$2 Trillion Valuation Target, $42 Billion Loss: The Anthropic Bet
Notes by SDG ·
The valuation is a reported target, not a completed IPO price. The loss is the reported 2025 net loss, including roughly $34 billion in financing-related accounting charges—not $42 billion spent running the business. IPO reporting · Loss breakdown reported by Reuters
Buying Nvidia gives investors a way to bet on demand for AI computing. Anthropic’s IPO would ask them to take a more direct position on the other side of that transaction: whether a company buying that computing can turn it into a durable, profitable business.
Investors in cloud and chip companies would have a reason to pay attention, even if they never buy Anthropic shares.
Recent reporting points to a possible November debut, with some prospective investors discussing a valuation of $1.8 trillion to $2 trillion. Those remain reported expectations; the timing and final offer price could change. Stocktwits, October 1, citing Bloomberg
At that valuation, investors would be betting on years of growth, customers who keep paying, and profits large enough to support the cost of staying competitive. Cloud providers and other suppliers are making expansion plans that depend on much of the same demand.
For much of the AI boom, public investors have been able to buy exposure through Microsoft, Amazon, Alphabet and Meta, where AI sits alongside established businesses. Nvidia offers another route through the infrastructure those AI businesses need.
Their share prices already reflect views about AI. What they don’t give us is a clean reading of the economics of an independent frontier lab: the cost of developing leading models, serving customers, winning distribution and doing it all again when the next generation arrives.
AI could become essential to its customers while its suppliers keep fighting over price.
Reuters’ review of Anthropic’s confidential IPO draft put 2025 revenue at roughly $4.6 billion, about twelve times the previous year, alongside an operating loss above $8 billion. The much larger net loss of almost $42 billion included roughly $34 billion from revaluing financing instruments that could become shares. Reuters, September 28, via CNA
Calling the entire $42 billion “cash burned” would give readers the wrong picture. An operating loss and cash outflow are different measures, too.
The Financial Times reported $11.5 billion in revenue for the second quarter of 2026 and positive adjusted operating income. That measure excludes costs including stock-based compensation. Financial Times, September 13
One quarter reportedly generated more than twice the revenue of the entire previous year. A valuation argument built only on 2025 sales would miss how quickly the business is changing.
The unadjusted results and cash flow would show how much of that growth translates into profit and cash available for expansion. One quarter of adjusted profit offers only a partial answer when the company faces several years of spending ahead.
If customers keep finding valuable work for Claude, spending could expand from individual subscriptions into substantial business budgets. A tool that helps complete useful work reliably has a stronger claim on those budgets than something employees occasionally experiment with.
But usefulness alone doesn’t determine how much the supplier can charge.
Imagine a business using AI to handle thousands of routine tasks. As models improve, it may get more work done. As competing models get better, it may also negotiate lower prices or move simpler tasks to a cheaper option.
The customer gets more work done at a lower price. For Anthropic, the outcome depends on whether usage and efficiency gains outweigh pricing pressure—and whether its products become valuable enough that customers have a reason to stay.
Anthropic could make itself harder to replace by becoming part of work that customers run every day. It would need business owners who can explain what Claude does for their teams and why it deserves a place in next year’s budget.
Anthropic already offers Claude Code and Cowork alongside connectors and enterprise controls such as access permissions, audit logs and data-retention settings. A lasting advantage would depend on how well those products and controls work together in the customer’s business. Claude Enterprise
Consider a hypothetical retailer using Claude to help review supplier agreements. The useful system would need access to approved documents, the retailer’s purchasing rules, examples of acceptable terms and a clear path for sending exceptions to a human. Procurement and legal teams would have to agree on what it can do, test its performance and measure whether it actually speeds up the work without creating expensive mistakes.
If that works reliably, replacing Claude becomes a decision about an operating process. A competing model might be cheaper, but the retailer would need to check whether the existing instructions, tests, approval steps and user experience still work as expected. Some of that could transfer easily; some might need rebuilding or revalidation. A small improvement on a public benchmark would not, by itself, tell the business whether switching is worth the disruption.
