How do foundation model companies make money beyond API access?
Whether you're building or investing in the foundation model market, our report covers everything you need to know: the players, the money, and what's coming next.
Summary
Foundation model companies make money beyond API access by selling the software, workflows, deployment control, audience access and transactions built around their models, where pricing power is often stronger than in raw token consumption.
The strongest businesses today are not the most speculative ones. Enterprise workspaces, consumer subscriptions and coding products already have substantial commercial evidence behind them, while outcome-based agents, marketplaces and commerce remain much earlier.
AI coding is the clearest proof that a foundation lab can move up the stack. Claude Code has become a multi-billion-dollar product because customers are paying for a development workflow and saved engineering time, not simply comparing the cost of Claude tokens with competing APIs.
Enterprise software changes the economics in another way. A model provider can charge for administration, security, connectors, identity, procurement and workspace access, then still collect additional usage revenue when employees run compute-heavy tasks.
OpenAI has the widest set of monetization options because consumer distribution creates businesses that enterprise-first labs cannot easily copy. The same user base can support subscriptions, enterprise products, coding, advertising, shopping and eventually transaction revenue.
Anthropic shows a different route. Its advantage is less about mass consumer distribution and more about turning professional workflows into high-value products, with Claude Code and large enterprise customers providing the clearest evidence so far.
Private and sovereign deployments are smaller in volume but can be unusually sticky. Cohere and Mistral can charge for dedicated infrastructure, deployment control, customization and long-term enterprise relationships where data location and security matter more than having the cheapest model.
Advertising has already moved from theory into a real OpenAI business, while commerce is still mostly an infrastructure story. Shopping integrations can eventually put an AI company close to the transaction itself, but disclosed transaction revenue still trails the much clearer evidence from subscriptions, enterprise software and ads.
Falling token prices probably make application ownership more valuable, not less. Raw inference becomes easier to compare and route between providers, while switching away from a deeply integrated coding tool, enterprise workspace or private deployment is much more disruptive.
The likely end state is not one universal business model. Consumer-scale labs can monetize audience and commerce, enterprise-heavy labs can sell workflows and agents, and open-model companies can push revenue toward licensing, deployment and support. The shared strategy is to capture more of the economic value created after the model produces an answer.
Written by the New Market Pitch teamWe track new markets so founders and investors can move faster. About us
Price is often what holds a market back. See how fast that's changing in our foundation model market report.
Why are foundation model companies trying to make money beyond APIs now?
Foundation model companies are moving beyond APIs because the bigger prize is increasingly the software, workflow or transaction built around the model.
Anthropic gives us one of the clearest examples. The company now says its run-rate revenue has passed $30 billion, up from roughly $9 billion at the end of 2025. At the same time, Claude Code has grown into a multi-billion-dollar product of its own. More than 1,000 Anthropic business customers are now spending at least $1 million a year on an annualized basis, twice as many as less than two months earlier.
OpenAI is moving in the same direction from a much larger consumer base. Its products now reach 2.5 million businesses, while enterprise contributes more than 40% of company revenue and is expected to approach consumer revenue. ChatGPT, Codex, enterprise products and advertising all sit alongside the developer API.
The same pattern shows up at smaller labs. Mistral sells paid assistant, coding and enterprise products. Cohere sells North as a workplace product and Model Vault as dedicated enterprise infrastructure. xAI has SuperGrok subscriptions, $30-per-user Business accounts, Enterprise contracts and Grok Bot.
The reason is straightforward. An API provider gets paid whenever somebody consumes its model. A company that also owns the application can charge for the interface, security, workflow, distribution and business outcome surrounding those same tokens.
What actually counts as foundation-model revenue beyond API access?
Revenue beyond API access means customers are paying the foundation model company for a product, deployment, workflow, license, audience or transaction rather than merely buying model tokens.
ChatGPT Plus is an obvious example because the user buys access to ChatGPT rather than a predefined quantity of GPT tokens. Claude Code subscriptions, enterprise workspace fees, dedicated Cohere deployments, commercial model licenses and ChatGPT advertising also qualify.
