Why are some AI companies hitting $10M ARR in six months?

Two ways to $10M, and this is the weird one

The piece opens on a puzzle the author keeps seeing: companies going from zero to $10-15M in yearly revenue in a matter of months. He splits them in two. The first group sells cheap software to a huge, ordinary market, the 'AI operating system for restaurants / law / HVAC'. That is vertical AI (an AI product built for one specific industry), and the numbers there are believable enough.

The second group is the strange one, and the only one the essay is about: companies selling infrastructure like memory, inference, security or developer tools, that hit '$12M ARR in six months since pre-seed'. ARR (Annual Recurring Revenue) is just the run-rate of subscription income, annualised.

Eight figures of it, that fast, from a plumbing-layer company nobody has heard of, doesn't register at first. The rest of the essay is one answer to 'how?'.

Sell to the frontier, not 'the enterprise'

The obvious guess is 'they sell to big companies'. The author's twist is which big companies: the AI-native firms already operating furthest ahead, the Anthropics, OpenAIs, Harveys and ElevenLabs of the world.

His reasoning is that those companies are on a treadmill where every few months they must ship a new capability and justify a higher valuation. Running that hard, they keep slamming into technical problems almost nobody else has hit yet, because almost nobody else operates at that scale or speed. If you happen to be in the room when they trip over one, and you've figured out a fix, you can be charging $100k a month, then $500k, off a single account.

The move isn't 'find a big market'. It's 'find the few customers living six months in the future, and solve the problem they just ran into'.

Why $500k a month isn't insane: the BATNA

The prices sound absurd until you ask what the customer is comparing you against. They are not weighing you versus another software vendor. They're weighing you against their own next-best option.

That option is the key idea, and it has a name: BATNA (Best Alternative To a Negotiated Agreement), the thing you'd do if the deal fell through. For a frontier AI company the BATNA is ugly. It's six months of their scarce engineers building it in-house, or a $20M-a-year compute bill they can't shrink, or a launch that slips. If your product cuts an inference bill from $20M to $13M, paying you $3M is one of the highest-return decisions they can make.

So the vendor is priced against engineering time, compute and lost growth, not against a competitor's sticker. Measured that way, a big number can be the cheap option.

The moat is information, not technology

Here is the sharpest line, and the one worth keeping: 'the moat isn't always technical... the information advantage comes before the technology advantage.'

Could a competitor build the same fix? Probably. But they won't even know the problem exists until they are in the same room with the same frontier customer. These problems aren't on Twitter, in Gartner reports, or anywhere the broader market can see them yet. Knowing a problem is real, months before it's public, is the edge; the code is almost secondary.

The same idea shows up across finance, security and code review under different names: the durable advantage is being trusted inside the system where the problem lives, not being marginally smarter than the next team. A better model doesn't hand your rival the knowledge that the problem exists.

Customer concentration as a feature, and category creation in reverse

Traditional software wisdom says depending on two or three customers for 80-90% of revenue is dangerous. The essay argues that early on it's closer to a feature. The handful paying you are the ones seeing the problem first; everyone else gets there eventually but isn't there yet, so your first few million can honestly come from three logos.

Then the flywheel. Land one marquee customer and their competitors hit the identical problem and come to you. Engineers move between companies and carry the knowledge with them. A case study goes out, and what started as one account becomes a whole category.

The author calls this 'category creation in reverse'. Instead of building a product and convincing a market it has a problem, you start with the few already living in the future, solve their problem, and wait for the market to arrive.

The honest counter: who wrote this, and who's invisible

Read it with one eye on the source. The author is an investor writing about the companies investors fund, so the framing flatters a portfolio thesis. 'Customer concentration is a feature' and 'just get in the room' are the kind of advice that sounds wise from the winners' chair.

The bigger tell is survivorship. Every company in the essay landed the frontier customer. The ones that got in the room and never converted, or bet the company on one account that then churned, don't appear, and there's no number for how often that happens. 'Get in the room' is a position, not a repeatable method, and the essay never says how an unknown gets access to those engineers in the first place.

By its own logic the moat is temporary: it lasts exactly until the problem becomes public and the rest of the market catches up. That's a real business, but a treadmill, not a fortress.

Vocabulary

• ARR (Annual Recurring Revenue) — The annualised run-rate of a company's subscription income.
• BATNA (Best Alternative To a Negotiated Agreement) — What you'd do if the deal fell through; it sets your price ceiling.
• Vertical AI — An AI product built for one specific industry, like law or HVAC.
• Moat — A durable advantage that stops competitors copying you.

If you're building — what to watch for

• The 'sell to the frontier' logic inverts for a solo or domain builder: your edge is the expensive, un-Googleable problems you personally hit as an early heavy user, and that list is your candidate product set.
• Price against the customer's real alternative, which is their engineering time, compute bill or a delayed launch, not against a competitor's sticker; that framing is what makes a large number look cheap.
• Treat information advantage as separate from technical advantage: knowing a problem is real before the market does is a moat a smarter model can't hand your rival.
• Customer concentration buys speed early but is a genuine fragility; don't read the essay's 'it's a feature' as permanent safety.

Reading it critically

• VC-authored and winner-sampled: every example landed the frontier customer, and the ones that got in the room and never converted are invisible, with no base rate given.
• 'Customer concentration is a feature' is genuinely fragile (one churn = 80-90% of revenue gone); the essay concedes 'eventually it's dangerous' but skips the timing entirely.
• 'Get in the room' is a position, not a method: there's no account of how an unknown actually gets access to frontier engineers, which is the hard part.
• By its own logic the advantage decays the moment the problem goes public, so it may be a fast head-start rather than durable value.