $20Bโ$80B depending on multiple applied | Denominator withheld | | Gross margin | Not disclosed | Inference cost + licensing + compute | Unknown whether positive | | Customer concentration | Not disclosed | S-1 core disclosure item | Absent from reporting | | Use of proceeds | Not disclosed | Compute / R&D / secondary split unknown | Determines whether this is growth capital or exit liquidity |
Every row in that table that reads not disclosed is a row that a registration statement would be legally required to fill. The absence is not an oversight by the reporting. It is a property of the information environment. A transaction described entirely by anonymous sources and never by a filed document has no denominator, and a valuation without a denominator is not a valuation. It is a headline.
Core: The Circular Ledger
Here is where the crypto-analytical toolkit earns its keep, because the capital structure being described is structurally identical to a pattern I documented in DeFi in 2020 and in the ICO market in 2017.
Consider the loop. Nvidia invests capital into a model laboratory. The laboratory deploys that capital, along with capital from other sources, into the purchase of Nvidia accelerators. Nvidia recognizes revenue. Nvidia's equity appreciates. Nvidia's appreciated equity and cash position give it greater capacity to invest in the next model laboratory.
This is not a conspiracy. It is a closed capital circuit, and closed capital circuits have a specific failure mode. I documented it at YieldFarm Alpha in early 2020, where the advertised annual percentage yield was not generated by trading fees paid by users but by the emission of the protocol's own token. The yield was real in the sense that the ledger recorded it. The yield was not real in the sense that it could survive the withdrawal of the capital that produced it. I ran the pool balances in Python and published the slippage curve: a five percent withdrawal would have moved the price beyond the point where the yield compensated for the loss. The protocol collapsed later that year. Roughly two million dollars in reader losses were avoided by people who read the slippage table instead of the headline APY.
The AI circuit has the same topology. The capital flowing into the laboratory originates, in part, from the entity that sells the laboratory its inputs. The revenue recognized by the seller is funded by capital the seller supplied. The valuation of the seller rises on that revenue. The risen valuation funds the next round of supply.
There is a legitimate version of this. Vertical integration through strategic investment is a normal industrial pattern. Automakers invest in battery manufacturers and then buy their batteries. The pattern is not inherently fraudulent. What distinguishes integration from circularity is the answer to one question: does the end product generate cash from a customer outside the circuit?
For an automaker, the answer is yes, because someone buys the car. For a model laboratory, the question is whether enterprise customers pay enough, at sufficient margin, to service the capital consumed by the compute. If yes, the circuit is an efficiency. If no, the circuit is a revenue-recognition device that transfers risk from the supplier's income statement to the investee's equity holders. And when the investee is public, those equity holders are retail participants with no visibility into the circuit at all.
The important disclosure to watch for in any registration statement is not the investment. It is the purchase obligation. If the anchor investment is accompanied by a multi-year take-or-pay compute commitment, then the transaction is not an equity investment with a strategic rationale. It is a forward contract disguised as equity, and the counterparty risk sits with the public shareholder. The reporting does not mention a purchase commitment. The reporting does not mention the absence of one either. That silence is the single most important gap in the entire story.
Core: The Anchor Allocation Is a Vesting Cliff
The Chinese term for what I am about to describe does not exist in the equity vocabulary, but the crypto vocabulary has a precise one: the unlock cliff.
In 2017, auditing the tokenomics of a hyped Ethereum infrastructure project, I found three critical defects in the vesting schedule. Each defect allowed early investors and the founding team to release tokens ahead of the community allocation. The mechanisms were technically compliant and economically predatory. The schedule was published, so the information was available. The information was simply not legible to anyone who did not model it.
An IPO anchor allocation with a lockup is the same instrument with different legal wrapping. The anchor buys at a pre-agreed price. The shares are restricted for a defined period, typically six months. During that period, the tradeable float is artificially constrained. Price discovery occurs against a supply curve that does not reflect the true share count. At lockup expiry, the supply curve shifts discontinuously.
The magnitude of that shift here would be unusual. If Nvidia's ten billion dollars represents ten percent of the raise, and if the raise represents roughly five percent of the post-money valuation, then the anchor position is a meaningful fraction of the tradeable float. Layer on standard institutional lockups and the founding and employee allocations, and the scheduled unlock events become material supply shocks with known dates.
