There's a moment after a demo lands when a customer's technical lead asks how quickly their team can adopt your product. That question makes a harder one concrete: what happens when the next Claude ships? The answer depends on where you sit in the stack, because your layer decides who your buyer is and what you can charge for.

This guide covers how each layer of the stack earns revenue, the four inputs that decide which layer you build on and where your model should come from.

What is a Generative AI Business Model

A generative artificial intelligence (AI) business model links the layer of the stack you sell from with the unit you charge for. It also defines what you own that your model provider doesn't. Your layer sets your buyer and your cost structure. Whether you charge for tokens, seats, resolutions or reserved capacity determines how revenue scales with usage, and what you own determines how dependent the business is on its model provider.

The layer decision comes first because it sets the ceiling on the other two: it narrows which units you can charge for and which costs you carry. Our guide to the AI business model works through those pricing structures and the margin math behind each one.

How Each Layer of the Generative AI Stack Earns Revenue

Every generative AI company sells from one of four layers, each with a different product, buyer and revenue shape. The four run from hardware at the bottom to finished workflows at the top:

• Compute and infrastructure: AI infrastructure companies sell rented graphics processing unit (GPU) capacity and the hardware under everything else. Buyers are AI labs, hyperscalers and enterprises running large models, and revenue comes from hardware sales and rented capacity, a market that passed $25 billion in 2025.
• Foundation models: Model labs sell access to the model itself and price it per token through an application programming interface (API), or per seat on enterprise plans. Roughly 85 percent of Anthropic's revenue comes from enterprise and developer customers.
• Developer tooling and orchestration: These companies sell the retrieval, evaluation, monitoring and deployment layer to engineering teams putting AI into production. Pricing usually follows a hybrid model that pairs a per-seat base with metered usage.
• Applications and vertical products: Vertical AI companies sell finished workflows to a named job or industry and bill per seat, per outcome or through a blend. AI software spending is forecast at $452 billion for 2026.

The ends of this list behave differently for a seed stage team. Capacity and model access can need hundreds of millions to billions of dollars to enter, while tooling and applications sell software a small team can ship in months.

How to Choose the Layer You Build On

Four inputs you already know decide the right layer for your company. They are what you can build that others can't, what data you can reach, how much you plan to raise and who signs your contracts. Most founders start from the layer that excites them and work backwards. Each input rules out at least one layer before you write a line of code.

Your Technical Edge

A technical edge is something you can build that a team with the same model access cannot, and most claimed edges fail that test. Prompt chaining, a retrieval pipeline or better latency gives you a temporary competitive advantage, since a stronger base model can absorb some of what thin products charged for last quarter. Portability tells you which it is: if swapping in another provider's model leaves the product effectively unchanged, the provider is creating what the customer values.

A genuine edge at the application layer usually lives in the loop around the model. Your product does the work inside a customer's process, keeps the record of what happened and handles the next action better for it. Competitors with identical model access can't reproduce that loop without your customers.

Your Data Access

Broadly scraped internet data is the weakest position at the foundation model layer, since every lab draws from the same sources and base models keep improving without domain data. At the application layer, customers create data that competitors cannot reproduce as they work inside the product. Each action refreshes the signal in a way a static dataset can't match.

A useful test before you claim data as your edge is whether a funded competitor could assemble the same dataset within a year. Answering yes turns it into a question for your capital plan.

Your Capital Plan

Most seed stage teams cannot afford to train a foundation model. Training costs have risen rapidly since 2020, and straight-line projections put a 2026 frontier run at roughly $1 billion in compute alone. A $2 million to $15 million round doesn't cover the training bill for a 2023-generation model, let alone a current one. It also doesn't buy a second attempt when the first run disappoints.

The application layer inverts that math, since you pay the same marginal API rate as any late stage company and your raise funds people and distribution instead of compute. Tooling sits between the two, with lower capital requirements than a model lab, but a longer road to revenue that depends on who owns the budget and signs the contract.

Your Buyer

Each layer sells to a different signer, and the signer constrains everything downstream. Infrastructure buying decisions that run through technology leadership commit you to technical diligence before the first dollar arrives. An application purchase owned by a line-of-business leader who has already allocated budget to the problem points your sales process somewhere else entirely.

Budgets are moving as well, and buyers increasingly compare application pricing against what the work costs in salaries rather than against another software line item. Chief financial officers fund projects that promise measurable operating outcomes. The layer you pick decides which of those conversations you'll have for the next five years, and it shapes where your model should come from.

When to Build on a Hosted Model vs. Train Your Own

Hosted APIs get expensive at scale, while premature training freezes you to a checkpoint that takes real work to update. Whether you're an AI wrapper is a separate question about what you own on top of the model. Volume, data and the dependency you're willing to carry decide which of the three routes below fits.

