AI Applications and Agents Industry Explained: The U.S. View
AI applications are where the money spent on chips, data centers and power has to come back as revenue. U.S. AI labs sell their models in three main ways: consumer chat apps with free and paid tiers, APIs that developers and businesses pay for by usage, and increasingly agents, software that carries out multi-step tasks such as writing and testing code or handling a workflow. Revenue is growing very fast. According to The Information, as reported by The Decoder in May 2026, OpenAI had annualized revenue of roughly $30 billion, first-quarter 2026 revenue of about $5.7 billion, and ChatGPT averaged 905 million weekly users; Anthropic's annualized revenue was nearly $45 billion. The companies may calculate run rates differently, so the numbers are not strictly comparable. Costs are large too: every request uses computing, and the same report put OpenAI's adjusted operating margin at minus 122% in the first quarter. Much of the labs' spending flows straight to cloud providers.
- Users
- ChatGPT averaged about 905 million weekly users in Q1 2026 (The Information)
- Revenue
- Anthropic annualized revenue nearly $45B; OpenAI roughly $30B (May 2026)
- Costs
- OpenAI adjusted operating margin about minus 122% in Q1 2026 (The Information)
How the chain works
Consumer apps, APIs and agents
| Consumer chat apps | API and enterprise | Agents | |
|---|---|---|---|
| Who pays | Individuals, via subscriptions | Developers and businesses, by usage | Businesses, by usage or seat |
| Scale signal | ChatGPT about 905M weekly users (Q1 2026) | Business revenue grows with usage | Writing and testing code is a common task |
| Cost driver | Free users still use compute | Tokens processed | Long tasks use many more tokens |
| What to watch | Paid conversion | Price per token | Tasks completed reliably |
Three ways to sell AI
Consumer apps reach the most people: according to The Information, ChatGPT averaged 905 million weekly users in the first quarter of 2026, short of a one-billion target. Most users pay nothing, so subscriptions from a minority have to cover compute for everyone.
APIs and enterprise contracts charge by use, which ties revenue directly to how much work customers send. Agents push usage further: instead of answering one question, they plan, call tools, check results and repeat, which can use many times more computing per task and lets them take on longer pieces of work if they are reliable.
Revenue is growing fast
The Decoder, citing The Information, reported in May 2026 that OpenAI's annualized revenue was roughly $30 billion and its first-quarter revenue about $5.7 billion, while Anthropic's annualized revenue was nearly $45 billion, with Anthropic projecting about $11 billion of second-quarter revenue and nearly $600 million of operating profit.
These figures are not audited, and companies may define annualized run rates differently, so they are best read as signs of direction and scale rather than precise comparisons.
Costs: compute is the bill
Every answer an AI model gives runs on accelerators in a data center. The same report put OpenAI's adjusted operating margin at about minus 122% in the first quarter of 2026, meaning it spent about $2.22 for each dollar of revenue, excluding items such as stock-based compensation.
That spending is revenue for others. Much of it goes to cloud providers, which is one reason Microsoft reported a $678 billion commercial backlog and Google Cloud a $514 billion backlog in mid-2026. Whether the application layer becomes broadly profitable depends on prices per task, falling compute costs and how much work businesses hand to AI.
How the U.S. view differs from the China view
The Chinese version of this page looks at Chinese AI apps and open-weight models, which compete on low prices and fast adoption inside China, often with limited access to the most advanced U.S. chips.
From the U.S. side the leading frontier labs and their largest customers are American, and the economic question is whether AI revenue keeps growing faster than the cost of computing. The answer matters for the whole chain, from cloud providers to chips and power.
2026: fast growth, big compute bills
By mid-2026, reports put OpenAI's annualized revenue around $30 billion and Anthropic's near $45 billion, with growth spread across consumer apps, APIs and agents.
Costs grew too. OpenAI's adjusted operating margin was reported at minus 122% in the first quarter, while Anthropic projected a small operating profit for the second quarter.
The difference shows that business models matter: enterprise usage priced per task can cover costs sooner than mass free consumer use. Either way, much of the spending flows to cloud and chip companies.
- AI application revenue is growing fast but is not yet broadly profitable.
- Compute costs are the main expense; prices per task decide margins.
- Business and agent usage may monetize faster than free consumer use.
Common misreadings
MythHuge user numbers mean huge profits.
RealityOpenAI's adjusted operating margin was about minus 122% in Q1 2026 despite 905M weekly users.
MythAI revenue is still tiny.
RealityReported annualized revenue was roughly $30B at OpenAI and nearly $45B at Anthropic in May 2026.
MythRun-rate figures are directly comparable.
RealityCompanies may define annualized revenue differently, and figures are unaudited.
MythAgents are just chatbots.
RealityAgents run multi-step tasks and can use many times more computing per task.
Pricing, costs and adoption: scenarios
| Scenario | Main effect | How it passes through |
|---|---|---|
| Agents become reliable for business tasks | Usage | Revenue grows faster than users |
| Price war on model access | Revenue per task | Lower margins, more volume |
| Compute costs fall | Margins | Applications become profitable sooner |
| Consumers resist paying | Subscriptions | Free usage keeps costs high |
| Regulation of AI use | Adoption | Slower rollout in some sectors |
| Labs cut compute purchases | Cloud providers | Slower cloud growth |
Supply and demand by segment
| Segment | Supply rigidity | Concentration | New demand | Policy role | What to watch |
|---|---|---|---|---|---|
| Frontier model labs | High (compute, talent) | High (a few labs) | Yes | Medium (AI rules) | Revenue, margins |
| Consumer apps | Low | High | Yes | Medium | Paid users |
| API and enterprise | Low | High | Yes | Medium | Usage, pricing |
| Agents | Low | Medium | Yes (new) | Medium | Reliability, adoption |
Supply rigidity: how long new capacity takes and whether money can speed it up. Concentration: whether output sits in a few countries or companies. The more rigid and concentrated, the more prone to shortages.
How to track it
| Indicator | What it tells you | Where to find it |
|---|---|---|
| AI lab revenue and user reports | Demand and monetization | Company announcements and press reports |
| Cloud revenue and backlog | Where AI spending goes | U.S. view: cloud and AI platforms |
| Software companies' AI revenue | Enterprise adoption | Quarterly results of listed software firms, free |
| AI compute | Cost side | U.S. view: AI compute |
Sources
- The Decoder citing The Information: OpenAI and Anthropic revenue and margins (2026-05-22)
- UpperEdge: Q2 2026 cloud earnings — AWS, Azure, Google Cloud (2026-08-13)
Last updated 2026-10-18.