The economics of generative AI are unusual in two ways worth understanding before any figure makes sense. Training costs are concentrated at a scale very few organisations can absorb, and consumer revenue is concentrated in a handful of applications. This page covers spend growth, consumer application revenue, the cost structure behind frontier models, and where value sits in the category. Adoption data is covered separately, because adoption and economics tell different stories.
Enterprise Spend Growth
Spending has grown faster than deployment, which is the central tension in the category’s economics.
The Spend Figure
Menlo Ventures estimated enterprise generative AI spend at approximately $37 billion in 2025, roughly triple the prior year.
Why Tripling Matters
Few technology categories triple annual spend in a single year. That rate reflects experimentation budgets rather than settled operating expenditure, which makes it less durable than it appears.
Spend Running Ahead of Deployment
The spend figure sits well ahead of measured deployment, which is documented separately in our AI adoption data. Our AI consulting services engagements plan against deployment rather than spend.
What This Implies for Pricing
Category pricing is being set during a period of unusual budget availability. Expect pressure as budgets normalise.
Consumer Application Revenue
The consumer side is where generative AI revenue is most measurable, and where concentration is most extreme.
AI App Revenue
AI applications generated $18.5 billion in revenue in 2025, growing 273% year over year.
Concentration in One Product
ChatGPT was responsible for 43% of that revenue, which means a single application accounts for close to half the category’s consumer earnings.
Downloads Versus Revenue
Generative AI apps reached 1.7 billion downloads and $1.9 billion in revenue in the first half of 2025, a ratio that indicates most usage is free.
Reach
Close to 700 million people used AI apps in the first half of 2025.
The Monetisation Question
Large reach with modest paid conversion is the defining challenge. Our mobile app development work treats AI features as retention drivers rather than direct revenue in most consumer categories.
The Cost Structure Behind Frontier Models
The training economics explain why the competitive field is narrow and why it is unlikely to widen.
Training Run Costs
A single frontier training run costs hundreds of millions, which restricts the field to organisations with either substantial capital or a major platform backing them.
Value Concentration
OpenAI and Anthropic together account for the large majority of total valuation in this category, reflecting those economics directly.
Why Compute Suppliers Matter
NVIDIA supplies the compute nearly every model developer runs on, which makes it structurally central to the category’s cost base regardless of which lab leads on capability.
The Inference Cost Distinction
Training is capital expenditure concentrated among few players, while inference is operating expenditure borne by everyone building on models. Our generative AI development work optimises the second.
Open-Weight Economics
The alternative economic model in the category, where the cost structure shifts from per-token pricing to infrastructure.
What Open Weights Change
Self-hosting replaces usage-based pricing with infrastructure and operations cost, which changes the economics substantially at high volume and worsens them at low volume.
The European Position
Mistral occupies a distinct position as both a frontier developer and the leading European AI company, with EU regulatory requirements potentially creating a structural advantage for European data sovereignty requirements.
The Licensing Caveat
Meta’s Llama is widely described as open source, and its licence restricts EU usage and includes a large-company clause. The Open Source Initiative does not accept it as open source.
Third-Party Inference Pricing
Providers including Together, Fireworks, and Groq frequently offer comparable quality at lower cost than first-party APIs, plus access to open-source alternatives.
Why This Belongs in an Economics Page
Licence terms and hosting choice affect cost more than model capability does at scale. Our AI and automation work models both before committing.
Reading Generative AI Figures Critically
This category produces more unreliable statistics than any other covered in our data pages, so a short method is worth more than additional numbers.
Separate Spend From Deployment
Spend figures and deployment figures diverge sharply. A spend number tells you budgets are available, not that systems are in production.
Check Whether Revenue Is Consumer or Enterprise
Consumer app revenue and enterprise platform spend are different markets with different growth profiles, and they are frequently conflated.
Discount Vendor-Commissioned Surveys
Many prominent figures come from AI companies surveying their own market. We excluded these from our data pages for that reason.
Expect Rapid Obsolescence
Model versions, pricing, and vendor share change within months. Our data analytics work treats any figure in this category as provisional.
FAQs
How much are enterprises spending on generative AI?
Menlo Ventures estimated enterprise generative AI spend at approximately $37 billion in 2025, roughly triple the prior year. That growth rate reflects experimentation budgets rather than settled operating expenditure, so treat it as less durable than the figure suggests.
How much revenue do AI apps generate?
AI applications generated $18.5 billion in 2025, growing 273% year over year, with ChatGPT responsible for 43% of it. Generative AI apps reached 1.7 billion downloads and $1.9 billion revenue in the first half of 2025.
Why is the frontier model field so narrow?
Because a single frontier training run costs hundreds of millions, restricting participation to organisations with substantial capital or major platform backing. OpenAI and Anthropic together account for the large majority of category valuation.
Does self-hosting open-weight models save money?
At high inference volume, frequently yes, since it replaces per-token pricing with infrastructure cost. At low volume it is more expensive once operations and engineering time are counted honestly.
Is Llama free to use commercially?
Not unconditionally. Its licence restricts EU usage and includes a large-company clause, and the Open Source Initiative does not accept it as open source. Mistral and permissively licensed alternatives are cleaner for European organisations.
Why do generative AI statistics vary so widely?
Because spend and deployment are conflated, consumer and enterprise revenue are mixed, and many prominent figures originate from AI vendors surveying their own market. Check what a figure measures and who produced it before relying on it.
