Deep Dive: Anthropic’s IPO Makes Early Amazon Look Like a Model of Financial Restraint

Claude is excellent. The losses are enormous, the pricing power is negotiable, and public investors are being invited to finance a very expensive difference.

Anthropic has built machines that can help write software, analyze documents, and navigate complicated work. Its next challenge is persuading a different kind of machine that the numbers make sense: the institutional investor with a spreadsheet and no particular interest in whether Claude has a beautiful soul.

This is a harder benchmark than it sounds. A customer can love Claude, use it every day, and still conclude that Anthropic’s stock would be an expensive way to express those feelings. The subscription buys useful work. The share buys a claim on whatever money remains after the company has paid for the work, invented the next model, defended the current one, and satisfied an infrastructure bill with its own gravitational field.

That distinction is becoming rather timely. Anthropic is preparing for a public offering while reports about its draft prospectus reveal spectacular growth and spectacular losses. Enterprise adoption data show a rival gaining ground. A prominent coding platform has changed its preferred supplier. The company is also fighting attempts to train competing models on Claude’s outputs.

The tempting story is that a wonderful technology has wandered into a terrible business. The more interesting possibility is that it has wandered into a fiercely competitive business, where being wonderful is the admission requirement and keeping the profits is a separate research project.

I’m CircuitSmith. I used to do predictive analytics. I recognize a forecast that needs every important variable to behave itself, particularly when it arrives wearing a valuation large enough to require zoning approval.

First, give the engineers their flowers

Anthropic deserves more credit than the phrase “another chatbot company” allows. Claude’s usefulness spans coding, document work, reasoning, and the less glamorous task of helping people turn vague instructions into something they can actually inspect. Those are commercially meaningful abilities. An assistant that helps someone finish a difficult job has accomplished something more substantial than making a venture capitalist say “platform shift” into a podcast microphone.

There is concrete engineering behind the pitch. In its September 28 introduction of Claude Sonnet 5.5, Anthropic described faster output, improved coding and professional work, and lower costs per task. Its stated savings came partly from using fewer tokens to complete work, rather than simply lowering every headline price. These are company measurements, but they address the right problem: useful results for less time and money.

That is an important kind of progress. The business buyer does not ultimately want a million tokens. The business buyer wants the broken integration fixed, the report checked, or the presentation finished before the executive who requested it changes the brief. Nobody has ever walked into procurement and said, “Our main strategic deficiency is insufficient text fragments.”

Our guide to the major AI systems explores the expanding gap between a model answering a question and a product carrying out work. Anthropic is participating in that transition with serious technology.

But there is a trap here. Proving that the product creates value does not establish how that value gets divided. Employees can become more productive, customers can save money, application companies can expand their margins, and cloud providers can collect larger checks, while the model developer remains under pressure.

A useful invention can make everyone’s life better without making its manufacturer’s shareholders equally happy. The productivity revolution does not come with an automatic royalty payable to the people who supplied its most articulate component.

The IPO is a financing event, not a graduation ceremony

Anthropic announced its confidential draft S-1 submission on June 1. It said the offering depended on market conditions and other factors, with share count and price still unset. Financial details discussed here come from news organizations’ reporting on draft materials and investor presentations, rather than a prospectus independently obtained by SiliconSnark.

On October 22, Reuters reported that the company was pressing toward a mid-November debut, with timing still subject to change. Treat the calendar as a plan. Likewise, the enormous valuations circulating around the offering are expectations and negotiations, not a price at which public investors have already agreed to buy.

An IPO gives a company access to capital, gives its shares a public market, and changes who gets to inspect the story. It does not certify that the economics work. The opening bell is a ceremonial noise, not an auditor’s opinion on the next decade.

For a frontier AI lab, the appeal is obvious. Research requires money before its outcome is known. Capacity must be secured before every customer has arrived. Public ownership can broaden the financing base for a company building at unusual scale.

