The headline trending on X reads: "Anthropic Hits $65 Billion Revenue Run Rate, Surpassing OpenAI."
But once I put both companies' numbers side by side, the $65 billion turned out not to be the most interesting part.
The real story: in the same quarter, one company made $559 million and the other lost $12.3 billion.
Here is everything I could verify, plus everything I could not.
The numbers, laid out
Anthropic's curve looks like this:
End of 2025: roughly $9 billion annualized run rate. April 2026: about $30 billion. May: past $47 billion. By the end of July: more than $65 billion.
That is more than a sevenfold increase from where it stood at the end of 2025.
For the most recent completed quarter: Anthropic booked more than $11.5 billion in revenue, against $787 million in the same period of 2025 — roughly 14x. Sequential growth of more than 50%.
On OpenAI's side: run rate above $40 billion. Second-quarter revenue of $6.7 billion, up 18% from $5.7 billion in Q1.
So on each company's most recently disclosed basis, Anthropic's run rate has passed OpenAI's for the first time.
But "run rate" is easy to misread
This deserves its own section, because it is where the arithmetic trips people up.
You may have noticed: Anthropic's $11.5 billion quarter times four is $46 billion, yet the reported run rate is $65 billion. Same for OpenAI: $6.7 billion times four is $26.8 billion, but the run rate is $40 billion.
Where does the gap come from?
Run rate is not "quarterly revenue × 4." It annualizes the most recent point in time — usually the final month of the quarter, sometimes the final weeks.
If a company is flat within the quarter, the two methods agree. If it is climbing steeply inside the quarter, "last month × 12" is necessarily far above "full quarter × 4."
Put differently: the gap between run rate and quarterly revenue is itself a measure of acceleration. Anthropic's gap ($65B vs $46B) is wider than OpenAI's ($40B vs $26.8B), which says it was climbing faster within the quarter.
One more caveat that has to be stated: Bloomberg flagged in its own reporting that the two companies may not calculate this figure the same way. One private company's self-reported number minus another private company's self-reported number is not a clean subtraction.
I would read this news as "the order of magnitude flipped," not "a precise $25 billion lead."
The real contrast is on the income statement
If I could keep only one set of figures, it would not be the run rate. It would be this:
Anthropic posted $559 million in operating profit in Q2 — the first operating profit in its history.
OpenAI's operating loss widened to $12.3 billion in Q2, from $9.3 billion in Q1. That figure includes stock-based compensation.
Which means: losses are expanding faster than revenue.
HSBC estimates OpenAI will lose roughly $14 billion across 2026, with cumulative losses near $44 billion from 2023 to 2028, and no profitability before 2030.
Same quarter, same industry, same wave of demand — one crossed into profit, the other grew its loss by a third. That carries far more information than who has the bigger run rate.
Why: the two businesses are shaped differently
This is not simply a question of who runs a tighter ship. The two companies sell different things.
OpenAI subsidizes hundreds of millions of free users worldwide. They pay no subscription, but every conversation consumes real compute. That is the classic consumer-internet playbook: capture users first, monetize later — and in that playbook, more users means bigger near-term losses.
Anthropic's revenue comes mostly from enterprises, and reporting points specifically to enterprise adoption of Claude Code as a major driver.
The economics of those two customer bases run in opposite directions:
Consumers are priced against attention, so you must burn cash to buy scale. Enterprises pay for outcomes, so more usage means more revenue.
And while inference costs remain high, the "more usage means more revenue" side crosses breakeven first, almost by construction.
Reporting also credits Anthropic with more efficient use of compute. Running a gateway, I have a very concrete feel for that sentence: on identical workloads, engineering choices can swing the bill by 2x without touching output quality. Before scale flattens gross margin, that difference lands directly on the income statement.
Three things I have to be clear about
First, none of this is audited. Anthropic is private and has no disclosure obligation. SiliconANGLE noted explicitly that the company did not say what method it used to calculate operating profit. The accurate phrasing is "figures the company told investors," not "financial statements."
Second, the two companies' definitions may not be comparable — Bloomberg said so itself.
Third, several claims in the auto-generated summary on that X trend could not be verified. Assertions like "Musk highlighted Anthropic topping RevenueBench" or "a talent influx from Workday and Box" have no primary source I could find, so they do not appear here.
There is a line under that summary box: Grok can make mistakes, verify its outputs. That line deserves to be taken literally.
The IPO is what these numbers are for
Why release these figures now? Because both companies are on the eve of going public.
Anthropic filed confidentially on June 1, 2026, targeting an October Nasdaq listing. OpenAI has also filed confidentially, but Anthropic may list first.
On valuation, two numbers must be kept apart:
The last actual valuation was $965 billion (Series H-1, May 2026).
"$2 trillion or more" is a reported target, not an accomplished fact. Some investors believe the current growth curve could support more than double the last round at listing.
These are categorically different numbers, and blending them misleads people.
Against that backdrop, the meaning of that first operating profit becomes clear: at the pricing table, a growth story that can prove it makes money and a growth story that can only prove it grows fast are valued by two different logics.
What I take away
One: the shape of the curve matters more than the endpoint. $9B → $30B → $47B → $65B — the most valuable information there is not the final number, it is the absence of an inflection. An absolute figure can be assembled from one or two large deals. Four consecutive observations that keep accelerating are very hard to fake. Slope tells you more than size.
Two: whether free users are an asset or a liability depends on when the bill comes due. For twenty years the default answer was "asset" — scale converts itself into money eventually. That $12.3 billion loss says that while inference costs stay high, every free user is a real invoice. The question is when you have to show your work. Going public is exactly that moment.
Three, and this is the part worth sitting with: what got validated in this round is not "whose model is smarter." It is "who found customers willing to pay for outcomes."
Behind Anthropic's curve stands a set of enterprises. They buy Claude Code not because it is cleverer, but because it saves engineering hours — and engineering hours carry an explicit price.
When a customer can compute how much you saved them, you no longer have to persuade them to pay. That is true for model companies, and it is just as true for the rest of us building AI products.
Sources: Bloomberg (Aug 17, 2026) reported Anthropic's annualized run rate surpassed $65 billion as of end-July, more than 7x its end-2025 pace (trajectory: ~$9B end-2025, ~$30B April 2026, past $47B in May), with the most recent completed quarter above $11.5 billion versus $787 million a year earlier. Bloomberg (Aug 13, 2026) reported OpenAI's run rate above $40 billion. Q2 comparisons (OpenAI: $6.7B revenue, +18% QoQ, $12.3B operating loss vs $9.3B in Q1, including SBC; Anthropic: $11.6B revenue, +50%+ QoQ, $559M operating profit, a first) come from Wall Street Journal reporting as relayed by SiliconANGLE (Aug 18, 2026). HSBC's loss estimates, Anthropic's June 1, 2026 confidential IPO filing targeting an October Nasdaq listing, the $965B last-round valuation (Series H-1, May 2026), and the reported "$2 trillion or more" target all come from public reporting; the target valuation is an expectation, not an accomplished fact. Both Bloomberg and SiliconANGLE caution that the two companies' revenue definitions may not be comparable, and that Anthropic did not disclose its operating-profit methodology. Claims in the X trend's auto-generated summary that I could not verify (Musk on RevenueBench; talent from Workday and Box) have been removed; that summary is Grok-generated and the platform itself flags it may be wrong.
Top comments (0)