An AI human behind
every counter.

Millions of people translate words into clicks. That is not a job. That is a bottleneck.

A cashier's core function is transcription. A person says what they want; the cashier converts it into an electronic record. That translation layer was never automated, because until about eighteen months ago it could not be. Voice was not good enough, latency was not good enough, and a synthetic face was uncanny.

All three changed at once. What used to require a person now requires a tablet.

What that is worth is not mainly a payroll line, and I would rather not sell it as one. A counter that serves ten people at once has no queue. A counter that speaks thirty languages does not lose the order — or the dignity — of the customer who speaks none of them. A counter that never gets tired offers the upsell on the worst shift of the week as reliably as on the best. And the person who was transcribing gets the part of the job a machine cannot do: know a regular, fix a problem, run a room.

The saving is the proof, not the pitch. It is what makes this transition inevitable rather than optional. But a merchant who buys this in order to have fewer people has bought the wrong thing, and in my experience most of them do not.

Everything reduces to one counter.

If a single counter works, the company is an execution problem. If it does not, nothing else matters. So here is the whole business in one table.

The last line is the one that matters most. Volume makes this business more profitable, not less. Most hardware-plus-service companies work the other way.

One tablet. Order, upsell, payment, kitchen ticket.

The face is the interface. A camera detects presence and the AI wakes. It takes the order conversationally, handles modifiers and a mid-order change, upsells every ticket, prompts the tip, takes the card and fires the ticket to the kitchen screen. One tap hands the customer to a human, mid-order, nothing lost.

These are acceptance criteria for the build contract, measured with synthetic probes on real Los Angeles networks. They are not aspirations.

Three ways a merchant pays. We win on all three.

Subscription today. A transaction-embedded take as the payments layer lands. The tiers below are the model, not today's revenue.

One deployment, three ways to monetise it. The same land-then-expand curve Stripe and Toast ran, applied to a category an order of magnitude larger.

How the take rate is actually captured.

This is the part most decks wave at. Here is the mechanism.

Of a 2.7% card-present fee, roughly 1.7–1.9% is fixed: interchange to the issuing bank and assessments to the card networks. Nobody negotiates that, including Stripe. The contestable slice is the processor's 0.8–1.0%.

We take 0.3–0.5% of it by becoming a payment facilitator under a sponsor platform rather than registering ourselves. The sponsor holds the card-network registration, the sponsor bank and the money-transmission licensing. We hold the merchant and set the retail rate. Two to four months and low six figures to integrate, against the $500K–2M and 12–18 months a self-registered facilitator costs. That question returns at roughly $500M of routed volume, and at that point the take approaches 1%.

Visa and Mastercard own the rails, and that is not the fight. We own the counter, and the counter decides which rail the transaction takes.

Ten places the same box earns.

One terminal on one counter, monetised ten different ways. Two earn today. One is funded by the next round. The other seven are plans, sized here so you can check the arithmetic rather than take an adjective. Every layer runs on hardware already sitting on the counter: no second install, no second sale.

The market column is the size of the intersection each layer sits in, not revenue we claim. Each layer reinforces the next: labor unlocks transaction routing, transactions unlock the identity graph, identity unlocks lending, hardware unlocks the scale that makes all of it worth doing.

Capturing one to three percent of the stack above at maturity is the arithmetic behind a hundred-billion-dollar outcome. Gated on execution, and on raising every round between here and there.

Bounded honestly, in three rings.

Kiosk penetration worldwide sits around 20%. The counter was not skipped because it was unattractive; it was skipped because the technology did not exist.

Everyone gets paid after the merchant pays us.

Three machines run in parallel, and not one of them requires payroll before revenue. That is why ten thousand counters inside twelve to eighteen months is a capital question rather than a headcount question.

What the round builds, cycle by cycle

The build is specified as five acceptance cycles with hard gates — no calendar deadlines, only order and proof on working software at a real counter. Each one unlocks a different part of the business, and they are not interchangeable.

How engineering is paid, and why I chose the expensive-looking option

I built the thing that is running. Alone, with AI tooling, on personal money. That is the part of this company that is already proven, and it is proven by a product in production rather than by a claim about a team.

What I am buying now is speed. Senior engineers, salaried out of this round under a modest vesting grant. Nothing is signed — I do not execute a salary I cannot fund yet, and I am not going to pretend a handshake is a hire. The people I have been talking to came to me, not the other way round — YC-backed, ex-Google AI, their own startup having not worked out, back in the Valley shortly. That is a better signal than a recruiter’s pipeline, and it is still not a signature. If they do not come, others will; at this level in this city they are findable, and the product does not stop existing while I look.