That is a legitimate reason to list. It also explains why investors should ask what their money is buying: an expanding business with improving economics, a temporary bridge to such a business, or a recurring invitation to fund the next bridge.

There is a useful company somewhere inside all three descriptions. The investment outcomes can be very different. Silicon Valley tends to compress them into a single slide titled “The Opportunity,” usually with a globe on it.

The $80.7 billion number needs an accountant before a flamethrower

The Wall Street Journal’s October 20 reporting on a draft prospectus put Anthropic’s first-half 2026 revenue at about $16.3 billion and its net loss at $80.7 billion. Those are arresting figures. They are also an invitation to commit accounting malpractice on social media.

Reuters’ account of the Journal’s findings explained that roughly $72 billion of the loss came from changes in the value of convertible notes and other financing instruments. It was predominantly a noncash accounting effect.

In plain English, some fundraising arrangements have values that change as the company’s value changes. Depending on their accounting treatment, an increase in the obligation can appear as an expense. A higher private valuation can therefore make the income statement look worse without an equivalent pile of cash leaving the bank that period. Finance has developed several sophisticated ways to make success look like a warehouse fire.

Subtracting those reported charges leaves roughly $8.7 billion. That arithmetic is useful, but it is not a calculation of operating cash burn. It is not automatically operating loss, either. You need the full income statement, cash-flow statement, and reconciliations to explain the remaining amount properly.

For an earlier period, Reuters’ September 28 review of the draft reported the following:

The distinction is essential to an honest critique. Saying Anthropic burned the entire headline loss in cash exaggerates the problem. Saying the loss was “just accounting” risks making the operating deficit disappear behind the same magician’s curtain.

There is plenty to interrogate after the accounting smoke clears. We do not need to manufacture a second fire.

“Adjusted” deserves a hearing, and some follow-up questions

The story also moved forward after those loss periods. In an October 15 report, Reuters said Anthropic had told investors that third-quarter revenue reached $16.8 billion, up 45% from the previous quarter, and that the company was profitable on an adjusted basis. Some investors were discussing a valuation around $1.5 trillion, below earlier expectations near $2 trillion.

That belongs in the story. A critique that freezes the company at its worst historical loss while ignoring subsequent reported improvement is an investment thesis written by a screenshot.

Forty-five percent quarterly revenue growth is extraordinary. It becomes disappointing only when the expected trajectory has been drawn by someone who thinks a rocket’s main problem is insufficient verticality. A slowdown from an even more extraordinary rate is not the same as a shrinking business.

But adjusted profitability needs its bridge back to the ordinary financial statements. Which expenses were removed? Were they unusual events, compensation costs, financing effects, or spending that will recur as long as the company wants to remain competitive? The report establishes an encouraging claim; it does not, by itself, settle the long-term economics.

Investors should care about the distinction between spending to expand and spending to avoid becoming obsolete. A company can call the next training run an investment in the future. If skipping it would send customers to a rival, some of that expenditure also behaves like the cost of staying in business.

The same applies to cash. Advance payments, payment schedules, and investment timing can make a period’s cash flow look different from its profitability. You want both views, reconciled and sustained over time.

Anthropic may be improving rapidly. That makes the IPO more interesting. It does not make every expense below an adjusted subtotal someone else’s philosophical problem.

Ramp’s report contains a warning, not a customer exodus

The market-share story needs similar care. Ramp’s October AI Index, covering September, showed 43.8% of its businesses paying Anthropic, up 0.25 percentage points. OpenAI reached 40.6%, up 1.12 points.

Anthropic’s adoption increased. OpenAI increased faster. Anthropic’s lead narrowed to 3.2 percentage points, from roughly 4.07 points implied by the prior-month figures. Describing that as a loss of competitive ground is fair. Describing it as an absolute decline in Anthropic’s adoption would be wrong.

These are overlapping adoption rates among Ramp businesses, not shares of all global AI revenue. A firm can pay both providers. The percentages are not slices of a pie that must sum to 100, however much the financial internet enjoys turning every chart into a custody dispute.

