Who Do You Want To Be

THE NUMBER: $10,000 per employee, per month. That’s what Higgsfield spends on inference. Four million dollars a month, across roughly four hundred people, and not one dollar of it trains a model. It runs one. Read the number as rent, not R&D — it’s the standing monthly cost of doing a thing nobody else can do as well. Hold it next to the other Higgsfield number: a billion dollars of ARR in eighteen months, faster than anyone on earth not named OpenAI or Anthropic. Keep both in your head. The gap between “we spend ten grand a head a month running someone else’s models” and “we’re the fastest company to a billion in the history of the category” is the whole argument. It’s the difference between a cost and a moat, and this week they turned out to be the same line item.

A billion dollars, no moat, on rented brains

Start with Higgsfield, because it breaks the model you’ve been carrying around in your head. They don’t own a frontier lab. They didn’t train anything. They rent intelligence by the token — four million a month of it — and point it at one thing: content. Studio-grade video, made at a volume a traditional studio can’t touch, pushed out by a machine of a hundred and fifty professional creators sitting on top of the rented models. Alex Mashrabov, who ran generative AI at Snap before this, will tell you the compute line isn’t a phase they’re going to grow out of. He expects it to climb toward a hundred thousand dollars per person per month. That’s not a company waiting for its costs to come down. That’s a company that has decided the cost is the business.

Every instinct trained on the last twenty years of software says this is a bad company. Gross margins upside down, a variable cost that scales with every unit sold, no proprietary technology underneath. By 2015 SaaS logic it’s a house of cards. And it’s growing faster than almost anything ever has. Something in the old model is wrong, and the thing that’s wrong is the assumption that the technology you own is the thing that protects you. Higgsfield owns none of the technology. It owns the behavior.

a16z drew the map and didn’t read it back

Here’s the part that made me sit up. The same weekend Higgsfield’s numbers went around, Andreessen Horowitz published a long essay on why they think OpenAI wins. Strip out the throat-clearing and it’s a clean piece of strategy. There are, they say, four levers you can pull to build a durable business at frontier scale: create new behavior, own distribution, price to capture value, and build high switching costs. Then they do something useful — they kill two of the four in a single paragraph, almost in passing.

Lever four, switching costs, is “basically out.” The model layer is a Red Queen’s race; you can swap one model for another task by task, and everyone gets better at the same time. Lever three, pricing, is what everybody’s chasing and nobody’s winning — whoever leads on price-to-performance is just whoever’s winning at scale anyway, and the whole industry is racing that number to the floor. Opus 5.5 runs forty percent cheaper than the model before it. OpenAI cut one model eighty percent in July and another fifty percent in September. Claude Sonnet 5.5 shipped yesterday morning, faster and cheaper again. You cannot build a wall out of a price that halves every quarter.

Which is our own argument coming back around, and I’ll say it plainly because we’ve been saying it since April: if you can’t price it and you can’t lock it, the labs are going to have to climb into the actual businesses of their customers to capture value. There’s nowhere else for the money to be. a16z just published the top half of that syllogism and stopped one line short of the conclusion.

The lever they waved off too fast

That leaves two levers standing, and a16z spends the essay on the one we’ve beaten to death — distribution. They’re right about it. I’ll come back to it. But watch what they do with switching costs. They declare lever four dead and move on. I don’t think it’s dead. I think it’s re-earned.

Switching costs don’t come from lock-in tricks anymore — no data gravity, no contract you can’t get out of, no proprietary format. They come from one place: being the only one who can do the thing. The second the behavior you unlock is something a customer can’t get anywhere else, the switching cost reappears on its own, because there’s nothing to switch to. Higgsfield has, for the moment, close to infinite switching costs — not because they trapped anyone, but because if you want what they make, at the quality they make it, there is no second vendor. That’s lever four, resurrected through lever one. a16z had the pieces on the table and didn’t snap them together.

Best is not only

Now the trap, because most founders fall in it. Being the best is not the same as being the only, and only one of those is worth anything.

I play tennis. So do Carlos Alcaraz and Jannik Sinner. Nobody is buying a ticket to watch me, and here’s the part that stings: nobody is buying a ticket to watch the fortieth-best player in the world either, a man who would beat me six-love, six-love before I won a point. “Best” is a rank in a crowded field, and the field re-sorts every season. Sinner was on top; then Alcaraz; the ranking is a treadmill, and the treadmill never stops. Build your company on being the best at a thing lots of people do, and you’ve signed up to defend a position that every competitor and every new model release is actively trying to take from you. You’ll spend everything you have staying one rung up a ladder that keeps growing rungs.

“Only” is a different kind of thing entirely. “Only” isn’t a rank. It’s a category of one. Higgsfield isn’t the best at a thing lots of people do — it’s the sole road to a thing you flat-out couldn’t do last year. The twelfth-best general-purpose model has no switching costs protecting it, because “twelfth-best at everything” is another way of saying “replaceable by any of eleven others and getting cheaper.” The only company at something has all the protection in the world and didn’t have to build a wall to get it.

