In April of this year, Anthropic removed the included tokens (the unit of AI consumption) from its enterprise per-seat pricing. The advertised price per seat went way down – from the $40-to-$200 range to a single, flat $20 a seat – but the real bill went up, because every use of AI a company’s people make now costs tokens. It’s the equivalent of your mobile phone company changing your unlimited mobile data plan to one where you pay a smaller fee for the line itself and then pay for every byte you transfer. The Register wrote it up on April 16. Nothing dramatic happened that day; it registered merely as another pricing change on the internet.
And yet, it reminded me of an idea I’ve been mulling over for the last decade.
Ten years ago, at Singularity University, I started writing a book I never finished. The idea was what I called “the Hourglass Economy,” and fashion was my canary in the coal mine. The GAP – the bastion of decent slacks at a decent price – was closing stores left, right, and center. So were Victoria’s Secret, Abercrombie, and many other high-street mainstays. Meanwhile, Gucci and Prada were doing a roaring trade, alongside the niche labels you had to know to know. And at the opposite end of the spectrum, H&M, Uniqlo, and Primark kept printing money. The top and the bottom of the market grew, while the middle got thinned out. Once you saw the shape in fashion, you saw it in consumer electronics, cars, package holidays, and many other places. The causes were clear enough: quality became table stakes, the internet killed the information asymmetry that trusted mid-market brands had been living on for decades, and infinite shelf space meant a pair of GAP trousers now sat next to a direct-to-consumer brand you’d never heard of.
Lately, I’ve been telling you that AI is bringing the hourglass back (I wrote about this in the radical Briefing just last week), but I’ve come to realize the mechanism is different.
The 2015 hourglass was a demand-side story where buyers walked away from the middle product. Nobody at the GAP failed at operations; customers simply stopped believing that “good value” was a reason to choose the GAP anymore. The AI hourglass is a supply-side story; it isn’t about what your customers want. Instead, it’s about what capability (or maybe more precise: intelligence) costs you. Same shape on the chart but completely different mechanics underneath.
The claim: among companies that have actually put AI into operations, through roughly the end of 2027, the cost of turning AI into operating advantage will be hourglass-shaped in company size – cheap for the very small, affordable for the very large, and worst for firms in the middle (call it fifty to a thousand employees). What matters is not AI adoption rates, or claims made by enthusiastic CEOs, but cost per unit of realized advantage – how deep a company actually got with AI, not just wether it uses AI at all.
The mechanism is deceptively simple.: A solopreneur or a five-person shop is buying intelligence at a price no vendor could sustain if everybody used it the way power users do. A company like radical, with all its six people, can buy a couple $100 Claude Max subscription and get nearly unlimited AI use out of it. We can MacGyver a throwaway workflow on our laptops on a Tuesday afternoon, ship it, and delete it Friday if it’s rubbish. Nobody has to approve anything; we just try, learn, and iterate. A very large company pays full metered freight for its AI usage, but spreads it across enormous volume, has the engineers to build the plumbing, and gets to amortize a data platform over ten thousand seats. The company in the middle gets the worst of both: too big to run on unvetted scripts and personal accounts – you now have an IT function, a security review, a SOC 2, and someone whose actual job is asking who owns this (all, of course, for good reason). And yet, too small to fund a real platform team. So they buy the same commercial tools their competitors buy, at metered prices, and they get whatever advantage a commodity gives them, which is very little.
The observation rests on an assumption, and it is admittedly already crumbling: consumer-tier AI is wildly, unsustainably subsidized. That subsidy is unwinding – the Anthropic repricing in April, OpenAI’s quiet throttling and re-tiering, and the whole industry discovering what an agentic workload costs when someone actually runs one. The pricing change I opened with is a crack in my own argument. If consumer pricing normalizes to something resembling the underlying compute, the bottom bulb deflates and the hourglass turns into a plain upward slope, at which point the price argument collapses. Keep an eye on the $20 and $200 tiers.
But even if the price advantage goes, small firms have a second one that stays: they can still move in ways a mid-sized or large company can’t, simply because they don’t run the same infrastructure or processes their bigger brethren have to. Whether that alone holds the bottom of the hourglass open, I honestly don’t know and only time will tell.
Which is why the number to watch is revenue per employee at firms under ten people – if I’m right, it pulls away from the middle band over the next two years (it’s the same underlying argument people like Sam Altman make when they say that we will see billion-dollar companies run by only a handful of people). If it stays flat, there’s no bottom bulb and this is just a ramp that favors size.
If you’re running a company in that middle band, the useful thing to do isn’t to follow my old Hourglass Economics prescription: go upmarket or go get big. That argument came from the demand-side hourglass, and it doesn’t transfer into our AI world. It’s also not the honored tradition of waiting for the vendors to package it – that’s how ERP, e-commerce, and cloud went, and the middle survived all three. But packaging arrives as a commodity, and commodities buy parity, not advantage. AI is different in a way that makes the old “wait and see” strategy a mistake. It shifts the question from “how do I use this tool to do the things I already do better?” to “how do I use AI to completely rethink what I am doing?”
The supply-side version is smaller and more practical: stop trying to be a mid-sized company doing AI, and start running a handful of genuinely autonomous cells that operate on small-company economics inside your corporate wrapper. Three to six people with their own tooling budget. Permission to build throwaway things that never touch production. A rule that most of what they make needs to survive contact with IT until it has proven it’s worth surviving. Pick one workflow this month, hand it to a small team with a credit card and no committee, and see what comes back in six weeks.
So, if the advantage you’re getting from AI right now is available to every competitor in your industry at the same price on the same website – what exactly did you buy?
@Pascal