Cloud distribution is more ambiguous. A customer using Claude through Amazon Bedrock technically does not call Anthropic's direct API, but Anthropic is still mainly getting paid when the customer consumes Claude. We therefore treat Bedrock, Vertex AI and similar channels as alternative ways of selling model access rather than separate business models.
Enterprise products increasingly combine the two. Anthropic, for example, can charge for access to its enterprise software while also billing variable consumption. OpenAI similarly mixes workspace access, contracted commitments and metered usage across its business products.
This hybrid structure is becoming common because fixed subscriptions are easy for customers to budget, while usage charges protect the provider when an employee suddenly asks an agent to perform hours of compute-heavy work.
| Revenue model | What customers actually pay for | Beyond API access? | How developed is it? |
|---|---|---|---|
| Consumer subscriptions | Access to an AI product | Yes | Very large |
| Enterprise workspaces | Software, security and administration | Yes | Very large |
| Coding products | Developer workflow and completed work | Yes | Very large |
| AI agents | Tasks and multi-step workflows | Yes | Growing fast |
| Private deployment | Dedicated software and infrastructure | Yes | Established |
| Commercial model licenses | Rights to use or self-host models | Yes | Established but uneven |
| Cloud resale | Model consumption through AWS, Azure or Google | Mostly no | Very large |
| Advertising | Access to AI users | Yes | Already meaningful |
| Commerce fees | Transactions influenced or completed by AI | Yes | Early |
| Marketplaces | Distribution of third-party products | Yes | Early |
| Implementation services | Deployment and customization work | Yes | Supporting business |
Want to avoid the mistakes that sank others? Our foundation model market report breaks down what went wrong.
Are ChatGPT, Claude and Grok subscriptions already big businesses?
Consumer AI subscriptions are already a serious business, especially for OpenAI, but scale varies enormously between foundation model companies.
ChatGPT sits in a different league because OpenAI says the product now reaches roughly 1.2 billion people each week. Only a fraction needs to pay for Plus or higher tiers for subscriptions to become an enormous revenue base.
Anthropic also sells paid Claude plans, while xAI currently charges $30 per month for SuperGrok and $100 for SuperGrok Plus. Mistral offers paid individual access around Le Chat and its coding products.
These subscriptions let labs package multiple capabilities together. Customers can pay one recurring fee for reasoning, search, voice, image generation, coding, file analysis, connectors and agents without thinking about the price of every million tokens underneath.
Flat pricing does create a problem. Somebody who asks ten short questions per day costs little to serve. A heavy Claude Code or reasoning user can burn through far more compute while paying a similar monthly fee. Providers are responding with higher tiers, rate limits, included allowances and extra usage credits.
Are enterprise AI subscriptions becoming more important than consumer plans?
Enterprise AI subscriptions are becoming one of the biggest businesses around foundation models, and OpenAI is already close to having enterprise revenue rival its consumer business.
OpenAI reported earlier this year that enterprise represented more than 40% of total revenue and was on track to reach parity with consumer revenue. Its products now reach 2.5 million businesses.
Anthropic's enterprise trajectory is just as striking. Two years ago, only around a dozen customers spent at least $1 million annually with Anthropic on a run-rate basis. The company said that figure had exceeded 500 earlier this year. More recently it passed 1,000. That is roughly an 80-fold increase in large customers in about two years.
What companies pay for goes well beyond access to a stronger chatbot. Enterprise plans add identity management, audit controls, connectors, company data, administration, security, analytics, support and predictable procurement. xAI's Business and Enterprise offers now follow the same pattern, including role-based controls, consolidated billing, data-retention options, encryption and dedicated infrastructure.
This is where foundation labs start looking surprisingly similar to traditional enterprise software vendors. Once thousands of employees use the model every day, companies care as much about who can access which documents, how spending is controlled and where data goes as they do about benchmark performance.
The economics can be better than a pure software seat because the provider can collect both a platform fee and additional usage revenue. Anthropic's recent enterprise material explicitly focuses on helping administrators control consumption-priced spending rather than pretending every user costs the same amount.
Private startups are closing in on listed giants faster than most people think. See who's next in our foundation model market report.
Is AI coding the first killer application built by a foundation model company?
