What Does an AI Company's Position in the Stack Tell You?

Most people meet large language models as products: a chat window, a coding assistant, a search box that answers in paragraphs. Behind each of those products sits an industry with its own supply chains, cost structures, labor markets and power struggles. Anyone who wants to work in that industry, advise it or compete with it needs a map of how the pieces fit together.

The most useful map has four layers. Infrastructure, models, platforms and applications form a stack, and each layer runs on its own economics, attracts its own kind of company and rewards its own kind of talent. Position in that stack is the first fact worth knowing about any AI company. It predicts margins, rivals and the shape of the jobs inside far more reliably than the brand name on the door.

That map rests on a handful of connected ideas: why the industry exists at all, how its layers make money, which advantages last, and what the vocabulary is for. Taken together, they turn a noisy field into something a newcomer can reason about.

Generality created the industry

Earlier waves of machine learning produced systems built for one task each: a model to flag suspicious transactions, another to rank search results, another to sort images. Each system served one buyer with one problem. Large language models changed the unit of value. A single model can write code, summarize a document, analyze a table and hold a conversation.

This generality is the economic engine. When one model serves thousands of uses, the cost of building it can be spread across customers in every sector, and a whole ecosystem forms around supplying, packaging and applying it. The pattern resembles the rise of the internet more closely than any single product launch: a general-purpose layer appears, and entire categories of business reorganize around it.

Generality also explains why the industry stratified so quickly. A broadly capable model needs enormous computing power beneath it and specific products above it. The layers of the stack are the natural division of labor that forms around one general technology.

Four layers, four kinds of business

At the base sits infrastructure: chips, data centers and cloud capacity. NVIDIA, AMD and Intel supply processors, custom silicon such as TPU and Trainium competes alongside them, and cloud providers rent the resulting capacity to everyone above. This layer is capital-heavy and physical, and its constraints run through manufacturing and supply chains as much as through software.

Above it sit the models. Frontier labs such as OpenAI, Anthropic, Google DeepMind, Meta AI, Mistral and xAI train large foundation models. Open-source labs release model weights that others can download, run and modify. Fine-tuning shops adapt general models to narrower purposes.

The platform layer connects models to the people who build with them. Companies such as LangChain, Hugging Face, Weights & Biases and Replicate supply frameworks, model hubs and developer tooling. At the top, applications turn capability into products for a specific user, from vertical AI software companies such as Harvey, Cursor and Perplexity to established firms adding AI features to products their customers already pay for.

Each layer earns money in its own way. Infrastructure sells capacity, and model labs sell access, usually priced by usage. Platforms sell convenience and speed to developers. Applications sell outcomes to end users.

A company's layer tells an observer which of these games it is playing and whose costs flow into its own.

Training and inference run on separate economics

Inside the model layer, two activities carry very different cost profiles. Training is the up-front cost of building: a frontier training run costs $100M or more, paid before the first customer sends a request. Inference is the ongoing cost of serving: every request consumes compute, measured in tokens, and the provider earns back its investment one response at a time.

The two activities reward different strengths. Training rewards access to capital, compute and scarce research talent. Inference rewards efficiency: hardware that serves each request cheaply, fast responses, and routing each task to a model of the right size. Many companies win at one and lose at the other.

This split gives anyone evaluating an AI business two separate questions. Who can afford to build the model, and who can afford to serve it at a price customers will pay? The second question shapes pricing per token, the margins of every application built on top, and much of the competition between labs once their models reach similar capability.

Six moats decide which leads last

Model capability spreads quickly across the industry, a dynamic known as model commoditization. Durable advantage comes from six other sources, and each can be checked company by company.

Distribution is reach into the places customers already work. The partnership between Microsoft and OpenAI is the defining case: a model lab gained a route to enterprise customers through a company whose software those customers already use. Data is the second moat, and proprietary data held by one company is becoming a central competitive edge. Talent is the third: the frontier is set by a small community of researchers, a few hundred people, and labs compete hard for them.

Compute is the fourth, since whoever controls GPUs controls the pace of training and the cost of serving. Brand and trust form the fifth, because enterprise buyers care which model sits behind the products they deploy. Ecosystem lock-in is the sixth: APIs, toolchains and developer habits that make switching costly.

Knowing which of these six a company holds, and which it lacks, tells an observer whether its lead is durable or a head start about to evaporate. The exercise works at every layer of the stack, from chipmakers to application startups.

Vocabulary as the price of entry

Parameters, tokens, context windows, inference, RAG, agents, RLHF, alignment: each term names a mechanism with consequences for cost, product design or risk. Parameters are the learned numbers inside a model.

A context window sets how much material a model can consider at once. Retrieval-augmented generation lets a model draw on documents supplied at the moment of a request. Reinforcement learning from human feedback tunes a model's behavior using human judgments of its answers.

Fluency in these words is what lets a newcomer follow an argument about strategy. A sentence about "lowering inference cost per token on a fine-tuned open model" carries a full business claim: about margins, about which layer captures value, and about which moat a company is relying on. Clear reasoning about AI starts with clear vocabulary, and the terms are learnable by anyone willing to study them.

Choosing a position in the industry

The same map that locates companies also locates careers. "Getting into AI" resolves into three choices: which layer of the stack, which function, and which company stage. The functions span research, engineering, product, go-to-market and policy. The stages run from AI-native startups, where the model is the product, to AI-integrated incumbents, where AI extends an existing business and customer base.

Each combination describes a different working life. A policy role at a frontier lab, a solutions architect at a cloud provider and a product manager at an application startup all count as "working in AI", and each draws on different skills and offers a different pace. Professionals arriving from law, marketing or banking bring domain knowledge that becomes most valuable once they can place it on the map.

The numbers that describe this industry move fast. Benchmarks, market size, enterprise adoption and hiring data all shift from year to year, so the durable skill is the habit of checking current figures against the structure beneath them.

That structure is the lasting asset. Locate a company in the stack, and ask how its layer makes money and whether it lives on training or inference. Check which of the six moats it holds, and read the vocabulary for the business claim inside it.

Position in the stack is the first question, and with that question in hand, a fast-moving industry becomes something a careful newcomer can read, compare and choose a place in.