Everyone is watching the headline. Nobody is watching the collateral.
A crypto-native outlet ran a piece last week asserting that Anthropic is moving toward an IPO. Five information points. Four of them were the same sentence wearing different clothes โ IPO as stress test, IPO as valuation reset, IPO as confidence signal, IPO as benchmark. Zero valuation figure. Zero revenue number. Zero underwriter. Zero named source. Zero publication metadata.
I spent four months of my early career modeling fund velocity inside the 2017 ICO fog, tracking on-chain flows across five hundred token sales, and I learned to read documents like this one the way a structural engineer reads a crack. What that piece actually was: an SEO reflex. A second-order echo of a rumor that probably originated somewhere else entirely, laundered through a vertical that has no AI-native reporting capacity. Its existence is not evidence that a filing is in motion. It is evidence that the keyword volume is.
So let me set the epistemic floor explicitly, because everything downstream depends on it. The claim that Anthropic is preparing to list is unverified. What is verified โ and what the aggregator completely missed โ is something more interesting. The first frontier lab to cross into public markets would not primarily be raising capital. It would be installing a discount rate on the entire private AI capital stack. That is the event. The listing is just the plumbing.
Trace the liquidity ghosts and the shape of the thing becomes clear.
The context you actually need
Anthropic's public revenue arc, per reporting I can cross-reference, runs from roughly one hundred million dollars annualized in 2023, to the billion-dollar order of magnitude in 2024, to several billion annualized in 2025, with internal expectations revised upward more than once. That is one of the steepest enterprise-software ramps on record โ the Snowflake and Databricks comparisons are the right ones, and they are flattering.
The composition matters more than the slope. Revenue arrives through three channels: enterprise API calls, coding-agent subscriptions, and cloud resale via AWS Bedrock and Google Vertex. The coding segment is disproportionately large. Its customer list is populated heavily by AI coding startups โ the Cursor class of companies โ which means Anthropic is, functionally, selling shovels to gold miners.
Then there is the governance layer, which almost no coverage touches. Anthropic is a public benefit corporation with a Long-Term Benefit Trust that appoints a subset of board seats and holds decision rights that can, in extreme scenarios, override shareholder preference. In private markets that structure is a narrative asset. In public markets it is a valuation discount, because public investors price uncontrollable governance at a premium.
And there is MCP โ the Model Context Protocol โ open-sourced in late 2024 and subsequently adopted by competitors including OpenAI and Google. That is the piece everyone is underpricing.
The real mechanism: a public price for a private asset class
Here is the part the aggregator gestured at and could not articulate. Every major frontier lab today โ OpenAI, xAI, Mistral, SSI, Perplexity โ is valued by private investors pricing a narrative. There is no daily clearing price. There is no audited unit economics. There is a term sheet and a story.
The moment one AI lab trades on a public exchange, that changes structurally. Every private AI valuation in the sector gets forced toward mark-to-market against a disclosed comparable. Not perfectly, not immediately, but directionally and permanently. The mechanism is not sentiment. It is the LP memo that now has to justify a 40x revenue mark against a public comp trading at 15x with an S-1 attached.
This is what a stress test actually means, and the aggregator misread it. The author framed it as: will investors pay for AI? That is not the question. The question is: can an AI lab survive quarterly scrutiny? Private markets tolerate indefinite losses wrapped in a grand narrative. Public markets demand predictable quarterly delivery. The test is not demand intensity. The test is the size of the gap between narrative and financials.
The circularity nobody is auditing
Now the part that should worry you, and the part my cross-border settlement background makes impossible for me to unsee.
Anthropic's compute is substantially subsidized by strategic investors โ AWS, Google, Microsoft โ through equity-plus-compute arrangements. This is not a scandal. It is standard practice across the sector. But it produces a loop: investment funds compute commitments, compute commitments become cloud and silicon revenue, that revenue supports the investors' valuations, higher valuations justify larger commitments.
I have seen this movie. In 2017 I documented that sixty percent of initial ICO liquidity was recycled within four hours, manufacturing the appearance of organic demand. The mechanism differs โ that was on-chain velocity, this is off-balance-sheet commitment โ but the epistemic failure is identical. In both cases, the visible demand is partially the system bidding against itself.
When Anthropic files, this loop becomes a disclosure item. Round-tripping, related-party revenue, resale margin splits between a shareholder and a channel partner โ these are precisely the line items short-sellers and SEC staffers attack first. Being first through the door means being the template defendant.