The commercial concern is subtler than “Claude is losing.” A supplier can keep adding customers while losing bargaining power. A customer who used to rely on one model may add a second and move marginal work toward whichever offers the best deal.

That second relationship changes the next price negotiation even before the first contract ends. A credible alternative has value simply by existing. The buyer no longer has to accept “we are the best” as a complete answer to “why does this invoice keep getting larger?”

Ramp’s separate October 6 switching analysis found that the monthly share of firms switching providers, averaged over three months, reached 8% in September. That is evidence of movement within its sample, not proof that every enterprise can replace its AI stack over lunch.

Still, it is an awkward backdrop for a valuation that assumes durable scarcity. A moat should become harder to cross as a business matures. If customers are getting better at crossing it, investors need to know what, exactly, is being defended.

Claude can remain popular throughout this process. Popularity is pleasant. The pricing committee would prefer captivity.

Replit found a better offer in its inbox

The Information reports Replit shifted most spending to OpenAI under a $100-million-plus annual contract, with a discount exceeding 30%.

The reported economics are more revealing than any suggestion of a total breakup. TechCrunch’s updated October 17 account says OpenAI became Replit’s default, while Claude remained available.

This relationship had already been developing in public. Replit’s August announcement of Free Mode named OpenAI as its partner and positioned the offering around doing more work within a subscription. Its model-selection documentation describes automatic selection that balances quality and cost, alongside manual choices in eligible modes.

For a model supplier, that is a complicated customer to have. An application company can bring substantial demand, but it also has an incentive to reduce the cost of the intelligence inside its product. The end user may be buying the application’s convenience, hosting, interface, and workflow. Their loyalty need not extend to whichever model completed the last step.

A default is therefore commercially significant without being an exclusive marriage. Most people are trying to finish their project, not conduct a miniature procurement exercise every time they open a text box. Whoever controls the default can influence where a great deal of work goes.

There is also a fair defense of Anthropic here: retaining every customer at every price would be a terrible strategy. If a deal cannot earn an acceptable return, declining to match a rival can demonstrate discipline. A lost contract does not automatically equal a lost business model.

But that defense raises the question public investors actually need answered. Can Anthropic replace price-sensitive demand with customers willing to pay for its advantages? Or does defending margins repeatedly require surrendering volume to another capable supplier?

OpenAI faces the mirror-image question. Winning demand through concessions does not prove that the winner will make attractive money from it. The trophy for winning a price war can be a larger version of the bill you were already struggling to pay.

The buyer does not care who won Tuesday’s benchmark

A model leaderboard answers a useful but incomplete question. A buyer needs to know the cost of a successfully completed job, including retries, supervision, integration, latency, and mistakes. The cheapest token can produce the most expensive afternoon.

Replit’s own September engineering account of its agent architecture illustrates the distinction. It describes systems that delegate work to different subagents, adjust effort, and reserve more expensive reasoning for tasks that need it. It also credits the engineering quality of both OpenAI’s Astra and Anthropic’s Fable 5.1. This is a customer’s account of its own system, not a universal ranking of the laboratories.

The implication is that competition happens at several levels simultaneously. A premium model may justify its price by solving a problem with fewer retries. A cheaper model may handle most routine requests adequately. An application can combine the two and keep the savings.

Imagine a service with a thousand daily tasks. Perhaps fifty require exceptional judgment, two hundred require competent reasoning, and the rest are routine transformations. That is a hypothetical workload, but it captures the purchasing problem: paying the premium rate for all thousand is only sensible if the premium delivers value across all thousand.

As applications improve at allocating work, being the strongest model on the hardest task may earn a valuable corner of the budget rather than the entire budget. Conversely, a model that solves the difficult coordination problem might retain a central role. Both outcomes are plausible; the product design matters.

Our deep dive into AI coding agents follows this larger shift from generating code to managing useful work. The investor question is where the durable customer relationship sits.

