The AI market expanded rapidly across all four layers (data, compute, models, and applications) in 2025. For example, NVIDIA’s data center revenue jumped from $115B to $194B in a single year; OpenAI reached about $13B in annual revenue; and Anthropic approached $7B in ARR.
We gathered revenue data from over 100 AI companies. Explore how revenues shifted across compute, data, models, and applications layers from 2023 to 2025.
AI stack revenue comparison
See the methodology to learn how we gathered AI revenues data.
AI revenue breakdown by subcategory
Note: We identified 21 subcategories under data, compute, model, and app layers. For simplicity, we included 7 subcategories with the highest revenue. See the dataset for all AI companies and subcategories.
Data layer revenues
For the data layer, Databricks ($4.8B), Snowflake ($4.68B), and MongoDB ($2.460M) have the highest revenue over the past 3 years. These three dominate because they own the foundational data infrastructure every AI application sits on: the lakehouse (Databricks), the warehouse (Snowflake), and the operational database (MongoDB). They capture AI demand regardless of which models or apps win.
Although the top of the data stack is dominated by data platforms, the middle and bottom layers tell a different story. Vector DBs (Pinecone, Qdrant, Weaviate) are all sub-$100M despite years of hype around RAG, and several companies were acquired before proving standalone viability (DataStax to IBM).
Compute layer revenues
The hyperscalers of AI (AWS: $128.7B, Google Cloud: $59B, and Microsoft Azure: $107.8B) and NVIDIA’s data center segment ($193.7B) account for approximately $490B in 2025 revenue, outgrowing all other players combined.
The interesting dynamic within compute is the emergence of a second-tier with CoreWeave ($5.1B), Lambda ($760M), and Together AI ($300M) in 2025. One possible explanation for the growing interest in cloud GPU players is that the existing GPU leaders (such as AWS, Azure, and GCP) are proving insufficient to meet market demand.
The open question is which moves faster: efficiency gains that reduce compute per query (smaller models, quantization), or new use cases that expand total demand (agents, video, enterprise rollouts). If efficiency wins, hyperscalers absorb the market; if adoption wins, specialist GPU clouds keep growing.
Model layer revenues
OpenAI ($13B) and Anthropic ($7B) are rapidly separating from the field. All other major players: Mistral ($400M), Cohere ($240M), xAI ($500M), ElevenLabs ($330M in voice) are clustered well below in comparison to OpenAI and Anthropic.
ElevenLabs in voice and Midjourney in image generation are two leaders in the model category, outperforming general-purpose foundation models. The hardest position in this layer is being a general-purpose model company without a major cloud distribution deal or an attractive consumer product. Mistral and Cohere both face that problem.
Application layer revenues
The pattern in the application layer shows that AI-native apps that replace an entire workflow are the best performing. One of the signals for this is the coding category, where Cursor, GitHub Copilot, Replit, Lovable, and Bolt collectively indicate that developers will pay more for tools that can both automate work.
Cursor’s jump from $1M to $1B in two years and Lovable’s increase from $1M to $400M in a single year are the most extreme growth figures in the dataset. They mark the shift from AI as a coding assistant to AI as the primary development environment, which fundamentally shifted the typical SaaS growth.
In terms of subcategories, healthcare AI (Abridge, Tempus) and fintech AI (Ramp, Brex) are growing as they operate in high-value regulated areas where the ROI of automation is easy to quantify.
Jasper AI’s revenue fell from $120M to $55M before partially recovering to $88M, still below its 2023 level. The drop shows that horizontal writing assistants without workflow lock-in are at risk of being displaced by both foundation models (such as ChatGPT) and embedded features in tools users already own (such as Notion AI and Google Docs).
For education, Chegg’s revenue nearly halved from $716M to $377M over two years, the sharpest sustained decline among companies in the dataset. A student paying $15/month for homework answers has little reason to keep doing so once ChatGPT offers the same service for free, and the collapse shows that owning a content library is a weaker moat than owning the workflow or the distribution channel.
AI revenues methodology
We gathered public data on AI revenues from research platforms such as Sacra, GetLatka, Macrotrends, and Crunchbase; company-owned sources such as investor relations reports, company newsrooms, official blogs, and SEC filings; financial media organizations such as Fortune, CNBC, Reuters, Bloomberg; tech media reports from TechCrunch; and regulatory/official sources for public companies such as SEC EDGAR filings.
Revenue figures reflect calendar year 2023, 2024, and 2025, or the fiscal year ending closest to those dates. Differences in fiscal calendars can also affect comparisons across different companies.
Note: For the many private companies in this dataset (for example, Anthropic, Mistral, ElevenLabs, and Cursor), the revenue figures are essentially informed estimates.
AI revenues dataset
Be the first to comment
Your email address will not be published. All fields are required.