None of today's foundation model companies have a revenue model that works. Training runs cost billions and depreciate in months. Raw intelligence sold by the token is a commodity, its price collapses with every release, and open-weight models keep resetting the floor toward zero. Subscriptions cap the upside of a technology that can do entire jobs. Even OpenAI makes most of its money from ChatGPT the app, not from selling intelligence through the API. The labs built the most valuable capability in history and they're monetizing it like a utility.
I don't think this means foundation models are a bad business. I think general intelligence is the wrong layer to monetize. The revenue model won't be solved once at the bottom of the stack. It gets solved thousands of times at the top, domain by domain, by companies that price the model's output the way their industry already prices outcomes.
Every industry monetizes outcomes differently
Advertising pays for performance. Sales pays commission on closed revenue. Law pays contingency on resolved disputes. Insurance pays for underwriting accuracy. Logistics pays per delivered package. None of these industries wants to buy tokens. They want to buy the outcome, and they already have decades-old machinery for pricing it. Each domain application of the foundation model will discover the revenue model native to its industry.
Outcome-based pricing and the training loop are the same apparatus. The number you bill on is the number you train on. If you charge for conversions, every conversion is both revenue and ground truth. Seat-based pricing doesn't just leave money on the table; it blinds you to your own training signal. The companies that figure out their industry's outcome model won't just monetize better than the general-purpose labs. They'll learn faster too.
Model developments eat the harness
The tempting conclusion is that the winners are the application layer: wrap the best general model in a harness of scaffolding and workflow and sell the result. The last three years ran that experiment, and the verdict is consistent: model developments eat the harness. Long context ate the retrieval pipelines. Native tool use ate the orchestration frameworks. Reasoning models ate the elaborate prompt chains. Agents ate the workflow builders. Every capability you build around the model is a bullet point on the next model's release notes. A harness is a patch on the current generation's weaknesses, with a lifespan of one release cycle.
Everything on that list was built from public knowledge, techniques anyone could read and eventually bake into the weights. The next release absorbs what it can see, and it cannot absorb what it has never seen.
The most successful harness company figured this out. Cursor built the best wrapper around other people's coding models, watched each release absorb more of the job, then went and trained its own frontier coding models on the data its editor generates. The harness was the beginning, not the moat.
The flywheel only spins if you own both ends
What compounds while everything else depreciates is proprietary data, and the only place to get it is your own application. Every customer interaction is training signal that exists nowhere else: what they accepted, what they rejected, whether the outcome landed. Data feeds model, model improves product, product wins customers, customers generate data. The flywheel only spins if you own both ends, the app that captures the signal and the model that learns from it. X feeds Grok. Kimi's app feeds Kimi's models. ChatGPT feeds GPT. Break the loop at either end and you're a distribution channel for someone else's intelligence.
The race from both ends of the stack
So a two-way race is on. The labs are racing up the stack, shipping agents and applications and interfaces, because applications are where the revenue models live and they don't have one. The application companies are racing down the stack, building post-training and then pre-training, because models are where the moat is and they don't have one. Each side is trying to grow the organ the other was born with.
My bet is on the application companies, and the reasons are structural. They're closest to the customer, so they see the outcome first and keep seeing it. They own the proprietary data because it's generated inside their product. And they already have the revenue model, the thing the labs are spending billions trying to find. A lab starts with a brain and has to earn a body in a thousand industries at once, against incumbents who know every buyer, workflow, and pricing convention. An application company has to build a brain in one domain, fed by data it already owns, funded by revenue it already collects. Distribution can be bought. Ground truth cannot.
An advantage isn't an outcome, though. Some application companies will hire the researchers, wire their outcomes into reward, and cross over into foundation model companies. Others will keep renting cognition while the margin migrates to whoever owns the weights. It's natural selection running on companies, and the transition is the filter.
Which strips the strategic question down to a single moat. Model weights depreciate and the app alone gets eaten, so the defensible position is the combination: a team that can keep building state-of-the-art foundation models for your domain, plus the distribution and data acquisition that come from owning the application your customers live in. The team keeps the brain at the frontier. The app keeps feeding it experience no lab will ever observe. The scarcest hire of the next decade isn't another application engineer. It's the team that can train.
Every animal has a brain
I keep coming back to biology, because biology settled this argument a long time ago. Every animal has a brain and a body, and neither survives without the other. The body is how an organism acquires data through its senses and energy through its metabolism. The brain turns that data into decisions that keep the energy coming. In four billion years, evolution never shipped an animal that outsourced its brain, because an organism that rents its cognition is not an organism. It is prey.
Companies are becoming organisms in this sense. The app is the body, the sensory surface that acquires proprietary data and the metabolism that converts outcomes into revenue, the energy that funds the next training run. The model is the brain, and it has to be yours, shaped by experiences only your body has had.
In the future, every company will be a foundation model company. Not every company will train a trillion-parameter general model, but every enduring company will train the state-of-the-art model for the domain it competes in, fed by data only its app can capture, monetized through outcomes its industry already knows how to price. An underwriting brain that has watched a hundred million claims resolve. A logistics brain that has watched a hundred million deliveries land on time or late. An advertising brain that has watched a hundred million creatives run and knows which frame lost the customer. None of them general, all of them unbeatable in their domain, because the general model has never seen what they've seen. That, finally, is the revenue model that works, and it belongs to the domain companies rather than the general labs. The labs will keep pushing the frontier of raw intelligence. Everyone building on top of it faces the choice every organism faces. Grow a brain, or become food for one.
—
This is the company we're building at Omneky, for advertising. The agent generates the creative, launches the campaigns, and buys the media, and every impression, click, and conversion flows back as ground truth into an advertising foundation model trained on outcomes pooled across every brand we serve. No single advertiser, and no general-purpose lab, could spin that flywheel alone. Advertising already prices outcomes. It's as clean a training signal, and as clean a revenue model, as any industry gets.