Great Claude, Weaker Empire. Why Anthropic's IPO Could Be a Trap

The author argues that Anthropic's planned IPO at a valuation of roughly $965 billion is a risky bet — not because of Claude's quality, but because of the price and the expectations baked into it. Anthropic has a top-tier model, but unlike Microsoft/OpenAI and Google it owns no mass distribution layer (OS, office suite, browser, developer ecosystem), and it faces pressure from cheap and open-weight models along with the uncomfortable cost structure of agentic AI. The piece weighs the counterarguments too, and concludes that a near-trillion-dollar price tag no longer values great technology but an as-yet-unproven victory — plain sobering-up would be enough to knock tens of percent off it.

Anthropic is one of the most interesting AI companies of our time. Claude ranks among the best models on the market, and Claude Code has become, for many developers, a tool that genuinely changes the way they work with software. The company has a strong technical reputation, enterprise customers, a good brand, and the image of a safety-serious lab.

That is precisely why its planned IPO is so risky.

Not because it is a bad company. The problem lies in the price and in the expectations. At a valuation of around 965 billion dollars, it is no longer enough to say: we have a top-tier model. Nor is it enough to say: we have loyal developers. A valuation like that assumes the company will become not merely a premium supplier of models, but one of the principal platforms of the future AI economy.

And that has not yet been demonstrated.

It is fair to consider the more uncomfortable possibility as well: OpenAI may be going public not only because it needs capital, but also because it does not want to leave the public market to Anthropic. That does not mean there is evidence of a deliberate "hit job". Such a claim would be too strong. But the strategic effect can be the same even without demonstrable intent.

If both companies go to market in a similar time frame, investors will not be buying abstract "AI". They will be comparing specific bets.

On one side stands OpenAI, with the mass-market ChatGPT brand and deep ties to Microsoft. On the other, Anthropic with Claude — excellent technology and strong partners, but without its own operating system, office suite, search engine, mobile platform, or a developer ecosystem the size of GitHub.

That is a fundamental difference.

Anthropic has capital and cloud relationships with both Amazon and Google. That helps it survive the race for compute capacity. But a partnership is not the same as owning the distribution layer. Microsoft can push AI through Azure, GitHub, Copilot, VS Code, Office, Teams, and Windows. Google has Gemini, Cloud, Android, Chrome, Workspace, Search, and Antigravity.

Anthropic has, above all, Claude.

That is not nothing. But at a near-trillion-dollar price tag, it may not be enough.

The Betamax-and-VHS analogy is simplistic, but useful. The history of technology repeatedly shows that the market does not always crown the best solution. Often the winner is the solution with better distribution, a lower price, an easier purchase, a stronger ecosystem, and wider availability.

For a subset of tasks, Claude may be better than the competition. In writing, reasoning tasks, long context, and programming it holds a very strong position. But the customer does not ask only about model quality. They ask other questions too.

Where do I already have an account? Who will sell it to me under an existing enterprise contract? Who will handle compliance, support, auditing, security, and integration into my workflow? Who will connect it to my documents, repositories, cloud, office suite, and internal data? And why should I pay a premium precisely here?

This is Anthropic's weaker spot. Not the model. Distribution.

Google Antigravity illustrates the problem very clearly. A developer can use an agentic environment built by Google and, inside it, reach for competitors' models. At that moment, value shifts not merely toward whoever has the best model. It shifts toward whoever owns the working environment.

The model can become a swappable engine. The platform remains the car, the service shop, and the dealership.

Kimi, Qwen, Gemma, DeepSeek, and other cheap or open-weight models add to the same pressure. They are not free. Someone has to pay for hardware, electricity, operations, latency, security, and administration. But economically they do one important thing: they reduce the bargaining power of premium API providers.

A customer will not route every task to the most expensive frontier model if a cheaper alternative handles the routine well enough. Ordinary summarization, classification, text transformation, simple programming steps, or internal assistants can run on a cheaper tier. The expensive model comes into play only for tasks where the lower tier fails or where quality is genuinely critical.

That is a rational architecture. And at the same time, unwelcome news for a valuation that assumes the premium model will be the default choice for a large share of the market.

The price per token may fall. The number of tokens per task may rise. Total AI spending may also rise. These three things do not contradict one another. They simply do not automatically imply that the largest margin will be captured by a standalone model company with no distribution platform of its own.

A common argument against AI companies goes: they sell tokens below cost, and so they will eventually go bust. That is too crude.

A far more accurate formulation is different. Inference itself may be grossly profitable for a portion of the services. The problem does not necessarily lie in any individual query. It lies in the overall capital account of frontier AI.

Training a new generation of models, R&D, data centers, specialized hardware, cloud commitments, customer acquisition, safety teams, litigation, and infrastructure amortization create costs that have to be recouped somewhere. And if the model layer is simultaneously getting cheaper, commoditizing, and increasingly routed through someone else's platforms, the return becomes far less certain.

In other words: the problem is not that every token must be sold at a loss. The problem is that the company has to spend enormous sums to stay at the frontier, while the customer asks ever more often whether something cheaper would do.

That is a dangerous combination.

For Anthropic, Claude Code is simultaneously proof of strength and a source of risk.

As a product it is very strong. It brought the company directly to developers and showed that Claude is not just a chatbot but a tool capable of genuinely working with software. But agentic programming burns tokens differently from ordinary chat.

An agent reads files, plans, runs commands, fixes errors, reloads context, compares alternatives, generates intermediate results, and returns to earlier steps. The user sees a single change in the repository. In the meantime, the model may have processed an enormous volume of tokens.