AI coding is the strongest evidence so far that a foundation model company can turn its own model into a huge standalone application business.
Claude Code passed $2.5 billion in annualized revenue after becoming generally available in May 2025. Anthropic said that figure had more than doubled from the beginning of 2026, while weekly users doubled and business subscriptions quadrupled. Enterprise customers accounted for more than half of Claude Code revenue.
The speed matters here. Anthropic built a multi-billion-dollar application business in roughly a year around a task that used to be one use case among many for language models.
OpenAI is chasing the same market with Codex. Earlier this year it reported three million weekly Codex users. More recent OpenAI enterprise data also shows how quickly coding agents are leaking into other departments: since February, weekly active enterprise Codex users rose 108-fold in legal, 41-fold in sales, 41-fold in recruiting and 26-fold in marketing, compared with fivefold growth in engineering.
That suggests coding agents are starting to become general computer-using tools rather than products used only by developers.
Coding also has unusually attractive pricing logic. A company deciding whether to spend another $100 on Claude Code does not care only about whether Claude tokens are slightly more expensive than GPT tokens. It can compare the cost with an engineer's time or the value of shipping software faster.
If you want more recent data on this point, please see our latest foundation model market report.
Can AI agents eventually charge for work instead of tokens?
AI agents could eventually let foundation model companies charge for completed work, but today most agent economics still depend heavily on subscriptions and consumption.
The usage is already changing quickly. OpenAI's enterprise data showed that agentic AI, measured through Codex output, represented 64% of combined Codex and ChatGPT output tokens among enterprise customers by June 2026. The same underlying models are being asked to spend more time working independently rather than simply answering individual questions.
Anthropic is pushing the same idea through Claude Code and Cowork. xAI now sells Grok Bot, which gets its own computer, can sign into tools and run recurring routines. Cohere's North packages agents, tools and orchestration into an enterprise environment.
Longer tasks create the opportunity. If an AI agent reconciles an account, prepares a sales report, fixes a software bug or processes a legal document, the customer's reference point gradually shifts away from the cost of the tokens involved. Human labor becomes the more relevant comparison.
We have not yet seen major frontier labs move broadly to pure outcome pricing. Most still protect themselves with subscriptions, credits or metered consumption because one agent task can use wildly different amounts of compute from another.
Investors react fast to big news. See what triggered the latest wave in our foundation model market report.
Can private and sovereign AI become a large business for Cohere and Mistral?
Private and sovereign AI can become a valuable business for Cohere, Mistral and similar labs because some companies and governments simply cannot use a normal shared AI service.
Cohere has gone unusually far in turning this requirement into a product. Model Vault lets companies run dedicated model instances with Cohere handling deployment and operations. Current published prices start around $2,000 per month for some retrieval configurations and reach $6,500 for larger ones. Generative models can cost far more on an hourly basis, with some Command configurations listed around $40 to $57.50 per hour.
Customers can also run Cohere's North application and control plane inside their own environment while using dedicated model infrastructure. For sensitive workloads, Cohere offers encrypted and private deployment options.
Mistral follows a similar route through private deployments, custom models, agents and enterprise workflows. Its European position also gives it a natural opening with customers that care about technological sovereignty and local control.
Governments are an important part of the same market. Cohere has expanded its U.S. public-sector distribution through Carahsoft, while its combination with Aleph Alpha strengthens a strategy built around sovereign enterprise AI. Anthropic offers Claude for Government in high-security environments, and OpenAI has developed dedicated government products and procurement routes.
The likely customers are banks, ministries, defense organizations, healthcare companies and other institutions where data location, isolation or security can decide whether a deployment happens at all.
These contracts can also create implementation revenue. Cohere uses solutions architects, applied machine-learning teams and customer-success staff to tailor deployments. Mistral provides customization around models and workflows. The economics are strongest when that work opens the door to recurring software and infrastructure revenue over several years.
If you want more recent data on this point, please see our latest foundation model market report.
Can open-weight foundation models actually make money?
Open-weight foundation models can make money when the free model pulls customers toward paid deployment, support, customization or commercial rights.
