Anthropic Could Be Worth $4 Trillion.

A software buyer can move from Claude to another artificial intelligence model without selling a factory, retraining a workforce or waiting for a lease to expire. In some systems, the change takes little more than an API call and a round of testing. That small technical fact sits underneath one of the largest valuation debates in business.

Anthropic was valued at $965 billion in a May funding round. Some investors now imagine an initial public offering worth between $1.5 trillion and $4 trillion, according to the Financial Times. The upper end would place a young AI company among the most valuable corporations ever built.

Can the Anthropic valuation make sense when customers can leave? Yes, but only if Anthropic turns model intelligence into something much harder to replace: a trusted place where companies do their work. Model quality may win an account. Workflows, distribution, security, data connections and dependable economics have to keep it.

Why the $4 trillion Anthropic valuation is even being discussed

The numbers feeding the optimism are extraordinary. Anthropic’s annualized revenue run rate topped $65 billion at the end of July 2026, up from about $9 billion at the end of 2025, Reuters reported. A run rate takes the latest pace of sales and extends it across a year. It is a snapshot rather than audited full-year revenue. Even so, the acceleration is difficult to ignore.

Anthropic has also projected roughly $190 billion to $200 billion in 2028 revenue, Reuters said. Investors willing to pay $4 trillion are buying that future in advance. At $65 billion of annualized revenue, the valuation would equal about 62 times sales. Even against $200 billion, it would still equal 20 times sales before accounting for dilution, execution risk or the cost of supplying all that computing power.

A multiple that rich carries a demanding story. Revenue must keep climbing. Gross margins must improve as the company serves more queries. Capital spending cannot consume every dollar that growth produces. Most of all, today’s users must remain valuable customers after several rounds of new models and lower prices.

The market is being asked to value Anthropic partly like a software platform and partly like a research lab. Those businesses deserve different multiples. Software platforms can enjoy recurring subscriptions, high margins and costly migrations. Frontier labs spend heavily to train the next system, then watch rivals release competing systems within months. A $4 trillion case needs the economics of the first business to emerge from the competitive habits of the second.

AI customers can switch one task at a time

Traditional enterprise software often buries its roots deep inside a company. A bank that replaces its core ledger or a manufacturer that replaces its planning system faces years of data migration, retraining and operational risk. The pain of leaving gives the vendor pricing power.

AI models sit in a looser position. A company may use Claude for coding, an OpenAI model for research, an open model for high-volume classification and a specialist system for images. Developers can put a routing layer between an application and several providers, then send each request to the model that offers the best mix of accuracy, speed and cost.

That makes churn granular. A customer does not have to cancel a company-wide contract on Friday and complete a migration on Monday. It can shift five per cent of traffic, compare the output, and keep moving. Winning one valuable job can be enough for a rival to gain a foothold.

The latest competition shows how quickly spending can move. After OpenAI released GPT-6 Astra in September, it accounted for about 13 per cent of enterprise AI spending tracked by Ramp, compared with roughly 8 per cent for Anthropic’s Claude Fable, according to Reuters. OpenAI also moved ahead of Anthropic in weekly spending on OpenRouter, a service that sends developer traffic across different models. Leadership can change during a single product cycle.

Open models add another source of pressure. A company with enough engineering talent can run one on its own infrastructure, tune it for a narrow task and avoid paying frontier-model prices for every request. The open alternative may trail on the hardest reasoning test and still win a large batch of routine work. That puts a ceiling on what proprietary providers can charge.

Fast revenue growth can hide fragile loyalty

Revenue growth and customer loyalty measure different things. A company can add new users faster than old users leave. It can also grow because existing customers are experimenting broadly, even if those customers have not decided which provider will handle their permanent workloads.

This distinction matters during a technology rush. Budgets swell while executives fear falling behind. Teams buy access to several models, run pilots and tolerate duplicated spending. Those conditions flatter every leading supplier. Renewal season is less forgiving. Finance departments ask which experiments reached production, which costs can be cut and whether a cheaper model can handle the same job.

Annualized revenue can magnify the confusion. It tells investors how fast the car is moving at this moment. It says less about the road ahead, the fuel bill or how many passengers will still be inside next year. For Anthropic, the strongest public evidence would include net revenue retention, gross margin by product, customer concentration and the share of usage tied to production systems rather than trials. Private-company headlines offer only pieces of that picture.

One large customer can also build its own substitute. Reuters reported that Meta, among Anthropic’s largest customers, has been looking to reduce its use of Anthropic models while developing more AI capacity internally. That is a special risk in this market: some of the best customers possess the money, chips and researchers needed to become competitors or suppliers to themselves.

The durable product may sit above the model

Anthropic does have ways to make leaving harder without trapping customers contractually. The most promising ones live above the raw model.

Consider a law firm that uses Claude inside document review. The model becomes more useful as the firm connects its matter files, writes evaluation tests, sets access rules, trains employees and builds an audit trail. Changing the underlying model remains possible. Reproducing the surrounding system takes time, and a small drop in reliability can carry legal or reputational costs.

