Why Vertical AI Wins

In June 2026, Cursor sold to SpaceX for $60 billion in stock. The number that made headlines was the price. The number that mattered was 41 to 26.

That was Cursor's market share collapse between June 2025 and May 2026. Eleven months, fifteen points of share gone, with Anthropic taking half the category through Claude Code. The acquisition was a smart exit. It was not a victory.

The structural lesson is not about Cursor. It is about what happens to companies that build on top of foundation models that are also their competitors.

The compute ceiling

When you build on foundation models, your dependency goes deeper than infrastructure. The labs sell intelligence. They also build products on top of that same intelligence. Every API call you make funds someone who can rebuild your product as a feature inside their next release.

This creates a ceiling that most builders do not see until they are already inside it.

• Your margin is their pricing decision.
• Your roadmap is their next release.
• Your differentiation is their next capability.

The economic logic only works in one direction. The lab gets cheaper inference, better models, and broader distribution every quarter. You get the privilege of paying for it.

There are only three exits from the compute ceiling.

• The first is to become a model lab yourself. This requires capital that almost no one has.
• The second is to be acquired by one. Cursor went to SpaceX. Windsurf went to OpenAI. Both were sold from the position of a company that had run out of room.
• The third is to build a workflow moat in a domain the labs will not enter.

Option three is the whole game.

Where the labs will and will not go

The model labs are not technically incapable of building vertical workflow products. They have the engineering talent and the capital to build a wealth management platform or a clinical operations system. They choose not to.

The reason is strategic. Their business is selling horizontal intelligence at scale to every industry. Building a vertical workflow product means competing with their own customers. It means hiring domain experts they do not have. It means taking on regulatory risk they have not modeled. It means getting distracted from the model race they are trying to win.

So the labs will sell intelligence into healthcare. They will sell it into financial services, legal, and insurance. They will not build the products that do the actual work inside those industries.

That gap is where durable companies get built.

Why specificity beats scale

For most of the software era, the playbook was horizontal. Build a tool. Acquire users. Expand across industries. Scale through distribution.

The AI era changes the equation.

In a world where intelligence is becoming a commodity, the differentiation cannot come from access to the smartest model. It comes from access to the deepest understanding of how work actually gets done.

A wealth advisor's workflow is not abstract. It involves portfolio construction, tax implications, private market access, fiduciary duty, accredited investor regulations, and behavioral patterns that take years to learn. A horizontal AI cannot replicate that knowledge by being trained on more data. It can only be built by people who lived inside the workflow.

The same is true of insurance underwriting, clinical operations, legal review, and dozens of other complex domains. Each one has its own data, its own rules, its own decision logic, and its own definition of what acceptable failure looks like. The labs are not going to build for these domains because the cost of going deep is too high relative to the return on horizontal capability gains.

This is why I keep returning to the same conclusion. The companies that will define the next decade of AI will not be the ones with the best models. They will be the ones that own the workflow.

The 24-month commoditization window

Vertical AI in domains the labs care about does not have unlimited runway. Coding tools, general productivity, customer support, sales automation, content generation — these are categories the labs are actively building toward. The companies in these spaces have somewhere between eighteen and twenty-four months of differentiation before the next model release closes the gap.

Cursor lost fifteen points of market share in eleven months once Claude Code matured. That is the cadence to plan around if you are building in a space the labs want to enter.

The inverse is also true. Regulated, narrow, complex domains where the labs will not build the workflow have much longer runway. The same reason the labs avoid these spaces is the reason durable companies can be built there. Compliance complexity creates moats. Domain expertise creates moats. Trust relationships with regulated buyers create moats.

• Wealth management.
• Private investing.
• Financial services workflows.
• AI-native fintech.
• Anywhere the workflow is the moat, not the model.

What workflow ownership actually means

Owning a workflow is more than building a vertical product. It means controlling the data layer that captures how the work gets done. It means owning the integrations into the systems where the workflow already lives. It means earning trust from the operators who execute the workflow today. It means absorbing compliance and regulatory risk in a way the foundation models cannot.

The companies that get this right end up with three things the labs cannot replicate.

The first is proprietary workflow data. Every cycle through the product produces structured information about what worked, what failed, and what the experts did to recover. That data is not in the labs' training sets. It cannot be generated synthetically. It compounds with every customer.

The second is distribution into regulated buyers. Healthcare systems, financial firms, law firms, and insurance carriers do not buy AI the way startups buy SaaS. They buy through trusted operator networks, references from peer institutions, and design partnerships that take months to mature. That distribution is a moat the labs are not built to compete in.

The third is workflow lock-in. A vertical AI product that has integrated into a wealth advisor's day-to-day decision-making, with audit trails, compliance review, and customer trust, is not easily ripped out. The switching cost is not the software. It is the workflow itself.

The economics have changed

There is one more shift worth naming.

For most of the software era, building meaningful companies required large teams. The cost of engineering, distribution, and operations meant that you needed significant headcount before you could approach significant revenue.

AI is changing that economics. The leverage available to small teams in 2026 is dramatically higher than it was in 2020. Engineers can ship faster. Founders can sell faster. Operations can scale faster. Workflows that used to require teams of fifty can now be run by teams of five.

This means the next generation of category-defining AI companies will be built by smaller teams than any previous software era. The capital intensity drops. The team intensity drops. What does not drop is the requirement for domain depth.

In the SaaS era, you could win with average product and great distribution. In the AI era, you cannot. The product has to actually do the work, and doing the work in a regulated industry requires people who know the work.

The conclusion

The Cursor exit was a smart move by a company that hit a ceiling. The lesson is not that vertical AI failed. The lesson is that vertical AI in domains the labs care about has a short window, and vertical AI in domains the labs will not enter has a long one.

The durable AI companies of the next decade will be built by operators who lived inside the workflows they are now rebuilding. They will own data, distribution, and trust in markets where horizontal intelligence is necessary but not sufficient.

This is what 434 backs. Not AI wrappers. Not features. Companies built by people who understand a workflow at the level of detail you can only get by doing the work.

Anywhere the workflow is the moat, not the model.