Mistral is one of the best tests of this strategy. The company has released models under open or source-available licenses while building paid products around managed inference, Le Chat, coding, enterprise deployments and commercial licensing.
Some Mistral licenses also draw a line between ordinary use and large commercial deployments. Companies above specified revenue thresholds can be required to negotiate commercial terms or use Mistral's hosted products.
Cohere has experimented with a similar balance by releasing some models openly while keeping paid enterprise infrastructure, North, Model Vault and customized deployments around them.
The business logic is comparable to open-source enterprise software. Making the technology easier to obtain can widen adoption, while customers with serious production requirements later pay for reliability, security, support or deployment.
There is still a ceiling to this strategy. If another provider offers an equally good open model with fewer restrictions, the model weights themselves have little pricing power. The surrounding product therefore has to become valuable enough that customers prefer paying the original developer.
Bigger than you think, or smaller? Our foundation model market report puts real numbers on it.
How much money can foundation model companies make through AWS, Azure and Google Cloud?
Cloud partners can bring foundation model companies huge volumes of enterprise business, although the cloud provider usually takes part of the economics in exchange.
Anthropic is the strongest example because Claude is distributed across Amazon Bedrock, Google Vertex AI and Microsoft Azure. The company recently said Claude is the only frontier model available through all three of the world's largest cloud platforms.
That reach matters. Large companies already have cloud contracts, security reviews, procurement systems and committed spending with AWS, Microsoft or Google. Buying Claude through an existing cloud account can be much easier than onboarding Anthropic as another strategic vendor.
The trade-off is margin. Reporting based on Anthropic's financial forecasts has suggested that cloud partners can capture a substantial share of the economics on sales they originate. That is the price Anthropic pays for distribution, infrastructure and access to customers it might otherwise have to win directly.
The tension explains why the same companies are building direct enterprise software at the same time. Selling through AWS can accelerate adoption. Getting the customer onto Claude Enterprise, Claude Code or another directly controlled product gives Anthropic a deeper relationship.
Can ChatGPT advertising become a major OpenAI business?
ChatGPT advertising is already a meaningful business for OpenAI, reaching a $1 billion annualized revenue run rate less than 200 days after launch.
The platform now has tens of thousands of advertisers and operates across more than 60 countries. OpenAI also says ChatGPT reaches roughly 1.2 billion people per week, giving its advertising product an audience no independent AI lab can currently match.
The speed is hard to ignore. Going from launch to $1 billion of annualized revenue in under seven months puts ChatGPT Ads far beyond an experimental monetization feature.
The underlying economics are attractive because advertising can turn free users into revenue without forcing them to buy a subscription. Someone researching a laptop, software package, holiday or home renovation is already expressing commercial intent. Advertisers routinely pay to appear during exactly those decisions.
OpenAI has continued expanding the format. It recently introduced more visual ads and new measurement tools, which makes ChatGPT advertising look increasingly like a proper ad platform rather than a few sponsored placements attached to conversations.
Other model companies will find this harder to copy. Anthropic, Mistral and Cohere do not have comparable consumer distribution today. xAI has a more plausible route because of Grok and its connection with X.
If you want more recent data on this point, please see our latest foundation model market report.
Revenue is easy to show, profit is harder. Our foundation model market report looks at who's actually making money.
Could AI shopping become even bigger than advertising?
AI shopping could eventually produce enormous transaction revenue for OpenAI, although the money is still far less proven than the shopping infrastructure itself.
ChatGPT can already research products, compare them visually, pull in merchant catalog information and, for some eligible products, let users complete checkout directly inside ChatGPT. OpenAI's Agentic Commerce Protocol connects merchants and product inventories to these experiences.
The merchant ecosystem has also become more concrete. Shopify product data can flow into ChatGPT, while OpenAI has introduced formal merchant-participation rules and product-feed infrastructure. A recent expansion with Albertsons is bringing Safeway grocery shopping directly into ChatGPT as part of a broader retail relationship.
That puts OpenAI unusually close to the purchase itself. Someone asking which laptop to buy might have spent twenty minutes explaining budget, intended use, preferred brands and trade-offs before reaching a decision. That is richer commercial intent than a two-word search query.