This is where trust can become economic value. Enterprise buyers care about data handling, uptime, contractual protection and predictable behaviour. Safety work also matters when a model is allowed to write code, operate software or touch sensitive records. Anthropic’s focus on safety may slow some releases, yet it can support retention if customers come to view the company as the safer operator for consequential work.

There is also a difference between technical portability and organizational willingness to move. An engineering team may prove that another model works in a week. Security officers still have to approve it, lawyers have to review the terms, and business owners have to accept fresh failure modes. Once an AI system acts inside a company rather than merely answering questions, those approvals become part of the switching cost. Anthropic needs to reach that stage before rival models become close substitutes.

Product design can add another layer. Anthropic recently combined Claude’s chat and Cowork modes and introduced document, presentation and design tools in one interface. The aim is to remove the user’s need to choose a tool for each task, Reuters reported. An assistant that knows a company’s files, preferences and recurring processes is stickier than a model sold by the token.

Distribution helps too. Claude is available through major cloud platforms as well as Anthropic’s own products. Cloud relationships put the model inside procurement systems that large companies already trust. They also cut both ways. A cloud marketplace that makes Claude easy to buy can make a competing model easy to test.

The best version of Anthropic, from an investor’s perspective, becomes the control point for AI work. Customers may still use other models, but Anthropic would own the interface, the agent, the permissions or the evaluation layer that decides how work gets done. In that business, model choice becomes a feature. In the weaker version, Anthropic remains one excellent model among several and must keep winning the same customer again with every release.

Safety creates value and a difficult clock

Anthropic’s identity creates a second valuation tension. Chief executive Dario Amodei argued in September that frontier companies should slow the rate of capability improvement so safeguards and outside evaluation can catch up. He called for embedded independent evaluators with access to the way models are trained and tested.

Days later, Reuters reported that Anthropic was considering another model release as GPT-6 Astra gained enterprise traction. The company was evaluating the new system’s safety and weighing the timing, according to people familiar with the discussions. There is no logical contradiction in testing a model carefully and releasing it. The commercial clock still matters. A long delay can send developers, usage data and mindshare to a rival.

Anthropic has begun turning its safety claims into procedures that outsiders can inspect. On September 18, it announced an embedded evaluation partnership with Accenture; each company expects to invest at least $1 billion over five years. The evaluators are meant to examine models during development rather than only after release.

Investors must decide whether this discipline becomes a moat or a tax. It becomes a moat when regulated companies choose Anthropic because the safeguards reduce deployment risk. It becomes a tax when customers praise the principles and send their largest workloads to whichever rival ships first. Contracts and renewal rates will supply the evidence.

What would justify an historic AI valuation

A $4 trillion valuation requires more than a sequence of benchmark wins. Benchmarks expire as competitors train new systems and customers change what they test. Durable value would show up in a handful of business measures.

First, existing customers would spend more over time even as model prices fall. That would show that usage is expanding faster than unit costs decline. Second, Anthropic would keep healthy margins after paying cloud providers and buying compute. Revenue that carries enormous serving costs deserves a lower multiple than subscription software revenue.

Third, customer concentration would fall. Depending on a few technology giants gives those buyers bargaining power and leaves revenue exposed when one builds an internal model. Fourth, more usage would move into recurring production work. A model embedded in customer support, coding pipelines or document operations is harder to displace than a chatbot used for occasional questions.

Finally, Anthropic would need to prove that its product layer can survive changes in model leadership. Winning every benchmark matters less than remaining the place where customers manage context, tools, permissions and outcomes. That shift would change the source of its power from temporary intelligence gains to accumulated customer investment.

A public filing could make this argument measurable. Investors should look for revenue retained from the same customer cohort, the portion of sales supplied through Amazon and Google, remaining obligations under long contracts, and the cost of serving an additional dollar of demand. They should also separate API consumption from subscriptions and professional services. Those disclosures would reveal whether Anthropic is building recurring economics or riding a wave of costly experimentation.

What investors are really valuing

The upper estimate for Anthropic asks investors to underwrite several victories at once: continued demand growth, better margins, lower concentration, credible safety and a product that grows more useful as customers use it. Each is possible. None follows automatically from a $65 billion revenue run rate.

Customer mobility should lower the price investors pay for each dollar of sales because it reduces the certainty of future cash flows. Anthropic can earn a higher multiple when a valuable, trusted and embedded service raises the operational cost of departure. Contractual obstruction would be a fragile substitute.

The small API change from the opening remains the cleanest test. If a customer can redirect a task and lose almost nothing, the model is a commodity for that task. If the switch breaks a carefully built workflow, weakens an audit trail or forces a team to rebuild its working memory, Anthropic owns something more durable.

Four trillion dollars is a bet that enough customers will reach the second condition before model intelligence becomes cheap and interchangeable. Until then, the valuation belongs to a forecast. The customers still hold the evidence.

Sources

• Financial Times: Investors warn Anthropic could struggle to sustain revenues post-IPO
• Reuters: Anthropic considers releasing new AI model ahead of IPO
• Reuters: Anthropic revenue run rate tops $65 billion
• Reuters: Anthropic folds Claude tools into one interface
• Dario Amodei: We Must Pace the Frontier
• Anthropic: Partnering with Accenture on embedded evaluation