AI's Real Turning Point May Not Be AGI. It May Be Profit.

Can an AI company make money the way a software company does

The oldest question hanging over generative AI was never really about the technology. It was about the accounting. Revenue has been exploding, but every time a user asks a question, a GPU has to actually do work. This isn't traditional software, where you build something once and then copy it at close to zero marginal cost.

For that reason, AI model companies have long been valued as businesses that look like software but sit economically somewhere between cloud computing and manufacturing. Bigger models mean higher training costs, and more users mean higher inference costs.

But a number that came out on September 13 is worth testing that assumption against. Reuters, citing the Financial Times, reported that Anthropic has told investors it expects positive adjusted operating income for two consecutive quarters. Reuters said it could not independently verify the report, and Anthropic did not immediately comment.

The 80% number matters less than what got subtracted to reach it

80%+ — Anthropic's gross margin, before accounting for partner revenue sharing and model training costs.

1,000+ — Anthropic enterprise customers spending more than $1 million a year, as of early 2026.

$2 / $10 — Price per million input / output tokens for Claude Sonnet 5.

Here is where caution is warranted. The gross margin above 80% that the FT reported comes before revenue sharing paid to distribution partners like Amazon, and before model training costs. In other words, it does not mean 80% is left over after every economic cost is accounted for.

So it's premature to say Anthropic has become a finished, high margin software business on the order of Microsoft Office. Until public financial statements are released, the exact costs excluded from "adjusted operating income" also need to be checked.

Even so, if two straight quarters of adjusted operating profit turn out to be accurate, it matters. At minimum, it points to the possibility that AI companies can escape a structure where more inference mechanically means bigger losses.

Why the economics may be improving now

The first factor is inference cost. The newest models don't automatically get more expensive. Anthropic prices Sonnet 5 at $2 per million input tokens and $10 per million output tokens, and says it delivers higher agentic performance at a lower price point than before. If the compute cost of handling the same amount of work keeps falling, then rising usage doesn't automatically translate into rising losses.

The second factor is a shifting customer mix. Anthropic says it passed 1,000 enterprise customers spending more than $1 million a year in early 2026. Enterprises aren't buying a $20-a-month chatbot subscription. They're buying outcomes in coding, finance, R&D and customer support that actually cut labor costs and save time. If AI saves someone several hours of work, what a business is willing to pay looks nothing like a consumer subscription fee.

The third factor is that not every query needs the flagship model. Simple requests can go to a smaller model, complex reasoning to a larger one, and coding to a specialized model, which lowers average compute cost. Users see one product, Claude, but behind the scenes the system is routing to whichever sufficient model is cheapest.

AI economics ≈ rate of revenue growth per user − rate of inference cost decline per user

That's not a precise accounting formula. But as an investment frame, it's useful. If the money a company collects per customer keeps rising while the compute cost of serving that same customer keeps falling, the gap between the two widens. That widening gap is what eventually shows up as gross margin expansion.

Why this number matters for Nvidia too

If AI labs actually start making money, the nature of GPU spending changes with it. Up to now, some investors have viewed GPU purchases as front-loaded investment pushed along by venture capital and Big Tech balance sheets, with a lingering worry that capex could eventually roll over if profits never caught up.

If instead AI labs can cover inference costs and still post operating profit from the money customers actually pay them, GPUs start to look less like a cost incurred to sustain losses and more like production equipment used to make money. If a single data center server can generate more revenue and more profit, that changes the quality of the entire AI investment cycle running from Nvidia and Broadcom through memory chips and power infrastructure.

The circular financing debate that flared up in the market a few days ago should be judged by the same standard. A structure where "a chip company invests in an AI company, and the AI company turns around and uses that money to buy chips" looks unsettling on its own. But if outside customers are putting in enough real cash, and the AI company is starting to generate operating profit and free cash flow, then the center of that circle stops being financial engineering and starts being actual demand.

Insight Times Editorial Desk