Several business models become possible from there: merchant fees, transaction commissions, paid distribution tools, checkout economics or financial-service partnerships.
For now, we should keep the scale in perspective. OpenAI has publicly quantified ChatGPT Ads at a $1 billion annualized run rate. It has not disclosed comparable shopping or transaction revenue.
Will OpenAI's marketplace turn into an App Store for enterprise AI?
OpenAI's enterprise Marketplace is currently more useful as a way to lock in customer spending than as an App Store-style commission machine.
The Marketplace now includes products from companies such as Adobe, CrowdStrike, Datadog, ServiceNow, Sierra, Vercel and Zendesk. Eligible enterprise customers can apply part of an existing OpenAI commitment toward qualifying partner products.
The current transaction flow is revealing. The customer contracts with the partner, and the partner invoices the customer directly. OpenAI then reconciles the eligible amount against the customer's commitment. There is no normal self-service marketplace checkout yet.
OpenAI still gains something valuable before collecting a large visible commission. A company that has committed millions of dollars to OpenAI can spend part of that budget across a broader software ecosystem, making the commercial relationship more useful and harder to replace.
Anthropic is taking a related route through partnerships rather than a marketplace. Its partner program supports implementation firms with training, technical help and co-selling so more companies can deploy Claude.
Market size is where every investment case starts. Get the full picture in our foundation model market report.
Which non-API business models are actually strongest right now?
Enterprise workspaces, consumer subscriptions and coding products currently have the best evidence behind them, while advertising has moved into the proven category much faster than expected.
Coding stands out because Claude Code has already created billions of dollars in annualized revenue. Enterprise software is broader: OpenAI says more than 40% of company revenue now comes from enterprise, while Anthropic's count of customers spending at least $1 million annually has climbed above 1,000.
Consumer subscriptions benefit from far greater distribution, particularly at OpenAI. Private deployments have less volume but can support large, sticky contracts in regulated industries.
Advertising has now crossed an important line because ChatGPT Ads has reached a billion-dollar run rate. Shopping, marketplaces and outcome-based agent pricing have credible products behind them but much less disclosed revenue.
| Non-API business | Evidence today | Pricing power | Our current view |
|---|---|---|---|
| Enterprise workspaces | Very strong | High | Core business |
| Consumer subscriptions | Very strong | Medium-high | Core business |
| Coding products | Very strong | High | Core business |
| AI work agents | Growing quickly | Potentially high | Likely major |
| Private deployments | Established | High | Strong enterprise niche |
| Advertising | $1B+ run rate at OpenAI | High at scale | Proven for OpenAI |
| Commercial licensing | Established but uneven | Medium | Useful secondary model |
| Professional services | Common around deployments | Medium | Supporting revenue |
| Marketplaces | Early | Unclear | Strategic today |
| Shopping and transaction fees | Early | Potentially high | Promising but unproven |
| Outcome-based agent pricing | Very early | Potentially very high | Longer-term opportunity |
If you want more recent data on this point, please see our latest foundation model market report.
Does selling applications protect foundation model companies from falling token prices?
Selling applications gives foundation model companies much better protection against falling token prices because customers become less sensitive to the price of the underlying model.
Raw model access is easy to compare. If several models perform well enough, developers can route workloads toward the cheapest provider. Better chips, smaller models, quantization and intense competition keep pushing inference costs lower.
Applications create more friction around switching. A business using Claude Code evaluates the whole coding workflow. A company deploying ChatGPT across thousands of employees cares about connectors, permissions, saved context, internal tools, security controls and employee familiarity alongside model quality.
Open-weight companies have even more reason to move upward. When customers can download the model, charging a large premium simply for inference becomes difficult. Mistral and Cohere consequently put more commercial weight behind enterprise software, customization and deployment.
Cheaper inference could actually help these businesses if their application prices fall more slowly than their cost of producing the underlying intelligence. The margin then migrates from the model call toward the product wrapped around it.
Is this market still open or already locked up? Our foundation model market report has the answer.
Do products like Claude Code and ChatGPT Enterprise actually have better margins than APIs?
Products can earn much more revenue per customer than raw APIs, but heavy AI workloads still carry real compute costs, so higher pricing does not automatically produce traditional SaaS margins.
