The AI layer and platform dependence

Dependency on LLMs is Death

One of the most useful ways to think about where AI startups are headed is to go backward for a moment.

I remember clearly when Dropbox decided to move off third-party cloud infrastructure and start running its own servers. At the time, a lot of very smart people in Silicon Valley said Dropbox was making a mistake. Their argument was that Dropbox wasn’t really a storage company; it was a UX company. Storage was a commodity. The interface was the value.That argument never made sense to me.

Files still have to live somewhere. And if that “somewhere” is owned by a third party, like Amazon Web Services, then that third party controls pricing, terms, access, and ultimately your margins. They can change the rules whenever they want and even compete with you down the road.

Dropbox understood that early. Storage wasn’t a feature. It was core to the product. And that decision is a big reason Dropbox is still around despite pressure from Google Drive and Microsoft OneDrive. They weren’t just defending a UI. They were defending control.

That same lesson is repeating itself now, just at a different layer of the stack.

The LLM Platform Trap

I think in 2026 we’re going to see a lot of AI-first startups struggle, not because AI demand slows down, but because too many of them are built on top of platforms they don’t control.

Take a hypothetical construction software company. General contractors love it. It automates billing, compliance, and change orders. Real value. But if 90 percent of that value is coming from calls to OpenAI, Gemini, or another LLM provider, then who actually controls the business?

The LLM provider controls pricing. They control performance. They control output quality. They control whether features stay available. And at some point, they can decide they don’t need your application at all because they can offer the same functionality directly.

You didn’t just build a product. You trained the platform that might replace you.

This is why VCs are now asking a harder question than “is this AI-powered?” They’re asking: if the underlying platform changes its economics or strategy, what happens to you?

Brand and distribution help. Having tens of thousands of users helps. But those things don’t always protect you when the intelligence layer you depend on is owned by someone else.

Flipping the Pyramid

What worked five years ago doesn’t work anymore. Back then, companies built the LLM first and trusted that value would emerge later. That made sense when the model itself was the breakthrough. Today, that playbook is mostly gone.

If you’re building now, you have to flip the pyramid. You build value first. You own a workflow deeply. You control data that’s proprietary and hard to replicate. You become embedded in how someone actually runs their business. Only after that do you start pulling more of the intelligence stack in-house.

Yes, that eventually means training or owning more of your own models. That’s expensive. But capital follows value. If you’ve proven that you control something essential, investors will fund the infrastructure required to defend it. Anything else is just renting intelligence.

The AI Layer

This brings me to something I’m hearing more and more in conversations with operators and advisors: the idea of owning your AI layer.

For years, companies outsourced core systems to platforms like Salesforce or SAP. Those were systems of record. They stored data. They ran workflows. That was fine. AI is different.

Once AI starts influencing pricing, underwriting, customer interactions, forecasting, or operations, it stops being a passive system and starts becoming the way your company thinks. That’s not something you want living entirely inside someone else’s platform.

You can use an LLM provider for compute and modeling power. That’s infrastructure. That should be swappable. What matters is that the intelligence layer—the workflows, logic, prompts, proprietary data, and decision frameworks, lives inside your organization. You should be able to change LLM providers without rewriting your business.

If switching models breaks your core workflows, then you don’t own your AI layer. Someone else does. This is the same realization Dropbox had years ago. Storage wasn’t just a backend detail. It was mission-critical. So they owned it.

AI is moving into that same category now. Not as a feature or plugin. But as core infrastructure. And the companies that survive the next wave won’t be the ones with the flashiest demos. They’ll be the ones that decided early what they had to own and built their businesses around defending it.