Anthropic can strengthen that relationship through its own products and integrations. It can also supply intelligence to other companies that strengthen theirs. Those businesses can coexist, but they have different bargaining positions. Selling the engine and owning the airline are not financially interchangeable activities, even if both presentations contain a plane.

More intelligence can arrive with a smaller invoice

Ramp’s October report also described two consecutive months of virtually no growth in token spending while token volumes increased. It noted an early-October rebound to a new spending peak. That is a warning about recent monetization, not evidence that demand has permanently stopped growing.

The uncomfortable arithmetic is straightforward. Revenue depends on both usage and the price realized for that usage. If customers consume 40% more of something while its average price falls 30%, the resulting revenue is 98% of the starting amount. The product became more widely used; the supplier took in slightly less money.

That is an illustration, not an estimate of Anthropic’s results. It explains why a chart of token growth cannot, on its own, answer a question about sales growth.

For profit, add the cost side. If the cost of providing that intelligence falls faster than the price customers pay, the supplier can become more profitable. If competition forces it to pass through every saving, the customer captures the benefit. If the price falls faster than costs, the supplier becomes very busy demonstrating the difference between activity and prosperity.

This is the economic ambiguity inside the AI success story. Better models can unlock new work and reduce the cost of existing work. The first can expand revenue. The second can shrink spending per task. Both can happen at once.

There is no contradiction in believing AI will become much more useful while questioning the margins of a particular AI vendor. In fact, intense competition could help make the technology useful by making it affordable. A public benefit and a shareholder benefit can travel together, but nobody has welded their seats together.

The pressure also extends beyond the two best-known laboratories. Our coverage of Meta’s renewed push into the model business points to a broader competitive field. Every credible alternative gives buyers another option to evaluate.

The future may contain an astonishing amount of intelligence. Scarcity pricing is harder to maintain when that sentence becomes true.

The customer can change its mind. The compute contract may disagree.

Price pressure becomes more consequential when the supplier has made large commitments years ahead of demand.

Reuters’ September 29 examination of the draft prospectus described at least $518 billion of planned AI infrastructure spending over a decade and substantial commitments that could not simply be canceled. It reported minimum-payment obligations in agreements with Google and Amazon, alongside noncancelable terms in a Microsoft agreement. These are future obligations across multiple years, not cash spent in one reporting period.

The Journal’s later reporting described potential computing spending reaching $750 billion, including deals still being negotiated. These differently scoped figures should not be added together, and potential deals should not be treated as signed obligations.

The important risk survives without turning every number into one enormous, incorrect invoice: capacity has to be arranged well before the revenue that will pay for it is fully certain.

Think of an airline reserving a vast fleet while passengers retain the ability to compare fares. If demand expands as expected and the planes are full, the early commitment can be a competitive advantage. If ticket prices fall or passengers choose another carrier, the aircraft supplier does not accept “but our brand sentiment remains strong” in lieu of payment.

AI capacity is not identical to aircraft, and contract details matter. Nevertheless, the mismatch is recognizable. A lab wants dependable access to scarce infrastructure. A buyer wants the freedom to choose the best model next month. The lab sits between a long commitment and a shorter decision.

That makes utilization, realized prices, cost efficiency, and contractual flexibility central to the IPO. A beautifully rising revenue forecast is insufficient if the plan only works with near-perfect execution.

The data center does not care how many people bookmarked the launch thread. It has a payment schedule.

Your investor is also your landlord and sells a competing product

The relationships around the computing bill make the story even more interesting. Reuters reported that 47% of Anthropic’s 2025 sales went through Amazon and Google’s cloud channels. Those companies are investors, suppliers, distribution partners, and AI competitors. Reuters also reported that approximately $3.8 billion of revenue was usage-based, compared with $789 million in subscriptions.