The arithmetic is simple. A share of gross margin looks free while there is no margin, and gets expensive precisely when things work: at a thousand counters the company clears roughly $400,000 a month in gross margin, and every point of that given away is permanent. Salary is a fixed cost that a thousand counters covers many times over. I would rather carry a known cost now than hand over a share of the thing I expect to compound.

What that costs me is honesty about the risk: salary starts before revenue does, which moves the risk from them onto me and onto this round. That is a deliberate trade, and it is the reason the round closes in October rather than whenever.

I can economise on almost everything in this company, and I do — no office, no advertising, no VP of anything, a founder on personal credit. I will not economise on what people need to live. Underpaid engineers are how a seed-stage company quietly dies: the good ones leave, the ones who stay build something you have to throw away, and you discover it at the worst possible moment. Either this team is paid properly, or a different team that is just as good gets paid properly. Those are the only two options I am willing to run.

That is what the $500K engineering line is — roughly a year of a team at this level plus the first in-house hires, not a placeholder.

I can reach a thousand counters without you. That is the point.

The honest version of the ask, because it is stronger than the usual one.

The model I ended up with does not consume capital to add a counter. A merchant who buys the terminal outright pays $1,500 at signing, which covers the hardware and the install with margin left over. The field force is paid out of money the merchant has already sent us — $100 on a collected deposit, then a percentage of collected revenue. Marketing is the idle screen itself. Every counter funds the next one.

Run that forward and $250,000 gets to roughly a thousand counters. I have modelled it, and it works on paper.

It also works with no margin whatsoever. One quarter of slower collections, one hardware batch stuck at customs, one bad month, and the whole thing stalls with no reserve behind it. It takes about three times as long. It puts the founder on personal credit for the duration, which is exactly the single point of failure a fund is right to underwrite against. And it means the field force stays on 1099 in California longer than it legally should.

The field force is one road to a hundred thousand. It is not the only one.

Everything modelled on the model page is a field-sales business: people knock, people install, growth compounds until headcount caps it. That is the conservative road, and it is the one the numbers describe.

There is a second road, and it is not in any figure I have shown you. A recognised face on a hundred thousand idle screens is not advertising we buy — it is advertising we own, and a merchant who already knows the name closes in one visit instead of three. Category demand pulled rather than pushed does not compound at twelve per cent a month; it steps. The same hundred thousand counters that take five years by headcount could take two or three if a brand-led wave lands, and the honest position is that I am building for that outcome and underwriting the slower one.

Which is why the ambition beyond it is stated plainly: 350,000 counters, then a million. Not as a projection — as the shape of the thing being built, and the reason the architecture, the partner terms and the cap table are all set up for a company that runs for decades rather than one that exits at the first good offer.

How marketing gets paid for after the first thousand

The field force gets us to a thousand counters. It does not get us to a hundred thousand, and no advertising budget I could raise would either. The screen is the inventory. Every idle counter is attention we own outright and pay nothing to acquire, which makes it something a consumer brand wants and something we can trade rather than sell.

I have already run this negotiation once. In the previous business I sat with top-tier talent and their management about building a brand jointly, structured on a percentage rather than a fee — the principal takes equity or a share of what the brand earns, and no cash leaves the company. That is the same structure I intend to use here, and it is the only marketing model I am interested in: partners on equity, not invoices.

The honest timing: at a thousand counters that conversation is not interesting to a principal of that calibre. At the valuation and footprint the next round implies, it is — which is precisely why it belongs in the plan for the thousand-to-hundred-thousand stretch rather than in this round's use of funds. Nothing is signed, and I am not pricing it in. What I am telling you is that the access exists, the structure has been negotiated before, and the inventory to pay for it is being installed right now.

Who is already behind it

Compute is credited. The face is on partner terms. The growth pays for itself.

I closed a working business to build this one. That is the whole argument.

And before that, I was him.

Each step is priced off the one before it. None of them is the destination.

The reprice at $150M is a waypoint in year one, not an outcome. Read the whole ladder before pricing the first rung.

These are ranges in which comparable companies raised, not valuations of this one. The right-hand column prices the subscription line only; the $20B figure below adds the payments layer those same counters already route. Every row after the first depends on the row before it, and on raising each round in between.

What could go wrong. Written by me, not by counsel.

You will find every one of these in diligence anyway. Better that you find them here first, with the answer already attached. Where there is no answer yet, it says so.