Revenue, tokens and compute: AI can't account for itself, and that matters more than it seems

By Michael Parekh (AI Ramblings, Substack) · October 12, 2026.

The key point: according to Bloomberg, as cited by Parekh, investors cannot compare OpenAI's and Anthropic's revenues on a like-for-like basis because the two companies account for them differently. Anthropic books sales of its models through cloud partners such as Amazon at the gross amount. OpenAI counts only its own share. Analysts tried to recalculate both on the same basis and concluded they don't have enough information.

The headline metric, ARR (annual recurring revenue, meaning the billing rate over a short period scaled up to a full year), is not a standardized accounting measure either. It can be calculated from the last week, the last month or the best period in a quarter. The figures in the article show the gap: OpenAI expects to reach or exceed a $70 billion annualized run rate by year-end, but to bring in about $35 billion in actual revenue in 2026, half that. Anthropic, according to the same source, took in about $4.6 billion in actual revenue in 2025, against a run rate of more than $9 billion at year-end. Parekh adds his own estimate, which is not a confirmed figure: a $100 billion to $120 billion annualized run rate for Anthropic by year-end. He expected Anthropic's S-1 (its IPO prospectus) within days, with a listing planned before Thanksgiving.

The second issue is cost. According to the WSJ, only 11% of some 400 companies surveyed could accurately forecast their AI spending. A study by researchers from Stanford, Carnegie Mellon, Berkeley and Microsoft Research covering more than 6,800 tasks found that in 32% of cases the cheapest model ended up costing more. In one example, Gemini 3 Flash took 952 steps, spent about $14 and failed, while Gemini 3.1 Pro solved the task in 85 steps for around $1.

The third is compute. According to the WSJ and SemiAnalysis, renting Nvidia H100 chips for a year costs 60% more than it did twelve months ago, and providers are demanding multi-year contracts with upfront payments of up to 30%. According to the article, Anthropic is leasing more than 300 megawatts from SpaceX in a deal worth up to $45 billion. Even individuals are hoarding: The Information reports on a former Microsoft executive with seven GPUs in his basement.

Our take: these aren't three separate anecdotes but a single thesis. When a technology grows faster than the tools used to measure it, the narrative sets the price. Before the IPOs, ARR works as a story. The S-1 will turn it into accounting, and that is where every figure will have to hold up. We agree with Parekh that net revenue is the honest comparison, with one caveat: without seeing the contracts, neither method is "the right one", and we don't have that information.

The finding on tokens is the most useful in the long run. If the cheap model turns out to be the expensive one a third of the time, the habit of comparing models by price per million tokens falls apart. What matters is the cost per task completed. In general, this shifts value toward whoever measures, caps and audits agent spending, a management layer that barely exists today. It fits with something we have long argued: margins are moving from the model to deployment and control.

On compute, the short-term message is harsh: scarcity, rising prices and an advantage for those who already have capacity. That concentrates power in a few hands and makes it more expensive for startups to get in, which is a real risk. But the flip side, also visible in the article, is that demand is driving investment in energy and chips. Home hoarding also shows an appetite for local AI, which fits with wider access to the technology.

A final, lighter note from the same piece: OpenAI says GPT-5 cut sycophancy by more than two-thirds compared with GPT-4o, and that outside researchers at CUNY, King's College London and Transluce have seen the improvement. These figures are attributed to the company and to those studies. Advocates such as the Human Line Project argue the problem persists with more than 900 million weekly users. Unlike what happened with social media, an "infinitely tunable" product is being adjusted early, and that is a good sign.

In short, until revenue, tokens and compute are all measured the same way, the most eye-catching figures deserve skepticism, and it is worth waiting for audited accounts. The transition will be messy. Long-term abundance does not depend on our being able to measure it today, but it does depend on our eventually learning how.