Who is actually losing money on AI, and where do the Chinese labs fit?
Ok, let us go down the list, because the picture is messier than "AI is a bubble" or "AI is printing money".
The US side
One top lab made an operating profit of about $974 million in one quarter, then flipped to an operating loss of about $567 million the next, on roughly $17 billion of revenue. The scarier number is the future: around $895 billion in signed compute commitments that have not hit the books yet. Its giant net losses are mostly accounting charges rather than cash going out the door; in the first half, $72.4 billion of an $80.7 billion loss.
Its biggest rival is projected to lose around $14 billion this year and not turn a profit until about 2030.
A social media giant saw its free cash flow go negative, minus $2.12 billion last quarter, while planning capital spending of $135 to $145 billion this year. It is now selling up to $25 billion of bonds, and a rating agency has moved its outlook to negative. The search giant, meanwhile, pays for its AI from profits made elsewhere.
And the Chinese labs are dirt cheap, so surely they are the ones winning?
The Chinese side
Prices per million output tokens tell the story: roughly $50 for a top US model, $15 for one flagship Chinese model, $4.40 for another, and under a dollar for the cheapest. Wow. But cheap to buy is not the same as cheap to run, and nobody outside really knows the real costs.
One report has the cheapest lab earning about $70 million in seven months while still posting a loss of roughly 715 million yuan. A giant tech parent spends about $10 billion a quarter on AI, sees profit down 38% and AI cloud revenue up 45%, and does not say whether the models pay for themselves. Part of the low pricing is simply a land grab for market share, helped by cheaper electricity.
So the honest scoreboard is: nobody is clearly winning yet, and the ones with the cheapest prices are also the ones telling you the least.