This was the week the AI boom stopped being an abstraction and started showing up on balance sheets. Anthropic filed to go public, Alphabet reported a quarter that spooked investors despite booming cloud growth, and a research report confirmed that the cost of running these models keeps collapsing while the companies buying them still struggle to prove a return. Three data points, one story: the capital flowing into AI is now so large that the interesting question is no longer whether the technology works, but who ends up paying for it and whether the returns ever justify the bill.
Anthropic's numbers are extraordinary, and that is the problem
Anthropic put its books in public view for the first time, and the growth is hard to overstate. The filing shows revenue of roughly $16.3 billion for the first half of 2026, up from about $1.16 billion a year earlier, which works out to around fourteen times in twelve months. The second quarter alone brought in $11.6 billion and the company turned its first operating profit of $974 million, even as the half overall showed a small operating loss. The widely reported $80.7 billion net loss sounds catastrophic until you read the footnote: the great majority of it is a non-cash accounting charge tied to the rising value of convertible notes, which is to say the loss exists mostly because the company became more valuable.
The parts worth slowing down on are the dependencies. Around 48% of first-half revenue came through the cloud marketplaces of Amazon, Google, and Microsoft, the same firms Anthropic lists as investors, customers, distributors, and competitors all at once. The top five customers accounted for about a quarter of revenue, and the company has committed to spend well over $200 billion on computing with Amazon and Google. So the business that grew fourteen-fold rests on a handful of giant clients, a handful of cloud landlords, and a forward compute bill larger than most national budgets. There is also the unusual disclosure that its own products could pose a catastrophic risk to humanity, a line that now sits in a prospectus next to the revenue tables. The growth is real. The durability is the open question, and the filing is honest enough to say so.
Alphabet shows the other side of the same ledger
If Anthropic is the company selling intelligence, Alphabet is one of the companies paying to build the factory. Its quarter captured the tension perfectly. Google Cloud revenue more than doubled to $30.7 billion, with cloud operating income tripling to $11.1 billion and margins expanding sharply. By any normal reading that is a triumph. Yet the stock fell roughly 10% after the report, because capital spending hit $51.4 billion in the quarter and free cash flow swung to negative $9.1 billion. The company held its 2026 capex guidance at $195 to $205 billion and told investors 2027 would be meaningfully larger, with some estimates pointing past $300 billion.
The market's reaction tells you where sentiment has moved. A year ago, enormous AI spending read as conviction. Now investors are scrutinising the cash line and asking when the depreciation on all that hardware starts to bite. Amazon showed the same pattern, with its free cash flow swinging by tens of billions on the back of AI investment while AWS posted its fastest growth in years. The hyperscalers are spending like the payoff is certain. The share price reaction suggests their owners are no longer so sure.
Cheaper models, flat profits
Underneath both stories sits an awkward fact from this year's State of AI Report. The cost of getting a given quality of answer from a model has fallen roughly thirteenfold a year since 2023, yet only about 37% of companies in McKinsey's latest survey said AI had added anything to their operating profit, essentially unchanged from a year before. Meanwhile the combined revenue run rate at OpenAI and Anthropic reportedly climbed from around $30 billion to about $105 billion in eight months. The labs are capturing enormous value. Most of their customers are not yet showing it on their own income statements.
That gap explains a lot of the quieter product news this week, from Oracle moving to bill for AI tasks rather than seats, to the broader death of per-seat software pricing as agents do work that no longer maps to logins. When a model gets cheaper every year, the margin shifts to whoever builds something defensible on top of it, which is why governance, data, and integration keep surfacing as the real bottlenecks to production.
So what
The through-line is straightforward. AI capability is abundant and getting cheaper; AI capital is enormous and getting scrutinised. Anthropic's filing is a bet that usage-based revenue keeps compounding fast enough to outrun a staggering compute bill. Alphabet's quarter is a bet that today's cash burn becomes tomorrow's cloud annuity. Both bets might pay off. But the market has stopped granting AI an automatic pass on economics, and that shift in posture, more than any single model release, is what made this week matter.
A first operating profit of $974 million in Q2 on $11.6B of revenue is the number I keep coming back to. Spending is huge, but at least one player is already showing the bill can be covered. Curious how much of that ramp you think carries over to the cloud names.