The applied advantage

Why post-training, continual learning, and domain expertise beat raw frontier scale.

The most valuable AI companies of this cycle will not be the ones that trained the biggest model. They will be the ones that took a capable model - anyone's model - and made it genuinely good at one kind of work.

This is a stronger claim than it looks. The frontier labs have the talent, the compute, and the brand, and their models will keep getting better and cheaper. Our view is that this is exactly why the durable value moves down the stack: to the companies that post-train general models on proprietary data, measure them against expert-defined tests, and keep improving them inside a specific domain. We call this the applied advantage, and we think it compounds in a way raw capability does not.

The application layer is already the biggest layer

The market has voted. Menlo Ventures estimates that companies spent $37 billion on generative AI in 2025, up from $11.5 billion the year before. More than half of that - $19 billion - went to the application layer: the products and software built on top of the models, not the models themselves. That is more than 6% of the entire software market, reached within three years of ChatGPT's launch.

The capital going into foundation models dwarfs the capital going into applications. The revenue does not. Enterprises are paying for work that gets done, and the work lives in the application layer.

Nobody loves a lab

Inside the enterprise, the model is already a swappable input. Menlo's mid-2025 market data shows how fast loyalty erodes: OpenAI's share of enterprise model usage fell from 50% in 2023 to 25%, while Anthropic rose to 32% and Google took 20%. Ramp's spending data tells the same story from the other side - businesses run multiple model providers at once and switch freely between them.

This is rational behavior, and it has a sharp implication. If the model is replaceable, the value sits in whatever is not replaceable: the workflow the product owns, the production data only it sees, and the tests it uses to know whether an answer was good. A company whose product is a thin layer over someone else's model owns none of those. A company that post-trains, evaluates, and improves inside a domain owns all three.

Post-training is becoming the product

Pretraining belongs to the labs. Post-training - fine-tuning and reinforcement learning on the data a product generates in real use - belongs to everyone, and it is getting cheap and accessible fast. Thinking Machines, founded by former OpenAI researchers, now sells a training API built on a simple observation: on specialized tasks, teams can train smaller models on their own production traces and beat large general models.

The legal platform Harvey shows where this leads. Harvey started as an application on top of frontier models. It has now released its first post-trained open-weight model alongside a legal agent benchmark built from real legal work. Read that again: a vertical software company now trains its own models and publishes its own quality tests, because the tests are the product. The expert-defined eval - what does a good contract review actually look like, judged by lawyers - is the hill every post-training effort climbs. The organizations with deep domain expertise are the ones that can define it. Everyone else is guessing.

Continual learning turns usage into ownership

Today's models are mostly frozen at deployment. That is changing, and the direction matters more than the current state of the research: models will increasingly learn continuously from the environments they work in. When that works, every completed task becomes training signal, and the product gets better with use in a way a competitor cannot copy by buying the same base model.

But the compounding only accrues to whoever owns the learning loop. If the improvement happens inside the model provider's stack, the enterprise is renting intelligence that gets smarter for everyone, including its competitors. If the improvement happens inside the enterprise's own deployment - its weights, its evals, its traces - the intelligence is an asset on its own side of the boundary. Continual learning is why "which model did you buy" is the wrong question and "where does your AI learn" is the right one.

The evidence from the field

The companies winning in vertical AI share a shape. Harvey serves more than 3,000 customers in over 70 countries and crossed $400 million in annual recurring revenue by going deep on legal work rather than wide on general assistance. In healthcare, Abridge topped a $5 billion valuation building ambient clinical documentation for hundreds of health systems, and OpenEvidence was valued at $15 billion on the strength of a medical answer engine reportedly used by around 40% of US physicians.

None of these companies trained a frontier model. All of them own a workflow, a corpus of domain data, and the trust of experts who check the answers. Their moat is not that they use AI. It is that their AI gets measurably better at one profession's work faster than a general product can follow.

Enterprises want sovereignty, and they should

The buyers have figured this out. An IBM study in June 2026 found that 68% of executives struggle to meet data residency and sovereignty requirements, and 91% say they do not fully understand their AI dependencies across vendors, models, and infrastructure. Enterprises have learned the risks of building critical work on a provider they cannot audit: a model version retires, prices change, terms change, and the capability they depended on is gone or different.

AI sovereignty, properly understood, is not about nationalism or running everything on-premise. It is about whether your AI answers to you: your data stays yours, the evals are defined by your experts, the fine-tuned weights sit in your boundary, and no single vendor can degrade your capability by changing a pricing page.

Poolside is the cautionary tale that proves the rule. The company bet from day one that enterprises and governments would pay for frontier models deployed inside their own walls, fine-tuned on their own code. The bet was right. The buyers came - defense departments, industrial giants. What killed the company was not the thesis but the capital structure: staying in the frontier training race required sums it could not raise. Its investor letter admits it had a six-week window to raise $2 billion for a cluster, missed it, and lost the ability to compete at the frontier. The company ended in a $6 billion licensing deal with Nvidia - the chokepoint vendor buying the capability outright.

The lesson is not that sovereign AI fails. It is that sovereignty cannot mean outspending the labs. The affordable version is the applied one: take strong open or licensed models, post-train them on your own data, evaluate them with your own experts, and keep the improvement. That path gets cheaper every quarter, while frontier pretraining gets more expensive every quarter.

What we look for

The durable applied AI companies we see share four traits. Domain expertise in the building, not just in the sales deck - the people who define what good work looks like are inside the company. An eval advantage - proprietary tests of real task quality that general benchmarks cannot replicate. A data flywheel - production usage that legally and technically improves the product. And model independence - the ability to route across providers or post-train open weights, so no lab's roadmap is a single point of failure.

We are more skeptical of two shapes: thin wrappers with no proprietary eval or data loop, whose margin is a prompt template; and new pure-play model labs without frontier-scale capital, which is the exact position that squeezed Poolside out of the race.

What could make this wrong

The labs are not standing still. If OpenAI, Anthropic, or Google bundle genuinely good vertical capabilities into their platforms at near-zero marginal cost, the procurement math for standalone vertical products changes overnight. A large capability jump could make general models good enough at specialized work that post-training stops mattering. The labs could also own the post-training tooling layer itself and tax the flywheel from below. And continual learning may be solved centrally, inside the frontier model, rather than at the edge.

Each of these is a live risk. But notice what they have in common: they all describe the labs absorbing the value of work they do not currently understand - the workflows, the evals, the expert judgment. That understanding is exactly what the applied companies own, and it is not for sale in any model release.

The labs will keep winning the race to the frontier. The frontier, increasingly, is not where the money is.