Anthropic told a small group of shareholders it will report an adjusted operating profit this quarter, marking a second consecutive period of profitability as the company prepares to go public. The Bloomberg report landed Sunday night. The company is projected to earn approximately $559 million in operating profit for Q2 2026. The metric strips out certain exceptional or one-time costs. That qualifier — adjusted operating income — is doing a lot of work, and anyone running production systems that call Anthropic APIs should understand exactly what it's not saying.
I run ten platforms in production. The construction ERP at coenconstruction.com calls Claude for estimate rewrites. The review dialer on valhalla-k9.com hits the API for SMS personalization. The agent on thepunchlist.ai uses it to parse change orders. Every one of those integrations assumes the vendor will still be operating, at roughly the same price, eighteen months from now when the next construction season starts. When I decide whether to keep calling Claude or rewrite everything to hit a self-hosted Llama variant, I am not betting on whether Anthropic can post a profitable quarter on an adjusted basis. I am betting on whether the API will still answer at a price I can pay, and whether the company will still be solvent enough that Amazon, Google, and Microsoft keep the GPU clusters online.
Adjusted operating income is a real number. It is also a carefully scoped one. This value does not take into account many expenses, such as payments to employees in stock. Stock-based comp is a real expense. It dilutes existing shareholders. It is how you retain the engineers who keep the models from hallucinating dosages and decimal points. The company accumulated cumulative operating losses of approximately $10–15 billion from 2021 through 2025. That is the actual stack of red ink the company has printed since founding. A business that reports $559M in adjusted profit for one quarter but carries $10-15B in cumulative losses is not profitable in the way a construction company or a SaaS with three years of trailing positive EBITDA is profitable.
The $80 billion commitment nobody puts on the income statement
The larger issue is not what happened last quarter. It is what happens over the next three years. Anthropic is projecting its cloud spend will reach $80 billion for the period through 2029, spread across Google, Microsoft, and Amazon Web Services. In April, Anthropic committed to spend more than $100 billion over the next ten years with Amazon Web Services. That promise is equal to about a fifth of AWS's backlog of contracted work, which reached about $496 billion in June. Those commitments do not show up as liabilities on a quarterly adjusted operating income slide. They show up as future obligations that the company must meet regardless of whether revenue keeps growing at the current pace or decelerates when the next model refresh costs more to train than the last one did.
I have written before about API pricing dropping 72% and GPU costs rising fifteen percent while model pricing stayed flat. Those are both vendor risk scenarios where the cost structure changes faster than the integration can adapt. Anthropic posting two consecutive quarters of adjusted operating profit does not eliminate that risk. It proves the unit economics can work at the current scale, with the current customer mix, at the current rate of compute cost decline. It does not prove the unit economics work when the company has to start recognizing $80 billion in cloud commitments as actual expense over the next thirty-six months, or when stock comp vests for the employees who built Claude 3 and 4 and are now being retained to build Claude 6.
The construction ERP does not call a model to rewrite every line item. It calls the model once to generate a template, caches that template in D1, and serves it until a project manager edits it and the edit is worth persisting. That architecture exists because I looked at a bill that said "4 million tokens" and realized I was paying for the same generation seven hundred times. The decision to cache aggressively is a hedge against the vendor raising prices, the model getting deprecated, or the API going away entirely because the company that operates it could not make the adjusted numbers turn into actual profitability before the investors stopped writing checks.
Adjusted operating income is the number you show investors when the quarter turned green but the five-year stack is still underwater. It is a real milestone. It is not the same as being profitable enough to survive the next three years of cloud bills.
Anthropic investors expect the AI company to go public in October at a valuation of $2 trillion or more, according to the Financial Times. That IPO will happen. The S-1, when it goes public, will include GAAP financials that show the $10-15B in cumulative losses, the $80B in future cloud commitments, and the stock comp expense that does not appear in adjusted operating income. Public market investors will price the company based on what those numbers say about 2027 and 2028, not on what Q2 2026 said about the last ninety days under a specific accounting treatment. The relevant question is not whether Anthropic turned a profit on an adjusted basis for two quarters. The relevant question is whether it can stay profitable on a GAAP basis for eight consecutive quarters while revenue growth decelerates from 10x to 2x and AWS, Google, and Microsoft start billing for the full contract value of those cloud commitments.
I run systems that depend on APIs from vendors who do not publish their adjusted operating income. Twilio publishes GAAP financials. Cloudflare publishes GAAP financials. Stripe publishes enough detail in their private financing rounds that you can back into whether they are burning cash or generating it. Anthropic, until the S-1 goes live, is a black box with two data points: $559M in adjusted Q2 profit, and $10-15B in cumulative losses since 2021. The gap between those two numbers is the difference between a company that posted a green quarter and a company that has closed the five-year stack and can pay back everyone who funded it. For a business that calls the Claude API in production, that gap is the difference between a vendor that will still be operating in 2028 and one that might get acquired, shut down, or repriced out of reach when the next fundraising round does not close.
Anthropic posting adjusted operating profit for two consecutive quarters is good news. It proves frontier AI inference can generate positive unit economics at scale, which was not obvious two years ago when GPT-3.5 was the default and every new model launch deprecated three endpoints. It is not proof the business is solvent on a five-year horizon. That proof will come when the S-1 drops and the GAAP numbers show whether the $80B in cloud commitments and the cumulative losses fit inside a revenue trajectory that can actually support them. Until then, every API integration is a bet on whether the adjusted numbers turn into actual profitability before the stack collapses under its own weight.