Two years ago this wasn't a contest. OpenAI had the consumer brand, the revenue lead and the headlines. Then, in April 2026, Anthropic's annualized revenue passed OpenAI's, and the gap has widened since. By late summer, trackers had Anthropic at a run rate near $65 billion and OpenAI near $40 billion.

Same product category, same customers, very different machines. The simplest way to see it is to follow a dollar through each company.

Historically, roughly 70% of OpenAI's revenue came from people paying for ChatGPT. That's the engine behind its scale: ChatGPT crossed 1 billion monthly users in June 2026. It's also the problem. Most of those users are free, and every one of them costs money to serve.

Read the chart from left to right. OpenAI booked $13.1 billion in 2025. Serving that demand cost about two thirds of it, leaving a 33% gross margin, well short of the 46% it had targeted. Then came the expensive part: training new models, research, talent and sales. The operating loss, $20.9 billion, was larger than revenue itself.

The company is working both ends. Enterprise now makes up more than half of revenue, and its revenue run rate grew 50% in a single quarter this summer. And it has added a line no other lab has: advertising. ChatGPT's ads business passed $1 billion in annualized revenue in under 200 days.

The bill is still growing. Inference costs are projected to rise from $8.4 billion in 2025 to $14.1 billion in 2026, and projected cash burn from about $27 billion this year to about $63 billion next year.

Anthropic built the opposite business. Roughly 75% to 85% of its revenue comes from usage-based API billing, and consumer subscriptions are only about 5%. Most of its money comes from companies embedding Claude in their own products, where Anthropic never meets the end user. More than 1,000 customers now spend over $1 million a year.

In the quarter, Anthropic projected $10.9 billion in revenue, up 130% from $4.8 billion in Q1, and its first quarterly operating profit, $559 million. The swing came from compute: the cost to serve each revenue dollar fell from 71 cents to 56 cents in one quarter.

Three caveats keep this from being a clean victory lap. A meaningful slice of Q2 revenue came from a compute resale deal with SpaceX, and the company told investors not to expect the result to repeat right away. It books revenue sold through AWS, Google and Microsoft on a gross basis, which inflates the top line relative to peers. And it carries an estimated $80 billion in committed cloud spend through 2029.

Strip the charts back and one use case is paying for the whole industry. An estimated 65% of lab revenue now comes from coding. Claude Code went from $500 million in run-rate revenue in September 2025 to about $8 billion by May 2026. OpenAI's coding and work products reached 20 million weekly users. The labs aren't just selling models to software companies anymore. They're selling the software engineer.

Follow the dollar one step further and it lands on people. OpenAI paid an average of $1.5 million per employee in stock compensation in 2025, equal to about 46% of its revenue. Outside of compute, nothing else on either lab's cost line comes close.

Neither lab publishes individual pay. What is public comes from court testimony, acquisition prices, visa filings and, since late September, the Anthropic IPO prospectus reported by Reuters. These are the largest figures on the record at each company. Equity stakes and deal prices are not annual pay, so read each row for what it is.

Two things stand out. The biggest numbers are never salaries. They are equity stakes and all-stock acquisitions, which cost the labs no cash and only hold their value while the valuation keeps climbing. And the most prominent hires of 2026, Noam Shazeer to OpenAI and Andrej Karpathy and John Jumper to Anthropic, came with no disclosed terms at all. For scale: Google paid about $2.7 billion in 2024 in the deal that brought Shazeer back, and he left for OpenAI less than two years later.

Mostly from Big Tech, not from each other. A LinkedIn analysis of 1,680 Anthropic engineers published in June found that 405 had worked at Google and 273 at Meta. Only 48 had worked at OpenAI. Half of the engineering team has a FAANG background, and the median engineer arrived with 12 years of experience.

OpenAI has not been counted the same way, but the shape matches. Google supplied roughly a quarter of its hires between January 2023 and March 2026, and Google, Meta, Apple and Microsoft together supplied nearly half. Apple is the sharpest case: its July lawsuit against OpenAI says more than 400 former Apple employees now work there.

Where the two labs recruit from the same place, Anthropic is landing more. SignalFire found in 2025 that OpenAI engineers were eight times more likely to leave for Anthropic than the reverse, and the 2026 numbers point the same way.

For a founder, the read is simple. The labs are staffed by senior infrastructure engineers from five or six companies. Those are the people you want for your first platform hires, and the labs are bidding for them with stock that is about to get a public price.

Both labs sell intelligence. One is still paying to acquire the customer. The other is starting to get paid by them.

The eight figures from this issue worth keeping.

Neither company publishes audited financials. Figures are from company statements, investor communications and press reporting, and estimates are labeled.

Nothing here is legal or investment advice.