Microsoft, I think, has a head start in companies already built around Microsoft 365. The documents, collaboration and access controls are already part of the working environment. Work IQ brings organizational context and permission-aware governance into AI workflows, while Copilot supports models from multiple providers. Microsoft’s Work IQ overview · Model choice in Copilot
That gives Microsoft a chance to retain the customer’s workflow even when the preferred model changes. The moat can sit around the model as much as inside it. Changing models would still require testing, but it need not mean replacing the broader platform.
Anthropic would need to earn a similarly dependable place in its customers’ operations. Customers would keep choosing Claude because the complete system works well for them, even when alternatives exist.
Business owners would need help identifying useful applications and deciding how their teams’ work should change. Training an engineer to connect systems is only part of the job. Someone in procurement has to decide which decisions can be assisted, which require approval, what counts as an error and who owns the outcome. Reusable industry workflows, implementation support and training for those business owners could make the difference between a successful demonstration and a renewed contract.
Customers would also expect predictable costs, dependable service, clear permissions, useful audit records and model upgrades that do not unexpectedly break a tested process. These requirements are not exclusive to Anthropic. Doing them consistently well could still make customers more comfortable expanding their use of Claude.
Anthropic introduced MCP as an open standard for connecting AI systems to tools and data. Openness can make adoption easier, but it can also make integrations reusable across providers. The company’s information and processes belong to the customer; connecting Claude to them does not automatically turn them into an Anthropic-owned advantage. Anthropic’s MCP announcement
Anthropic would have to keep earning its position through performance on the customer’s actual work, product quality and the cost of getting a reliable result. Repeat spending from customers who can switch would provide better evidence of a moat than the number of integrations.
The cost of serving each additional customer matters, alongside the ongoing research and training needed to remain competitive. Better efficiency could improve profits, but competition could force Anthropic to pass much of the saving to customers through lower prices.
A vast market can support a valuable company. It can also support several companies working extremely hard to undercut each other.
If every new deployment needs months of bespoke engineering and support, impressive customer wins could come with disappointing returns. Reusable deployments and capable partners could reduce that burden. The financial payoff would be existing customers adding more workflows and spending more, without Anthropic’s cost of winning and supporting that business rising just as quickly.
Computing capacity often has to be arranged before the demand arrives. Reuters reported roughly $518 billion of future cloud, computing and infrastructure obligations in the draft. That covers coming years, rather than a single year’s expense. The payment schedule, conditions and flexibility of those commitments are essential to understanding the number. Reuters, September 28, via CNA
Securing capacity early could be a major advantage if demand keeps outgrowing supply. Having customers ready to pay and no computing capacity to serve them would be an expensive problem of its own.
Anthropic may have to commit to paying for capacity years before it knows how much customers will buy.
Reuters also reported that two unnamed customers each accounted for 12% of 2025 revenue, and that many large customers lacked long-term contracts. Reuters, September 29
On October 5, Investing.com, citing The Information, reported that Microsoft and Meta were reducing employees’ internal reliance on Claude and encouraging their own AI tools. Microsoft reportedly lowered its projected internal Claude spending by more than a third. Meta’s internal Claude Code user count reportedly fell from about 60,000 to 30,000, with layoffs contributing to the decline. October 5 reporting
The same report said customer spending on Anthropic models through Microsoft’s platforms continued growing. An internal spending reduction does not establish a retreat from offering Claude to customers. Investing.com
Microsoft and Meta also have resources and competitive incentives that an ordinary retailer does not. Their moves are not a verdict on enterprise demand overall. They do show why Anthropic has to earn its stickiness: some of its substantial customers can fund alternatives of their own. Nor should we assume these are the two unnamed customers in the filing.
A customer can reduce usage, switch providers or find a more efficient way to complete the same work. A lab’s committed capacity may be much harder to reduce.
Expansion can still be worthwhile. The contract terms will determine how much room management has to adjust if growth falls short of the plan, and public investors will need to understand that flexibility.