This is not speculation. It is arithmetic. The float is a function of share count and lockup schedule, both of which will be disclosed in the registration statement. What will not be disclosed is the behavioral response of a retail base that has been told the listing is a validation of artificial intelligence as a civilizational project. That base will buy the lockup expiry as a dip.
The precedent I would point to is not an IPO. It is the token generation event. The mechanics of price discovery around a restricted-supply event are the same whether the instrument is a share or a token. An anchor investor accepting a lockup is not a confidence signal. It is a liquidity-suppression mechanism with a scheduled expiration.
Core: This Is a Depreciation-Financing Event
Here is the insight I have not seen stated anywhere in the coverage, and it is the one I would lead with if I were writing the registration statement.
An IPO raising one hundred billion dollars is not a hundred billion dollars of growth capital. It is a hundred billion dollars matched against an asset base with a finite useful life. GPU clusters are depreciating assets. Data centers are depreciating assets. Power purchase agreements are fixed obligations with escalation clauses. The entire capital structure being financed is a stack of depreciating hardware and long-dated commitments.
Why does this matter? Because software businesses and hardware businesses have different terminal values, and the valuation being discussed is a software multiple applied to a hardware cost structure.
A software business with eighty percent gross margins and negligible capital intensity can sustain a sales multiple of twenty-five times. A business that must consume capital continuously to maintain its competitive position has a fundamentally different economics, and the market eventually applies a different multiple. Semiconductor manufacturers trade at lower multiples than software companies for exactly this reason. The capital intensity is not a phase. It is the business.
Frontier model training is capital intensive at the frontier and cost-declining behind it. The strategic problem is that the revenue-generating capability sits at the frontier, and the frontier moves. A cluster that trains today's flagship model is a depreciating asset the moment the next flagship is trained on newer silicon. The depreciation schedule and the competitive schedule are in a race, and the depreciation schedule has a fixed clock while the competitive schedule does not.
If the use-of-proceeds table in the eventual registration statement shows that the largest line item is a compute-related commitment rather than general corporate purposes, then the correct analytical frame is not venture financing at all. It is project finance for a depreciating asset base, and it should be valued the way project finance is valued: on cash flows against useful life, not on narrative against total addressable market.
That frame produces a different set of questions. What is the average useful life of the deployed compute? What fraction of revenue is consumed by depreciation? Is the depreciation schedule accelerated to reflect competitive obsolescence, or straight-line to flatter the reported margin? I have audited depreciation schedules in traditional equity research. The schedule is where management discretion lives, and the schedule is what a two-trillion-dollar price tag tolerates least.
Core: The Provenance of the Two-Trillion Number
In 2021 I traced the deployer wallet history of an NFT collection that claimed exclusive provenance rights for its holders. The claimed origin story was fabricated. Three addresses linked to the deployer had been previously flagged in money-laundering investigations. I published the ledger analysis step by step, and the floor price fell forty percent within a week.
The technique applies here, and the technique is simple. Ask where the number comes from.
A two-trillion-dollar valuation does not emerge from a discounted cash flow model with a withheld denominator. It emerges from one of four sources: a comparable-company analysis against Nvidia and Microsoft, a private-market markup from the most recent financing round, a negotiating anchor positioned above an acceptable settlement, or a misunderstanding propagated through aggregation.
The comparable-company route is the most superficially rigorous and the least applicable. Anthropic does not have Nvidia's installed base, pricing power, or manufacturing position. It does not have Microsoft's recurring revenue, distribution, or cash generation. Comparing a company to a firm it does not resemble produces a multiple, not a valuation.
The private-market markup route is more plausible. If the most recent private round valued Anthropic in the hundreds of billions, a two-trillion-dollar IPO target implies a multibillion-dollar step-up in a compressed timeframe. Step-ups of that magnitude in that timeframe occur in crypto every cycle, and they are called re-ratings, and they are followed by re-ratings in the other direction.
The negotiating-anchor route is the most probable. A reported valuation is a claim, not a price. The difference between a claim and a price is the difference between a rumor and a filing, and the entire reported transaction sits on the rumor side of that line by its own admission.