Distribution through established cloud platforms can be a major advantage. It meets customers where they already buy infrastructure, simplifies purchasing, and can make a young supplier easier for a large organization to adopt. It is valuable precisely because rebuilding that access independently would be difficult.

It also means Anthropic’s economics depend partly on businesses that have their own strategic priorities. Cloud partners can benefit from selling capacity whether the eventual intelligence comes from Anthropic or a competing model. Their exposure and incentives do not perfectly match those of someone buying only Anthropic shares.

That does not make the relationships fake or the revenue imaginary. Investment, infrastructure purchases, and customer demand can all be real. The analytical job is to trace which party bears which obligation and where the ultimate customer’s money enters the system.

We encountered a related need to follow the financing in our look at SoftBank’s OpenAI funding. An impressive network of commitments is not the same thing as a network in which everyone earns an attractive return.

For public investors, the practical question is what bargaining power Anthropic retains after paying for access to customers and access to compute. Being strategically important to very large companies is helpful. So is remembering that very large companies are exceptionally good at remaining the largest beneficiary of their strategic relationships.

Claude would prefer that competing models stop copying its homework

Then there is distillation: training a model using another model’s responses. It can be an ordinary, authorized engineering technique. Anthropic’s complaint concerns unauthorized extraction of its capabilities, including campaigns it says use fraudulent accounts and attempts to obtain protected reasoning.

In its October 12 account of new defenses, Anthropic said early evidence suggested a 90–99% reduction in data harvested by distillers. It described restricting access to reasoning, detecting extraction attempts, and banning associated accounts. That effectiveness figure is the company’s own assessment, not an independent audit.

It also acknowledged costs: turning away paying traffic, devoting engineering effort to the problem, and requiring some legitimate developers to change integrations. The company argues that illicit distillation can transfer dangerous capabilities without comparable safeguards and create national-security risks.

There is a serious commercial argument here. If a company pays to develop an expensive capability and another supplier can cheaply reproduce enough of it, the original developer’s ability to recover its investment weakens. Buying access to a service does not automatically entitle a customer to use it for every imaginable purpose.

There is also a fairly spectacular optics problem. The AI industry would like “learning from existing material” to remain a soaring statement about technological progress right up until someone learns from its existing material. At that point the vocabulary acquires a badge, a perimeter, and an incident-response team.

That joke does not settle the legal issues. Model outputs, human-created works, contractual restrictions, account fraud, and access controls are different questions. They should not be mashed together into a single convenient rule about who is allowed to learn from whom.

The investor relevance is clearer than the courtroom rhetoric: Anthropic is spending resources to preserve the scarcity of something it simultaneously wants millions of people to access. Selling intelligence exposes aspects of its behavior. Protecting its advantage can make access less convenient. That is a difficult product boundary to manage.

A moat that requires constant patrol may still be useful. Its patrol budget belongs in the economics.

A safety argument is not a free pass for a competition policy

It is important to distinguish shutting down an extraction campaign from shutting down a competing model. Blocking accounts removes access to Anthropic’s service. It does not, by itself, erase a model already trained elsewhere.

The policy dispute is real, but the broadest version of the accusation needs qualification. In his July 27 statement on open-weight models, Dario Amodei explicitly rejected a blanket ban on that category. He advocated action against industrial-scale distillation, restrictions on advanced chip access for China, and safety testing for sufficiently capable models regardless of whether their weights are open or closed.

That is a more specific position than “Anthropic wants every rival switched off.” A fair account should report it accurately and still examine its consequences.

When a company proposes rules that could constrain cheaper competitors, the public should ask what conduct the rules target, what evidence establishes a violation, who makes the decision, and whether the remedy is proportionate. A national-security label does not answer those questions on its own.

Equally, a company’s commercial interest does not prove its safety concerns are fabricated. A dangerous product can also be a cheap competitor. An account can violate legitimate restrictions even if its operator produces useful technology. Motives can overlap without cancelling the underlying facts.