According to Reuters, 47% of Anthropic’s 2025 sales went through Amazon and Google’s cloud channels. That is a distribution figure, not a claim that those two companies themselves purchased nearly half of Claude’s output. They also act as investors, compute suppliers and AI competitors. Reuters, September 29
It is quite a relationship to fit into one slide.
For an enterprise buyer, purchasing through an existing cloud relationship can be easier than adding a new supplier. Anthropic itself has highlighted its availability across AWS, Google Cloud and Microsoft Azure as part of its expansion. Anthropic, May 28
These relationships can provide computing capacity and access to customers that Anthropic’s own sales team might struggle to reach. Anthropic has established a Claude Partner Network and a Services Track that evaluates partners on trained practitioners, production deployments and customer references. Anthropic’s partner program
Experienced implementation partners could turn a general-purpose model into repeatable solutions for particular industries. Fewer failed deployments and customers willing to recommend the results could help win the next sale. Partners would have to demonstrate those results, and they can support competing AI providers too.
If the workflow, business context and operating controls all sit in somebody else’s platform, with Claude serving as an interchangeable model underneath, much of the stickiness may belong to that platform. Anthropic would need to balance the reach of its partners with building products customers value specifically because they are using Claude.
A cloud company can benefit from its stake in an AI lab, from supplying computing capacity, and from helping distribute its products. It can have several reasons to support the relationship.
A public investor buying Anthropic shares has a more specific economic interest: the return available from owning Anthropic.
A strategic investment can make sense to a partner even when the same valuation requires more demanding assumptions from an outside shareholder.
These arrangements can support real customer demand. Customer revenue, investment capital and future purchase commitments nevertheless measure different things. To judge Anthropic’s economics, investors need to follow the money through to what remains with the lab after it delivers the service.
Otherwise, the same growing ecosystem can look like several separate confirmations of success when some of those businesses depend heavily on one another.
Anthropic’s IPO could become a referendum on the financial assumptions behind much of the AI economy.
If Anthropic can demonstrate durable margins, repeat customer spending and a credible way to finance its commitments, that would strengthen the case that a frontier lab can become a substantial business in its own right. Investors valuing OpenAI and other private labs would gain a public comparison. Suppliers would gain evidence about the economics of a major customer.
Investors could remain enthusiastic about usage and become less convinced about margins. AI adoption might continue growing while more of the benefit goes to customers through lower prices. Cloud providers, application companies and infrastructure suppliers could face very different outcomes. “AI is winning” would become an increasingly incomplete investment thesis.
I suspect the financing pressure could reach suppliers before end-user demand visibly weakens. If raising fresh capital becomes more costly, a lab may reconsider future capacity expansion. That could affect the expectations of businesses planning to supply it.
A signed contract provides more visibility than a hopeful forecast. Investors still need to consider who will pay, when, and from what source of cash.
Strong results and access to capital could give Anthropic more freedom to secure capacity, develop products and compete. Financing can influence which companies get the opportunity to become the winners investors are trying to identify.
The opening-day share price will also reflect IPO pricing, the number of shares available to trade, interest rates and enthusiasm for a scarce listing. A successful debut would establish an initial price; the following quarters would begin testing the assumptions inside it.
A stronger investment case would emerge if customers return and spend more, revenue becomes less concentrated, profits hold up after distribution fees and ongoing model development, and cash generation covers more of the company’s commitments.
Customers moving from pilots into recurring production work, expanding across departments and renewing because they can demonstrate results would show how far Anthropic has become a supplier they depend on. That dependence could give it some protection from price competition and help justify a higher valuation.
Anthropic cannot represent every business in the AI economy. Its execution, products and deal terms will matter. But a public listing could become one of the clearest places to examine a question that runs through the entire boom:
How much of the value created by AI will the companies building the intelligence actually get to keep?
You can believe AI will change the world and still be uncertain about the answer. Anthropic’s IPO could put a very large price on that uncertainty.