Everyone Reading Anthropic’s 335-Page S-1 Is Looking at the Wrong Numbers

Everyone Reading Anthropic’s 335-Page S-1 Is Looking at the Wrong Numbers

Wall Street analysts spent the last forty-eight hours hyperventilating over Anthropic’s gross margins and compute burn rate. They are reading a 335-page S-1 filing like it belongs to a traditional cloud software vendor from 2014.

It doesn't. And if you evaluate an artificial intelligence infrastructure giant using standard SaaS metrics, you are going to miss the biggest capital allocation trap of the decade.

The consensus breakdown of Anthropic’s filing focuses entirely on the wrong tension. Financial commentators love to frame this IPO as a morality play versus capitalism: a Public Benefit Corporation (PBC) trying to balance existential safety charters with hyper-growth revenue targets. They obsess over the sheer volume of capital flowing straight back out the door to Amazon Web Services and Google Cloud. They treat the dual-class voting structure and the Long-Term Benefit Trust as quaint governance quirks.

They are completely missing the structural reality buried in the footnotes.

Anthropic’s S-1 is not a prospectus for a software company. It is the financial blueprint of a regulated utility masquerading as a high-growth tech stock, wrapped in a governance shell designed to insulate management from the exact market forces that govern public equities.

The Compute-Revenue Illusion

Let’s dismantle the biggest myth floating around financial media right now. When pundits look at Anthropic’s top-line annual recurring revenue (ARR) growth, they apply a standard enterprise software multiple. They see massive enterprise adoption of Claude across Fortune 500 workflows and assume classic operating leverage is just around the corner.

Software has near-zero marginal cost of replication. Once you write the code, selling the ten-millionth license costs almost nothing.

Frontier AI models have punishing marginal costs. Every single token generated requires physical silicon, megawatts of electricity, and liquid cooling. When you read the cost of revenue lines in Anthropic’s S-1, you aren't looking at temporary growing pains. You are looking at the permanent physics of inference.

I’ve sat in boardrooms where enterprise CTOs blew eight-figure budgets trying to build proprietary wrappers around foundation models, convinced that inference costs would drop to zero within eighteen months. They didn't account for test-time compute. As models shift from instant text generation to multi-step reasoning chains, the compute required per query explodes. Even as the cost per floating-point operation drops due to better hardware, the total compute consumed per user task skyrockets.

Anthropic’s gross margins are structurally capped compared to legacy giants like Salesforce or Adobe. Calling Anthropic a "SaaS company" is a category error. You wouldn't value a natural gas pipeline using Microsoft’s price-to-earnings ratio, yet institutional analysts are doing the exact equivalent because the product is delivered via an API.

The Hyperscaler Round-Tripping Trap

Look closely at the risk factors section regarding cloud infrastructure commitments.

When Amazon and Google poured billions into Anthropic, the financial press celebrated the validation. What the S-1 lays bare is the circular mechanics of those deals. A massive percentage of the venture capital raised by Anthropic never actually sat in a traditional bank account earning yield. It existed as cloud credits and multi-year compute commitments flowing directly back to the investors' own balance sheets.

In accounting terms, this creates an optical illusion of scale.

Imagine a scenario where a landlord gives a tenant ten million dollars to start a bakery, on the strict condition that the tenant must spend nine million dollars renting the landlord's kitchen equipment and buying the landlord's flour. The bakery looks like it has massive capital backing and high operational throughput, while the landlord books millions in guaranteed equipment revenue.

That is the fundamental dynamic between frontier labs and hyperscalers. Anthropic’s S-1 reveals just how deeply intertwined its cost structure is with its primary investors—who also happen to be its fiercest distribution competitors through AWS Bedrock and Google Cloud Vertex AI.

If Anthropic wants to train next-generation models, it has to buy compute from the very companies building competing internal models. That isn't a partnership. That is a hostage situation with clean accounting.

The Public Benefit Corporation Misdirection

Every mainstream breakdown of the S-1 treats Anthropic’s Public Benefit Corporation status and its Long-Term Benefit Trust as a heartwarming commitment to AI safety.

Read the actual corporate governance clauses again. Strip away the altruistic branding and look at the mechanics of control.

