For a long time, I (among others) have argued that regardless of what you think about contemporary GenAI's value, ethics, or utility, it's fundamentally financially-unsustainable in its current form (and nobody's presenting a way to change that).
Alt
Graph of Adjusted Operating Margin trending from -221% in Q1'25 to +19% in Q2'26.
Anthropic's S1 registration for Q2 of this year superficially seems to imply that I'm wrong: their "adjusted operating margin" has for the first time trended positive, meaning that - for this particular measure of profitability - they're actually managing to sell AI services for less than it costs to provide them, which could be the fist time ever that this has happened for any "AI company".
Does this mean that I'm wrong? Maybe... But I wouldn't be convinced yet. In Anthropic's case, at least, the profitability (by this metri)c is absolutely overshadowed if you strip away their "adjustments": debt repayments, dividends, and the ever-growing cost of model training in an increasingly-competitive environment aren't included. And within the report itself there's a ton of warnings: the rising cost of hardware and energy, the fluctuating cost of computation, etc.
Also: Anthropic still made a net loss for the period in real terms. Even though they're one of the least-unprofitable AI companies on paper1.
So no, I'm not convinced yet that large-scale general purpose LLMs can be made financially-viable under the current operating strategies, and I've not yet seen a route by which they can be made to be so. I suppose that perhaps I'll link back to this post at some point in the future as the point at which I should have seen the tide turning on the commercial value of GenAI! But I wouldn't count on it.
Footnotes
1 Contrast e.g. Amazon or Google's AI "adjusted operating margin" on the same terms, which would be closer to -1000%.
I think thereβs two halves to the discussion.
Can LLM services be sold for a profit? Looks like the answer here is a guarded yes, particularly when itβs sold as an API service.
Can new models be created and trained in a way that allows for a return on investment? This is where it gets much harder, especially with some of the open weight models being released for free that are almost as good.