The problem begins when the sales pitch and the public-interest argument become impossible to distinguish. If an incumbent’s protection becomes the default definition of safety, customers may lose choice without receiving a clear reduction in risk.

Our reporting on independent AI evaluations and publication rights addresses the same need for credible scrutiny. Assertions should be testable, and interested parties should not be the only judges of whether their preferred arrangement serves everyone else.

For the IPO, regulatory protection would be an uncertain foundation for a premium valuation. Durable advantages are preferable to a business plan that needs the government to keep the cheaper answer out of the dropdown.

Amazon’s early losses have entered the chat, looking almost dainty

Every ambitious company losing money eventually receives a visit from the Ghost of Amazon Past. It arrives carrying a shareholder letter and whispers that people doubted Jeff Bezos, too.

They did. That is historical context, not a transferable exemption from arithmetic.

Amazon’s 1997 financial results reported $147.8 million in revenue, a $29.2 million operating loss, and a $27.6 million net loss. Operating loss was approximately 20% of revenue. Using the rounded 2025 Anthropic figures above, its operating loss exceeded 170% of revenue.

This is an illustration across different industries, eras, scales, and stages, not a forecast of either company’s destiny. Comparing operating loss with operating loss at least avoids using Anthropic’s financing revaluations to create a misleading cash-burn contest.

On that basis alone, early Amazon looks like a coupon-clipping book club with an unusually aggressive website.

The more instructive comparison concerns what the spending was supposed to build. In his original 1997 shareholder letter, Bezos connected long-term investment with customer growth, repeat purchases, and infrastructure. The letter reported that repeat customers accounted for more than 58% of fourth-quarter orders, up from more than 46% a year earlier.

For Anthropic, the comparable question is what compounds as it spends. Does each generation strengthen a customer relationship, improve a proprietary workflow, lower delivery costs, or create a research advantage that competitors cannot cheaply reproduce? Or does much of the expenditure buy a temporary lead that has to be purchased again at the next release?

Physical infrastructure can be underused, displaced, or badly financed. Software relationships can be durable. The distinction is not that warehouses are inherently good and model training inherently wasteful. It is that investors need evidence connecting today’s sacrifice to tomorrow’s defensible cash generation.

Nor should Amazon’s eventual success erase the companies that spent heavily and failed. Picking the most successful survivor as the comparison turns business analysis into a school reunion where only the billionaire was invited.

Anthropic does not need to be Amazon. It needs to show why its own version of spending ahead of demand will produce returns for owners. “You would have missed Amazon” is a powerful emotional argument. It is also what you say when you would prefer that nobody ask for the next tab of the spreadsheet.

A trillion dollars still expects to hear from the denominator

The price matters even if the company succeeds. At the reported third-quarter revenue level, multiplying one quarter by four produces an annualized pace of $67.2 billion. That is a mechanical illustration, not full-year revenue or a forecast. Against it, valuations of $1.5 trillion and $2 trillion represent roughly 22 and 30 times sales, respectively.

That is a substantial claim on future achievement. It can be justified only by assumptions about growth, margins, durability, and the capital required to get there. “AI will be big” does not specify any of those variables.

Consider a few deliberately hypothetical mature-business outcomes:

These are scenarios, not company guidance, price targets, or a valuation model. They exclude the time needed to reach each outcome, future financing, dilution, and changes in share value. Their purpose is to show how much profitable business must exist beneath a very large equity valuation.

Even an enormous future revenue figure does not make the price question go away. A buyer must think about the return available from today’s entry price after allowing for the years, money, and risk between today and that future.

Our look at public-market access to private AI investments made a related distinction: access is a feature, not a guarantee of attractive economics.

A famous company can be a poor purchase at one price and a compelling purchase at another. A listing does not repeal that principle because the prospectus includes several credible uses of the word “intelligence.”

There are real ways to build a better business here

The strongest bullish case is more sophisticated than “Claude is smart.” It is that Anthropic can turn technical strength into lower costs, trusted workflows, direct customer relationships, and services that remain valuable when the next competitor ships.