That is why limits and billing are so sensitive. If a company lets heavy users consume too much compute on a flat rate, it risks its margin. If it tightens the limits, the user feels the subscription has turned into a taximeter. And once a customer starts thinking twice about every agent run, the product loses part of its magic.

With classic SaaS, the more a customer uses the product, the stronger the relationship and the better the economics tend to be. With agentic AI, that need not hold automatically. The most active user may also be the most expensive user.

Ahead of an IPO, the public market will want to see hard numbers: gross margin after inference, the cost of heavy users, retention, enterprise contracts, cloud commitments, and a real ability to convert usage into profit.

Litigation belongs among the risks as well. Here one must neither exaggerate nor sweep it under the rug. Copyright, training data, licensing, and liability for outputs will be a long-term problem for AI labs.

It would not be honest to write that Anthropic will certainly go under because of lawsuits. That is not supported today. But it is honest to say that legal uncertainty raises the discount an investor ought to demand on such a stock. Especially for a company whose valuation already assumes a near-perfect scenario.

If the market for training data shifts from contested fair use to large-scale licensing, costs will rise. If courts or regulators restrict the use of certain data sources, the pace of development may slow. And if large settlements recur, that is another line item in the capital account.

For an ordinary technology company this would be an uncomfortable risk. For a near-trillion-dollar valuation it is a fundamental question.

At a valuation of 100 to 200 billion dollars one could say: a top-tier company, fast growth, a high-quality model, a good enterprise position, high risk, but a comprehensible bet.

At 600 billion it is already stretched.

At 965 billion almost everything has to go right.

Claude would have to remain substantially better than cheaper alternatives over the long run. Enterprise customers would have to pay a premium and not drift toward a mix of models. Agentic workflows would have to become an enormous and profitable market. Compute costs would have to fall faster than prices. Legal and regulatory risks would have to leave training and distribution substantially untouched. And the company would have to create sufficient lock-in despite not owning a platform the size of Microsoft's or Google's.

That is a great many conditions at once.

What is more, the public market will not listen only to a story about model quality. It will ask more simply: what is the margin, what are the commitments, how fast are costs growing, how many customers actually pay large sums, what is retention, and why should a standalone model vendor be worth nearly as much as the future operating system of the AI economy.

For the analysis to be honest, the other side exists too.

Anthropic may be right. If agentic AI becomes the economy's new layer of work, the decisive number will not be the price of a token but the value of a solved task. If an agent saves a lawyer, an analyst, or a developer several hours of work, the token bill may not matter much.

Enterprise customers, moreover, often do not buy the cheapest solution. They buy support, auditability, security, stability, compliance, reputation, and predictability. In these segments Claude may hold a premium longer than the skeptics expect.

Nor do cheap models necessarily take everything. The routine, yes. But frontier reasoning, long context, sensitive decisions, reliable agentic control, and complex corporate processes may remain a premium market.

And finally: if inference costs fall faster than prices, margins may improve. In that case today's worries about compute may in time look overblown.

That is the best argument for the optimists. But even it defends a high-quality company rather than automatically a good stock at any price.

This is the main distinction that often gets lost in the debate.

Anthropic need not go bankrupt. Claude need not lose. Claude Code need not be a dead end. The company may keep growing, win large customers, and remain one of the world's most important AI labs.

And even so, an IPO at a valuation of around 965 billion dollars may be a bad buy.

A stock is not bought according to whether the product is likeable or technically excellent. It is bought according to the relationship between price and future earnings. And here the bar is set extremely high.

All it takes is slower growth, pressure on margins, a price war, stronger open-weight competition, an unpleasant court ruling, weaker monetization of agentic products, or a more successful OpenAI IPO that takes most of the retail "I want to buy AI" narrative. At such a price, no catastrophe is needed. Ordinary sobering up is enough.

A decline of tens of percent would then be no shock. It would be expectations returning to reality.

Anthropic has outstanding technology. That has to be said plainly.

But a nearly trillion-dollar valuation is no longer a valuation of outstanding technology. It is a valuation of victory. And victory is precisely what has not yet been demonstrated.

The company has strong partners but does not own a mass distribution layer. It has an excellent model but faces cheap alternatives. It has a fast-growing product, but agentic AI has an inconvenient cost structure. It has an enterprise story, but the public market will want numbers. It has a chance to become one of the decade's winners, but the price already looks as though that has happened.

The possibility that OpenAI's IPO will weaken Anthropic's investment story is real. Not as a proven conspiracy, but as a market effect. If investors have to choose between ChatGPT with Microsoft behind it, Gemini inside Google, and Claude as a premium specialist, Anthropic will be in a harder position than the quality of the model alone suggests.

Betamax wasn't bad either.

It just didn't win the market.

And that is the biggest warning for Anthropic ahead of its IPO.

Sources and notes:

Reuters: reporting on the Anthropic and OpenAI IPOs, Amazon's and Google's investments in Anthropic, and the litigation over training data.

Google Antigravity: official product and pricing information.

Microsoft: official information on the partnership with OpenAI.

Transparency of creation:

The concept, structure, and editorial line of the article are the work of the author, who prepared the content outline, established the key theses, and directed the entire creative process. Generative AI — ChatGPT by OpenAI — was used as a working tool for research, sourcing, checking wording, and developing the author's outline.

The author edited the outputs throughout, verified the key findings, and approved the final wording. No part of the text was published without human review. Factual data were verified against publicly available sources.

This procedure accords with the principle of transparency in the use of generative AI and moves toward the requirements of Article 50 of EU Regulation 2024/1689 (the AI Act) on labelling content generated or modified by AI. #poweredByAI