Who is building the AI stack — and where the value is moving

“AI company” now describes businesses operating at very different layers of the technology stack. Understanding those layers makes product news easier to interpret.

Compute and infrastructure

At the base are chips, data centers, networking and cloud infrastructure. Training and serving capable models requires substantial compute, making infrastructure economics part of the AI story.

Foundation models

Model developers turn compute and data into general-purpose systems exposed through consumer products and APIs. Competition here involves capability, efficiency, multimodality, context, tools, safety and distribution.

Developer platforms

A growing layer helps teams connect models to data, evaluate outputs, observe agents, manage prompts and control costs. Some of this functionality may be absorbed by model providers; other parts can remain valuable precisely because they work across providers.

Applications

Applications translate general model capability into a specific job: coding, design, support, research, sales, media or operations. Their defensibility may come from workflow, proprietary data, distribution, trust and integration rather than exclusive access to a model.

Why the map keeps changing

AI layers overlap. Model companies add applications, application companies train models, and cloud providers offer both infrastructure and model access. CortexLab follows these boundary shifts because they often explain why products converge, prices move and once-distinct categories disappear.

Developer infrastructure

Between models and end-user applications sits a growing layer of infrastructure: inference platforms, model gateways, vector and retrieval systems, observability, evaluation, security and agent tooling. These products help teams turn a capable model into a reliable service.

This layer can become more important as raw model capability becomes easier to access from multiple providers. Operational control, data connections and evaluation may be where a company creates durable differentiation.

Applications own the workflow

At the application layer, value comes from solving a specific job: coding, design, customer support, research, sales, media production or vertical professional work. The underlying model matters, but distribution, proprietary context, interface design and integration can matter just as much.

Why the layers keep moving

AI companies frequently expand up or down the stack. Model providers ship applications. Cloud platforms offer models and agent tools. Application companies train or customize models. The boundaries are strategic rather than permanent.

For readers and buyers, the useful question is therefore not only who has the strongest technology today. It is which layer controls the customer relationship, the workflow, the data and the economics as the stack evolves.

Follow the control points

Value tends to gather around scarce resources and durable control points: compute capacity, differentiated models, proprietary data, distribution, workflow ownership and trusted customer relationships. Which point matters most can change as technology becomes cheaper or standardized.

For example, if capable models become easier to substitute, the application that owns the workflow may gain leverage. If inference remains constrained or specialized, infrastructure economics can dominate.

Watch vertical integration

Companies increasingly cross stack boundaries. Model providers build consumer and enterprise applications; cloud platforms add model marketplaces and agent tooling; application companies train specialized models or operate their own inference. These moves can improve products while also changing dependencies for partners.

Use the stack as a research map

When evaluating a company, ask which layer it primarily serves, which suppliers it depends on, what it uniquely controls and how easily customers can switch. That framework is more durable than treating every AI announcement as an isolated event.