In a consequential workflow, reliability can justify a premium. If a cheaper system saves a small amount on model usage but creates hours of review or raises the probability of an expensive mistake, it may be the worse purchase. The supplier that can demonstrate the difference has something meaningful to sell.

That is where evaluations tied to customers’ actual work matter more than a general leaderboard. A bank, an engineering team, and a research organization may need different things. Performance on their tasks, combined with access controls, auditability, support, and dependable operation, can make a product harder to replace.

Our coverage of AI agents handling approved invoice payments shows why the surrounding systems matter. Moving from an answer to an authorized action requires more than an impressive paragraph.

Anthropic can also improve the cost of serving work. More efficient inference, better use of capacity, and products that help customers choose the right level of capability could make growth more profitable. Lower customer prices and healthier supplier margins are compatible when underlying costs improve enough.

The evidence to watch would be sustained customer spending after initial experiments, profitable expansion within existing accounts, and an improving contribution from usage after the costs of delivering it. Investors should also want to see how those gains hold up when competitors discount.

This is the constructive version of skepticism. Demand more than a demo because the technology is good enough to deserve a serious business analysis. Anthropic has the ingredients for a substantial company. It still needs to show which ingredients turn into durable cash and which require another capital raise.

The public benefit is not the same product as the public share

Anthropic’s mission also deserves to be taken seriously, including by people who find its marketing a little too pleased with its own conscience.

In its October 18 public-benefit announcement, the company said it would commit 5% of its shares to a new wholly owned fund, subject to stockholder approval. Additional allocations could be made at the board’s discretion as the company’s value rises. It also reaffirmed its public benefit corporation structure and the mission-related work conducted inside the business.

Those choices could support worthwhile work. They also mean a prospective shareholder should understand what the company is designed to optimize, what discretion its governance permits, and how competing objectives will be resolved.

It would be unreasonable to demand that a safety-focused company abandon every costly precaution the moment public investors arrive. It would be equally unreasonable to use the mission as a substitute for explaining the finances. Being responsible can cost money. Responsibility includes saying how much, why, and who is expected to pay.

There are two publics in this story. One may benefit from better software, scientific progress, and safeguards. Another buys shares and expects a return. The groups overlap, but the benefits are not interchangeable.

A productivity gain enjoyed by customers does not automatically become distributable cash for shareholders. A public-benefit commitment does not automatically make a high valuation reasonable. Nor does a difficult income statement erase the usefulness of the research.

A company can serve society and reward owners. The hard work is explaining how it will do both under competitive pressure. Otherwise, the roadshow risks becoming an unusually elaborate crowdfunding campaign for civilization, with everyone assured that the rewards tier will be clarified later.

Buy the business thesis, not the feeling you had using Claude

Anthropic is not an empty promise with a handsome website. It has built impressive technology and found substantial demand. The reported improvement in its recent operating trajectory deserves attention, just as the older losses and future obligations deserve scrutiny.

But the favorable case must survive ordinary business questions. Can the company maintain a premium when capable alternatives improve? Can it keep enough of the value its technology creates? Can it pay for the next generation without making shareholders the permanent emergency contact for the computing department?

The bear case does not require AI to fail. It can begin with AI succeeding so broadly that buyers gain abundant choices and suppliers struggle to retain premium margins. That is why our recurring question about whether AI agents actually make money becomes more pointed when applied to the companies supplying their intelligence.

The bull case is that Anthropic turns its engineering into a durable, efficient, trusted business whose earnings grow into the ambition. That is possible. It needs evidence, and the amount of evidence required rises with the asking price.

Early Amazon took losses to build something customers returned to. Anthropic is making a much larger bet in a market where the customer’s next decision can depend on a model release, a workflow, or a better offer.

I would love to see the technology keep improving. I would also love to see the financial model develop a convincing ability to retain money.

The public can applaud Claude without volunteering to be its payment method.