Distribution changed roads

Back to the lever a16z loves, because it just did something under our feet and almost nobody has named it. Distribution used to mean one thing: reaching human eyeballs. SEO, the ad auction, the feed, the homepage, the thing you optimized so a person scrolling would land on you. That road is dying. More than one person who used to run a large social platform has said out loud this year what the numbers already show — the ad-supported web is all but dead. Media traffic is down a third as AI summaries answer the question before the human ever clicks. The eyeballs aren’t going where they used to go.

So where’s the new road? Being readable by the machine that now does the looking. When your customer’s agent goes out to shop, to book, to research, to write the report, the question stops being “did a human find you” and becomes “could the agent parse you, use you, choose you.” Agentic readability is the new distribution. It’s SEO for a reader that never scrolls, never sees your banner, and doesn’t care what you paid for placement. Amazon understood this the second it walled its catalog off from Meta’s shopping agent — we wrote that up last week. The same move, run forward: the winners of the next cycle will be the ones the agents can reach, the way the winners of the last cycle were the ones Google could index. Distribution didn’t die. It swapped roads, and the on-ramp is machine-readable.

The oldest lever is the deepest one

Which brings me to the lever a16z leads with and I think is actually the floor under everything else: create a new behavior. Peter Drucker said it fifty years ago — the purpose of a business is to create a customer. Not to serve an existing one better. To make one that didn’t exist. a16z’s whole case for OpenAI rests on this: OpenAI’s real talent isn’t the best model or the best chip, it’s a track record of finding the simple, obvious-in-hindsight thing that makes people do something they’d never done. ChatGPT itself was an accident they almost walked away from until Thiel told Altman: no, this is it, you’ve created a new kind of customer, nothing else matters as much.

Here’s why I think this is the most important sentence in the whole pile, and it’s not a business point, it’s a personal one. Ask it about yourself. What can you do today that you could not do, or would not make time to do, yesterday?

I’ll give you mine. I love taking in information. I love podcasts. And I never had the time, or was never willing to make the time, to get through the ones I wanted. So I used Claude to build myself a system that rips the signal out of a podcast and hands it to me as a written email. It’s eighty percent as good as listening, in eight percent of the time. That’s a trade I’ll take every single day, and I keep adding shows to the list because the behavior is now cheap enough to be a habit. That didn’t exist for me a year ago. Nobody sold it to me. A new behavior got unlocked, and I reorganized a slice of my life around it.

That’s the question worth carrying out of here, whether you’re allocating capital or running a shop or raising a kid: not “is there a better model,” but “what did this just make possible that wasn’t, and am I the road it runs on.” Higgsfield answered it with infinite video. I answered it with podcasts I finally have time for. The scale is different. The move is identical.

The quadrant, with names in the boxes

So draw the two axes — Harry drew them this morning, skeletons and all. One axis: did you unlock a behavior that didn’t exist yesterday, or are you just doing an old chore faster? The other: can only you do it, or can anyone?

Lower left is the graveyard, and it’s crowded. Doing a known task, that anyone can do — that’s general report-writing AI, the spreadsheet-filler, the “we added a chatbot” feature. Switching costs of zero, price racing to the floor, one prompt away from being replaced. Most of what shipped this year lives here and doesn’t know it.

Lower right — a new behavior, but anyone can deliver it — is the fast-follower’s purgatory. You opened a door, and because there’s no wall, the crowd walks through behind you and the price collapses. Nice while the lead lasts.

Upper left — old behavior, but only you — is the legacy moat: a regulated utility, a patent, a thing that’s yours because of history, not because it’s new. Durable, but it’s not going to double in eighteen months.

Upper right is the only square that matters. A behavior that didn’t exist yesterday, and you’re the one who can do it. New and only. That’s Higgsfield. That’s the box every dollar of enterprise value is trying to reach, and almost nobody’s standing in. It’s the promised land precisely because it’s so hard to get to and so lonely once you’re there.

Everywhere else, you’re renting GPUs at eight bucks an hour to sell a token somebody undercuts tomorrow. The upper right is where you rent the same GPUs and charge for something that has no comparison price, because there’s nothing to compare it to.

So, who do you want to be

The models will keep getting better and cheaper. That’s not the game; that’s the weather. DeepSeek discounts the frontier, Sonnet ships on a Tuesday, Opus drops forty percent, and none of it settles who wins, because the thing getting better and cheaper is the commodity, not the moat. The moat is the behavior you unlock and whether anyone else can unlock it too.

Stop asking if your model is better. Ask what your customer can suddenly do — and whether anyone else can let them do it.

Sources:

• Harry Stebbings on Higgsfield — $1B ARR, $4M/mo on models, 150-person content team (X)
• 20VC: $1BN ARR in 18 Months — The Untold Story of Higgsfield, with Alex Mashrabov
• OpenAI Understands Something Important and Rare — a16z
• Akamai Lands $11.6 Billion Anthropic Deal as CPU Inference Defies GPU Consensus — Tech Times
• How GPU Prices Can Double While AI Gets Cheaper — Tomasz Tunguz
• Nvidia unveils security platform to rein in AI agents and $150bn stock buyback — The Guardian
• Claude Sonnet 5.5 announcement — Anthropic (@AnthropicAI, Sep 28)