When to Build on a Hosted Model

A generative AI startup should begin with a hosted API and keep it while volume remains modest and the task benefits from frontier-class quality. Prompt engineering and retrieval will establish what the hosted model can deliver before you spend anything on fine-tuning. The route trades upfront flexibility for provider dependency, since you inherit the provider's pricing and model-lifecycle decisions and behavior may change from one month to the next. Teams that build for provider portability from the start keep that dependency cheap to unwind.

When to Fine-Tune an Open Model

Fine-tuning an open-weight model becomes practical when you have a narrow, closed-vocabulary task with thousands of labeled examples, a stable output schema and steady daily volume. You gain more consistent performance on a specific task at a lower inference cost. Ongoing model operations come with it, since someone has to run deployments and evaluations. Adopting a newer base model also means repeating the tuning and validation work.

When to Train Your Own Model

Training your own model makes sense in two situations: you hold a data asset no base model covers, or you're a funded lab whose product is the model itself. Both put you at the foundation model layer. Bloomberg trained its own finance model on decades of proprietary financial data no base model covered, and that dataset justified the run.

CRV-backed Reflection AI makes the second profile concrete. Its flagship model had not shipped when the company raised $2 billion in an open frontier model round led by Nvidia in October 2025. A compute agreement with SpaceX worth up to $6.3 billion followed, at $150 million per month. If your capital plan doesn't resemble that, your model should come from someone else.

What Generated Output Changes About Selling Software

Whichever layer you choose, what you deliver is generated output, and that breaks four assumptions traditional software contracts rest on. Each one surfaces first in a redline or a renewal conversation:

• Output quality varies per request: Traditional warranties describe repeatable software behavior, while generated artifacts are probabilistic and identical prompts may not return identical results. Buyers respond by negotiating accuracy-based service level agreements, which pushes sellers to name hallucination and drift in their terms.
• Rights and provenance attach to what the model produces: Copyright for generated output depends on human-authored expression appearing in the result, and prompting by itself does not establish copyright. Before signing, enterprise buyers examine indemnity exposure, whether either party may reproduce outputs for others and which party receives whatever rights exist.
• Customers can regenerate instead of renewing: A customer who can get the raw model's answer cancels the wrapper, which puts customer churn pressure inside the product rather than out in the market.
• A model release can reprice your product overnight: Chegg's subscribers fell 31 percent year over year to 3.2 million in the first quarter of 2025. The company cut 22 percent of its workforce after free AI tools replaced what students had been paying for.

These constraints hit hardest where the product sits closest to raw generation. Owning the workflow around the model gives sellers more control over quality and contract terms, and it protects renewal value.

How Investors Test a Generative AI Business Model at Seed and Series A

Investors meeting a pre-revenue team start by asking whether the founders chose their layer deliberately and can say why. They then test whether the team has modeled infrastructure cost per customer before it shows up in gross margin, since inference-driven gross margins carry a cost traditional software never did. Expectations keep rising too, and the revenue most teams now show at a Series A sits well above what the 2021 cohort needed.

The question underneath all of it is what you own that survives the next model release. Investors fund teams that answer with a first-party feedback loop, an operational dataset or a signed customer base. Teams that answer with a feature list are describing the model provider's roadmap. We've written separately about what makes a company AI native, and the layer question is where that identity starts.

What the Strongest Generative AI Business Models Have in Common

The strongest generative AI business models we see combine deliberate choices across the business. Founders in that group picked their layer from what they already had rather than from what the prototype could do, and model sourcing followed the economics, hosted until volume and data justified otherwise.

CRV invests at the seed and Series A stages, and we've watched mobile, cloud and AI reset which layers were worth owning. Our view is that founders underweight the layer choice relative to how much it constrains everything downstream. If you're an early stage founder looking for support pressure-testing your layer choice and compute economics, reach out to us to see if we'd be a good fit.

Frequently Asked Questions About Generative AI Business Model

How do generative AI companies make money?

Generative AI companies charge for one of four units: reserved capacity, tokens consumed, seats or completed outcomes. The unit follows from where a company sits in the stack, which is why the layer decision sets the revenue model rather than the reverse. Most companies end up blending two of the four, usually a seat base with metered usage on top.

Which layer of the generative AI stack is most profitable?

No layer is reliably more profitable than the others. Profitability tracks how much control you have over cost per request and how much capital you had to commit to serve it. A company that can cut cost per request without degrading output keeps its margin as it grows, while a company that cannot, is exposed every time usage rises.

Can a startup compete with foundation model providers?

A startup can compete without matching a model lab's spending. Application companies can build on the labs' own models and compete through deeper workflows and distribution. Accumulated customer context adds another advantage. Legal and coding products compete on that architecture while leaving frontier-model spending to the labs.

Do you need proprietary data to build a generative AI business?

Proprietary data helps only when it's the kind a competitor can't assemble. Operational signals from customers' live workflows and labeled outcomes from real decisions are two examples, and context accumulates the longer the product runs. Scraped or purchasable data confers little advantage, since every competitor draws from the same sources. Plenty of strong AI businesses started with no proprietary data at all and built it through the product.