When AI labs say they are profitable, ask what they stripped
1 Oct 2026 · Buying AI
This month, reporting on Anthropic put four very different money numbers on one business. Each one leaves something different out. The habit to steal: ask what is still in the number.
When an AI lab says it is profitable, ask which costs are still in that number. The answer changes what "profitable" means.
Anthropic, the maker of Claude, makes a handy example. Everything below comes from news reports of investor documents and a draft IPO prospectus. Anthropic declined to comment on the prospectus, the prospectus is not public, and figures could change. This post is about the habit, not a verdict on the company or its IPO.
Four numbers, one company
- A net loss of about $42 billion in 2025. This is the all-in result, on revenue that grew 12-fold to nearly $4.6 billion, per the Reuters report (republished by CNBC).
- An operating loss of more than $8 billion. Same report: this leaves out write-downs of liabilities, mostly tied to earlier fundraising. About $34 billion of the net loss was an accounting charge on financing that could turn into shares. Reuters describes that charge as not money the company spent running its business.
- Positive "adjusted operating income." Documents seen by Bloomberg showed it for the second quarter of 2026 (CNBC summary), and the Financial Times reported Anthropic told shareholders to expect it again this quarter. The FT says the measure strips out costs including stock-based compensation. No dollar amount was reported, and I could not find the full list of what it excludes.
- Gross margins above 80%. Per the FT (via Irish Times), that is before revenue shared with distribution partners such as Amazon, and before the cost of training its models.
Not every exclusion is a trick. Number 2 drops a non-cash accounting charge, which is a reasonable thing to separate out. Numbers 3 and 4 are different: they skip things a business pays for again and again.
Training versus inference, quickly
Training is the cost of building a model. It comes in big, finite bursts, one per model. Inference is the cost of running the model every time someone asks it something, and it grows with usage. IBM has a plain definition of inference, and DigitalOcean lays out the cost shapes side by side.
So "profitable on inference" can be a meaningful claim. It suggests each answer pays for itself. It does not show that the whole business does, because someone still has to fund the next model. In the Prof G Markets episode on 30 September, one guest called leaving training out "earnings before bad things." That is a guest's opinion, not a finding. I could not verify the on-air figures about inference-only cash flow, so I left them out.
One more gap. Reuters reports Anthropic spent $7.33 billion on compute and infrastructure in 2025, out of $12.65 billion in total operating expenses. The coverage I found does not say how much of that went to training versus serving.
Questions that travel to any vendor
- Is this a standard accounting number or an adjusted one? Is there a bridge between the two?
- Are the missing costs one-time or non-cash, or do they come back every year?
- Does the number cover the whole business, or one slice (gross margin, inference only, one product)?
- Who produced it, and has anyone outside the company checked it?
The same questions work on a vendor's "cost per task" claim, a startup's "profitable" slide, or an AI price list.
Simple rule: When anyone says an AI business is profitable, ask which costs are still in the number: training, staff pay, partner shares, or only the cost of running the model.