In a traditional Delaware C-Corp, the board of directors has a strict fiduciary duty to maximize shareholder value. If a hostile bidder offers a 40% premium for the company, the board generally has to take it seriously or face brutal litigation from institutional shareholders.

Under a Public Benefit Corporation structure, the board is legally obligated to balance the pecuniary interests of stockholders with the best interests of those materially affected by the corporation’s conduct and the specific public benefit stated in its charter.

In plain English: management can legally ignore short-term profit maximization whenever they want, and shareholders cannot sue them for it.

Combine that with the Long-Term Benefit Trust—an independent body with the power to elect and remove directors over time—and you have created the ultimate anti-activist poison pill.

Silicon Valley founders spent the last twenty years using dual-class super-voting shares (like Mark Zuckerberg at Meta or Larry Page and Sergey Brin at Alphabet) to maintain total control while taking public money. Wall Street finally started pushing back against dual-class structures, with major index providers threatening exclusion.

Anthropic found a much smarter mechanism. Instead of asking public investors to trust greedy founders with super-voting stock, they asked public investors to trust an unelected safety board with moral authority. The financial result is identical: common shareholders have zero real say in capital allocation, strategic pivots, or acquisition offers. If the Trust decides that pausing a lucrative product launch is necessary to satisfy its safety charter, the stock price can crater by 30% overnight, and institutional investors have zero legal recourse.

Answering the Questions Investors Are Actually Asking

When retail and institutional investors search for breakdowns of this filing, they usually ask the wrong questions because they are anchored to outdated tech playbooks. Let’s look at the actual realities behind the common queries.

Is Anthropic actually profitable?

No, and asking when it will hit GAAP profitability misses the point entirely. Frontier AI development is an arms race with no finish line. The moment a lab stops reinvesting every dollar of operating cash flow into the next massive training cluster, its current models become obsolete within six months. Expecting near-term GAAP net income from a frontier lab is like expecting a pharmaceutical company to stop funding R&D so its quarterly earnings look prettier. The burn isn't a bug; it's the entry fee to stay in the game.

Does Anthropic have a durable moat?

Not in the model weights. Model weights are a depreciating asset with a half-life measured in weeks. Open-weight models from Meta, Mistral, and DeepSeek routinely close the capability gap on standard benchmarks within months of a proprietary release. Anthropic’s real moat isn't intelligence superiority—it’s enterprise switching costs, compliance integrations, and developer muscle memory. Once a regulated financial institution builds its internal compliance workflows around Claude’s specific context window and tool-use behaviors, ripping it out to save 15% on token costs is a nightmare no VP of Engineering wants to manage.

How does Anthropic’s valuation compare to OpenAI?

Comparing private secondary market valuations to a public S-1 is apples to hand grenades. Private valuations reflect the fantasies of late-stage venture funds trying to mark up their portfolios before raising a new fund. Public S-1 pricing reflects the cold reality of institutional liquidity, quarterly earnings scrutiny, and discounted cash flow models. Any comparison between the two that ignores the liquidation preferences of private preferred stock is mathematically useless.

Why Enterprise Buyers Should Care More Than Stock Pickers

If you are a CTO, a VP of Engineering, or a founder building on top of Anthropic’s API, this S-1 is the most important vendor risk document you will read all year. Forget the stock ticker. Read the risk disclosures like a procurement officer.

When a private company goes public, its incentives undergo a violent, permanent mutation.

Up until today, Anthropic’s primary goal was capturing market share and proving frontier capability to justify massive private funding rounds. Subsidizing API calls to win developer mindshare was an acceptable loss.

Once those shares start trading on the open market, every single quarter is graded in public. When institutional investors start demanding margin expansion, Anthropic will have only three levers to pull:

  1. Aggressive price tiering: Expect the cheapest, fastest models to remain cheap, while the truly capable reasoning models get locked behind punitive enterprise pricing tiers.
  2. Quantization and silent downgrades: Public tech companies are notorious for quietly optimizing backend infrastructure to save compute costs, often resulting in subtle degradations in output quality that only power users notice.
  3. Enterprise lock-in features: Watch for Anthropic to aggressively push proprietary workflow tools, memory layers, and orchestration frameworks designed to make your application state completely non-portable.