Core: The Infrastructure Ceiling
The capital circuit has a physical constraint that the coverage ignores entirely, and it is the constraint that ultimately determines whether any of the numbers are achievable.
Two trillion dollars of valuation implies a compute build-out at a scale that must be matched by electricity, land, cooling, and interconnection capacity. These are not financial quantities. They are physical quantities with permitting timelines measured in years. A GPU order can be placed in a quarter. A substation cannot be energized in a quarter. A transmission corridor cannot be permitted in a quarter.
The constraint is not the accelerators. It is the megawatts. Every frontier training run of the scale implied by these valuations requires power density that existing data center campuses were not designed to deliver. The industry has responded by signing power purchase agreements years in advance, by locating near stranded generation, and by exploring behind-the-meter arrangements that bypass standard interconnection queues. All of these are real. All of these are also bottlenecked by equipment lead times for transformers and turbines that are measured in years, not months.
Add a second constraint that is political rather than physical. Advanced accelerators are export-controlled. Any change in the control regime alters the supply picture for every buyer in the queue. A model laboratory that has committed to a multi-year training roadmap and financed that roadmap through a public offering has taken on a supply-chain risk that is not primarily commercial. It is geopolitical, and it does not have a hedge.
The infrastructure layer is where the two-trillion-dollar narrative meets the physical world, and the physical world does not negotiate. The ledger does not lie, but it forgets โ and what it forgets here is that a valuation is a claim on future output, while a megawatt is a claim on present steel.
Core: The Governance Tension Nobody Wants to Price
A public listing converts voluntary commitments into enforceable exposures.
Anthropic's Responsible Scaling Policy commits the company to staged capability evaluations and to deployment decisions gated by those evaluations. It is a governance document, voluntarily adopted, and it functions as a reputational asset. Post-listing, it functions as a constraint on the growth curve that public shareholders are buying.
This produces a specific tension that I have not seen quantified. If a capability threshold triggers an RSP-mandated pause, the revenue associated with that capability is deferred. Deferred revenue against a growth multiple is a re-rating event. The shareholders who bought the validation narrative did not buy a company that might voluntarily stop selling. They bought a company that grows. The governance document and the equity instrument are in structural conflict, and the conflict is not disclosed in any prospectus because it is a policy, not a liability.
The regulatory layer compounds it. Training runs above established compute thresholds trigger reporting obligations under frontier model frameworks in the United States and general-purpose AI provisions in the European Union. These obligations are not optional and they are not free. They require documentation, evaluation infrastructure, and legal review. A public company must disclose the cost of compliance. That disclosure will convert what is currently a reputational asset into a line item.
There is a third exposure that the reporting has not touched. Training data provenance is a live litigation risk across the industry. Privately, that risk is a reputational concern that can be managed through settlement and silence. Publicly, it becomes a quantified contingent liability requiring disclosure and, if material, reserve. The transition from reputation risk to balance sheet item is a one-way door.
The safety brand is real, and it will be priced by the market. The question is whether the market prices it as an asset or as a constraint. My read is that during the growth phase it is priced as an asset, and during the first slowdown it is priced as a constraint. The repricing between those two states is the largest unmodeled risk in the transaction.
Core: Where Anthropic Actually Sits
The competitive matrix below is built from public information about ownership, distribution, and positioning. It is not a capability assessment, because capability rankings change on a quarterly cadence and any ranking I publish today is stale on publication.
| Dimension | Anthropic | OpenAI | Google DeepMind | Meta | |---|---|---|---|---| | Model tier | Frontier cohort | Frontier leader | Frontier cohort | Open-weight leader | | Enterprise channel | Strong (Bedrock, Vertex) | Mixed, consumer-heavy | Moderate, cloud-native | Weak, open distribution | | Compute assurance | Moderate-strong (AWS + Nvidia) | Strong (Microsoft + Nvidia) | Strongest (in-house silicon) | Strong (owned clusters) | | Capital backing | Strong (Amazon, Google, reported Nvidia) | Strong (Microsoft, Nvidia) | Parent balance sheet | Parent balance sheet | | Safety brand | Strongest | Moderate | Moderate-strong | Weak | | Listing status | Reported IPO candidate | Private | Parent listed | Listed |
The matrix makes one thing clear. Anthropic's strongest differentiated asset is the safety brand, and the safety brand is the asset least convertible into revenue at scale. Its distribution is strong but mediated by two hyperscalers whose interests are not identical to its own. Its compute assurance is dependent on a supplier that is reportedly about to become a shareholder, which converts a commercial relationship into a governance relationship.