Revenue quality, honestly assessed
Two structural facts compress the multiple.
First, customer concentration. If the top ten customers skew toward venture-funded coding tools whose own profitability is unproven, public markets will apply a channel-dependency discount immediately. That is not a hypothesis; it is how comparables get priced. And it creates a second-order risk that nobody models: if the coding-tool layer consolidates or thins out, Anthropic's revenue takes an indirect hit through no fault of its own.
Second โ and this is the one that determines the terminal valuation โ this is not a SaaS business. Software has near-zero marginal cost. Anthropic has a business model best described as compute resale plus an intelligence markup. Marginal cost scales roughly linearly with usage. Reporting suggests inference gross margins were negative in 2024 and improved into mid-range territory in 2025, with management guiding toward healthy margins by 2027-2028. Those are unaudited projections resting almost entirely on one assumption: that inference cost per token falls faster than price per token. If that slope is wrong, the margin gets squeezed from both ends simultaneously.
There is also the alignment tax. Constitutional AI and RLAIF are genuine original contributions, and their real value is engineering efficiency โ reduced human-labeling dependence, better training scalability โ not ethics. The market keeps misreading this as a values story. But the constraints do bite: in open-ended creative and jailbreak-sensitive contexts, conservative behavior costs usability. Enterprise coding tolerates it. Marketing and content workflows do not. That caps TAM quietly.
And there is no multimodal generation. In the single largest consumer-facing AI revenue pool, Anthropic is structurally absent. A filing priced as a general-purpose AI platform has a hard wound here.
MCP is the asset, and it is mispriced
Here is my strongest conviction on this name. Anthropic's durable strategic position is not model leadership โ that lead is contested quarterly, with Claude strong on software engineering benchmarks and long-context instruction following, structurally absent on generation, and alternating with rivals on competition math. Leadership is a renting position.
MCP is different. If the protocol consolidates as the default connective tissue between agents and tools, Anthropic occupies something closer to Android's position in mobile โ it captures rent from protocol placement rather than from winning each model generation. In interoperability markets, the standard-setter outlives the product leader. I have watched this exact dynamic play out in cross-chain infrastructure, where the winning bridge was rarely the best-engineered one and almost always the one that became the default assumption.
The open question โ and it is unresolved โ is governance. If MCP ends up controlled by a neutral foundation, it becomes a public good and the moat evaporates. If Anthropic retains control while competitors depend on it, that is a genuine structural asset. Nobody is pricing this because nobody can price an unsettled governance question.
The contrarian read: this is bearish for the AI token complex
Here is where I diverge from every take I have read this week.
The consensus treats an Anthropic listing as bullish for AI broadly. I think it is quietly bearish for the AI-crypto convergence trade, and here is why. Right now, dozens of AI-adjacent tokens trade at multiples of revenue against no disclosure regime at all. They are priced on narrative because narrative is the only available input.
Introduce a public benchmark with an S-1, audited financials, named customers, and a disclosed margin structure, and you hand allocators a comparison they did not previously have. Capital does not need to leave the sector. It just needs a denominator. When the first real disclosure arrives, the tokens trading at triple-digit revenue multiples against zero transparency stop being interesting and start being embarrassing. That repricing will not be announced. It will show up as a slow bleed in the AI-token index while the equity comp grinds higher.
The second contrarian point: listing erodes the neutral-supplier story. Anthropic's differentiation rests on being the vendor that does not compete with its own customers. A public float with AWS and Google as major holders invites exactly the related-party scrutiny that undermines that positioning. Customers who compete with those clouds will re-underwrite their supplier-lock risk.
And the third: the private narrative premium era ends. Whether the stock pops or sags, one outcome is fixed โ the AI capital stack now has a public reference price, and every private round from here gets benchmarked against it.
Positioning
Do not trade the headline. Trade the filing date, because that is the day the discount rate changes.
Watch three things in the S-1 and nowhere else: net revenue retention, top-ten customer concentration, and the resale margin split with strategic shareholders. Those three numbers will tell you more about the AI capital cycle than every valuation round reported this year. And watch MCP governance โ if that standard neutralizes, the moat story dies quietly and the multiple follows within two quarters.
The bubble breathes. But this time it is going to have to breathe in public, on a quarterly cadence, with an auditor in the room. That is not a stress test for Anthropic. It is a stress test for everyone who priced an AI asset while nobody was checking the plumbing.