I have watched dozens of startups build their entire core value proposition on top of a single foundational API, only to watch their unit economics implode the moment the underlying vendor went public and tightened the screws on margins.

If your product relies on a single prompt sent to Claude, you do not have a software business. You have a marketing wrapper around someone else’s balance sheet.

The Structural Flaw in the AI Bull Case

Let’s step back and look at the macro argument driving this entire IPO. The bull case for Anthropic rests on the assumption that foundation models will capture the lion's share of the economic value created by artificial intelligence.

History suggests the exact opposite happens during major infrastructure buildouts.

During the telecom boom of the late 1990s, companies like WorldCom and Global Crossing raised astronomical sums of capital to lay millions of miles of fiber-optic cable. Analysts used glowing metrics about internet traffic doubling every hundred days to justify insane valuations. The infrastructure builders were right about the technology—internet traffic did explode—and completely wrong about who would capture the profits. The fiber-optic companies went bankrupt, and the value accrued to the application layer companies (like Google, Amazon, and Netflix) that bought up the distressed bandwidth for pennies on the dollar.

Look at the capital expenditure numbers buried in Anthropic’s S-1. Look at the multi-billion-dollar training runs that require dedicated power substations.

There is a massive, unacknowledged risk that foundation models are the fiber-optic cables of the 2020s. Training a frontier model costs hundreds of millions of dollars, yet the output—intelligence delivered over an API—is rapidly commoditizing. When three or four well-funded competitors can produce models that score within two percentage points of each other on coding, reasoning, and summarization, pricing power evaporates.

Anthropic’s S-1 tries to combat this reality by highlighting its safety reputation and enterprise reliability. Those are real differentiators. A major bank will gladly pay a 20% premium to a vendor that guarantees zero data retention and strict indemnification against copyright lawsuits.

However, a 20% pricing premium does not offset a 500% increase in training compute requirements for the next generation of models.

How to Actually Read the S-1

If you want to understand whether Anthropic will survive the next five years as an independent public company, ignore the glossy executive summary at the front of the filing. Skip the charts showing exponential token growth. Go straight to three specific metrics:

  • Net Dollar Retention (NDR) among cohorts spending over $100,000 annually: If large enterprises are expanding their spend organically after the initial pilot phase, the workflow lock-in thesis is working. If NDR is stalling, it means companies are treating Claude as an experimental budget item rather than core infrastructure.
  • The ratio of R&D expense to Gross Profit: Watch whether R&D is scaling faster than gross profit dollars. If the cost of training the next model grows faster than the gross profit generated by the current model, the business is running on a treadmill set to an accelerating incline.
  • Customer concentration: Look closely at how much of Anthropic’s revenue comes directly from AWS and Google Cloud resellers versus direct enterprise contracts. Reseller revenue carries lower margins and puts a middleman between Anthropic and the end customer.

The downside to taking this skeptical view is obvious: if Anthropic achieves a genuine capability breakthrough—an architectural leap that leaves competitors stuck in the mud—all traditional financial modeling goes out the window. In a winner-take-most intelligence scenario, paying any multiple for the winning lab looks cheap in hindsight.

Betting on a miraculous architectural breakthrough, however, is venture capital, not public equity investing.

The Verdict Nobody Wants to Hear

Wall Street wants a clean story. They want to believe Anthropic is the disciplined, safety-conscious adult in the room that will quietly print high-margin enterprise software profits while the rest of the AI sector burns cash on consumer gimmicks.

The 335 pages of Anthropic’s S-1 tell a much more uncomfortable truth. They reveal a brilliant research organization trapped inside a brutal capital-intensive business model, dependent on its direct competitors for compute, and shielded from shareholder accountability by a governance structure designed for a non-profit.

Stop reading the filing looking for the next Salesforce. Start reading it like you're underwriting a nuclear power plant—because the capital requirements, regulatory risks, and existential stakes are exactly the same.

AY

Aaliyah Young

With a passion for uncovering the truth, Aaliyah Young has spent years reporting on complex issues across business, technology, and global affairs.