The pattern of a supplier taking equity across multiple downstream customers is worth naming. A supplier that invests in several competing laboratories is not expressing conviction in any one of them. It is hedging across them while locking the entire cohort into its product roadmap. That is a rational strategy for the supplier. For the investees, it means that shareholder support is non-exclusive, and non-exclusive support does not create competitive advantage. It creates a floor under everyone.
Contrarian: What the Bulls Have Right
I have spent this piece dismantling a number. Now I will argue the other side, because the other side is stronger than the skeptics admit, and the skeptics are making an error of the same kind they are accusing the bulls of making.
First, the two-trillion-dollar figure is probably not a valuation. It is an anchor in a negotiation. Anyone who reads it as a committed price and dismisses the entire event is committing the same sin as the person who reads it as a committed price and buys the event. Both are treating a leak as a term sheet. The professional response is to record the leak, note the direction of travel, and wait for the filing. Dismissing a transaction because its leaked valuation is implausible is not skepticism. It is pattern-matching with extra steps.
Second, the circular capital structure is not prima facie fraud, and treating it as fraud is intellectually lazy. Vertical integration through equity investment is a legitimate industrial pattern with a long history. What makes it legitimate is whether the integrated entity sells to customers outside the circuit. Anthropic does. Enterprise API revenue, cloud distribution, and developer tooling are external cash flows. The circuit amplifies them. It does not replace them. The bulls are right that the structure is not fictional.
The real question the bulls are not asking is the one I raised earlier: at what price does a legitimate integration become an illegitimate capital structure? Vertical integration financed at a rational multiple is a competitive advantage. Vertical integration financed at a multiple that implies demand which does not yet exist is a transfer of risk to whoever buys the equity. The structure is not the problem. The price is the problem.
Third, the safety positioning is a genuine commercial asset in regulated industries. Financial services, healthcare, and government procurement all require vendors with a compliance story that survives audit. Anthropic's governance framework is more legible to a procurement officer than any competitor's. That is a real moat in a real market, and it is a moat that does not depend on benchmark superiority. It depends on being the vendor that a risk committee can approve. I have watched enterprise software businesses win on exactly this basis for two decades. It is durable.
Fourth, the Nvidia strategy of investing across the cohort is defensible. A supplier with a dominant position in a fast-moving market faces a portfolio problem: it cannot know which downstream architecture wins, so it hedges. That is what a rational monopolist-adjacent supplier does. The bulls are right that this is not evidence of a bubble. It is evidence of a supplier optimizing for optionality. The bubble question is separate.
So the honest position is this. The reported transaction is real in structure and unverified in scale. The underlying company is real. The circularity is real and largely legitimate. The price is the unverified variable, and the price is the thing everyone is arguing about while pretending to argue about the technology.
Takeaway: What to Watch and What It Will Mean
The instrument to watch is the registration statement. Not the leak, not the aggregation, not the commentary. The filing.
Four line items will resolve the entire question. The revenue figure, which supplies the denominator and makes the multiple computable. The gross margin, which reveals whether the business generates cash or consumes it at the operating level. The use of proceeds, which reveals whether the raise funds growth or refinances depreciation. And the customer concentration, which reveals whether the revenue is diversified or dependent on the two shareholders who also control the distribution channel.
If the largest use-of-proceeds line is a compute commitment, the transaction is depreciation financing and should be valued as project finance against a useful life. If the largest line is general corporate purposes, the disclosure is being managed and the market is being asked to price an undisclosed plan.
One more thing is worth stating plainly, because it is the reason I wrote this piece instead of a shorter one. The question is not whether Anthropic is a serious company. It is. The question is whether a public market can price a rapidly depreciating asset base against a narrative that has no depreciation schedule, using an instrument whose holders were sold validation and received an operating business. That question has been asked before, in 2017 with tokens, in 2020 with yield, in 2022 with reserves. The answers were all the same, and the answers all arrived late.
The ledger does not lie, but it forgets. What it forgets, every time, is the denominator.
