For a couple of weeks now the same figure has been doing the rounds: a company that accidentally spent half a billion dollars on Claude in a single month, having neglected to put usage limits on its employees' licences. If you haven't seen it, just scroll your LinkedIn feed for a few minutes and you will.
It is a wonderful number, large enough to feel scandalous but attached to no one in particular. Every retelling traces back to a single sentence in Axios, sourced to an unnamed consultant describing an unnamed client.
A friend sent me Cameron Murray's Charts that lie go viral at around the same time as I first saw that $500m claim. It's a wonderful article, encouraging folks to apply a bit of old-fashioned critical thinking and call BS on numbers that seem to fit a particular narrative too conveniently. Quantium's CEO once said when looking at numbers about Quantium's website traffic "I don't know much about websites, but I work with numbers for a living and those numbers are bullshit". (Further investigation revealed some technical errors in the tracking which meant the numbers were, in fact, bullshit).
I've decided not to dwell on whether the share was meant as inspiration, a comment on my personality, or both. 😬
Instead, I built a small calculator to run the sniff test properly, and work backwards from $500 million to the organisation that would have to sit behind it.
Trivially, if you take a "reasonable" assumption for an extreme spent per person per month, let's say $20,000, then you can deduce that the organisation must have had 25,000 people all using AI at that rate. If this were an engineering team, that narrows it down to a comparatively small number of companies, all of them very well-versed in the concept of consumption-based pricing (from traditional cloud work), with mature FinOps functions that would likely have known better.
But spend in a large organisation is never evenly distributed; it follows a Pareto curve, a long tail of light users behind a small group of people running agents around the clock, smashing out prototypes at a rate that their typing speed is back to being the bottleneck again. Put in sensible assumptions — a floor of $100 a week for the lightest user, an ordinary heavy skew, and the input-heavy token mix that agentic coding typically produces — and half a billion dollars a month resolves to roughly 270,000 active Claude users, each averaging about $1,850 a month. A quarter of a million people, every single one of them a committed power user. (For reference, the largest employer in Australia is Woolworths Group which employs just a touch over 200k staff, so we can rule out that the anonymous company came from AU).
Of course, the average per user is interesting, but what we really want to know is how many tokens the power users are burning through. The single heaviest user would be getting through something like $6.3 million a month on their own — around 2,500 passes through the entire Linux kernel source, cover to cover. Around 360 per day, assuming they are of course working weekends (if your employer is funding you to the tune of $6m per month in tokens, it's the least you can do). That is roughly forty times the largest individual bill anyone has actually reported, the $150,000-a-month Anthropic developer the New York Times wrote up earlier this year.
None of this proves the half-billion didn't happen, it could have been for example a runaway fleet of unattended agents. But "incompetent engineer forgets to set up proper monitoring on newly deployed automation" is a much less interesting story.