Enterprise software provides the cleanest improvement. A model company can charge for administration, security, connectors and workspace access before the customer uses another token. Heavy activity can then be billed separately.
Dedicated deployments can also produce attractive economics because customers pay for reserved infrastructure and operational guarantees. Cohere's current Model Vault pricing makes this unusually visible: the customer buys a dedicated instance rather than paying only for occasional requests.
Advertising changes the equation in another way. OpenAI can generate revenue from people who would otherwise use ChatGPT for free, helping offset the inference cost of serving a very large unpaid audience.
Agents and coding products are trickier. They can justify much higher customer spending because they perform valuable work, yet they also tend to consume far more compute than short chat interactions. Anthropic's enterprise documentation now devotes substantial attention to controlling consumption-priced spending for exactly this reason.
Which foundation model companies are best positioned beyond APIs today?
OpenAI currently has the broadest non-API business, Anthropic has the clearest proof that a specialized AI application can become huge, and Cohere and Mistral have stronger positions in private and sovereign deployments.
OpenAI can monetize the same model infrastructure through ChatGPT subscriptions, enterprise software, Codex, advertising, shopping and a growing enterprise marketplace. Its consumer scale gives it options that enterprise-first labs simply do not have.
Anthropic has built a different kind of strength. Its company-wide run-rate revenue now exceeds $30 billion, while Claude Code has shown that a foundation lab can create a multi-billion-dollar application around one high-value workflow. The rapid increase to more than 1,000 million-dollar business customers also gives Anthropic a particularly strong enterprise base.
Mistral has less distribution but more freedom to play the open-model and sovereignty angle. Its combination of open weights, commercial licensing, enterprise products and private deployment gives customers several ways to use the same underlying technology.
Cohere is even more concentrated on enterprise infrastructure. North, Model Vault, private deployments and retrieval products make sense for customers that care more about security and control than about having the world's most popular consumer chatbot.
xAI has a potentially powerful combination of Grok subscriptions, enterprise seats, Grok Bot and distribution through X, although there is still less public evidence that its business products have reached OpenAI or Anthropic scale.
| Company | Strongest businesses beyond API access | Biggest advantage | Main challenge |
|---|---|---|---|
| OpenAI | ChatGPT, enterprise, Codex, ads, commerce | Massive consumer and business distribution | Huge compute requirements |
| Anthropic | Claude Code, Claude Enterprise, Cowork | Strong enterprise and professional adoption | Smaller consumer audience |
| Mistral | Enterprise, coding, licensing, private deployment | Open models and European sovereignty | Much smaller distribution |
| Cohere | North, Model Vault, private AI | Enterprise security and deployment control | Little consumer scale |
| xAI | SuperGrok, Business, Enterprise, Grok Bot | Consumer distribution through Grok and X | Less proven enterprise scale |
| Meta/Llama | Ecosystem distribution | Huge installed ecosystem | Direct model revenue is secondary |
In many markets, a handful of startups take most of the money. Our foundation model market report shows who's getting the biggest slice.
Will foundation model companies eventually make more money from products than from APIs?
Yes. The strongest foundation model companies are increasingly likely to make more money from applications, enterprise software, agents and other products around their models than from straightforward API access alone.
The transition has already started. Coding has produced a multi-billion-dollar business inside Anthropic. Enterprise contributes more than 40% of OpenAI revenue. ChatGPT advertising has crossed a $1 billion annualized run rate. Private-deployment companies are charging for dedicated infrastructure and enterprise control rather than only token consumption.
APIs will remain enormous. Millions of companies want programmable access to models, and every new application creates additional inference demand. But raw model access sits in the part of the stack where customers can compare performance and price most easily.
Owning the actual product changes that relationship. A developer working in Claude Code, an employee using ChatGPT Enterprise, a bank running a private Cohere deployment or a shopper completing a purchase inside ChatGPT is buying something broader than model intelligence.
Several endgames are already visible. OpenAI is becoming a consumer and enterprise platform with subscriptions, advertising, commerce and software. Anthropic is building around professional work and agents. Cohere and Mistral are leaning harder into enterprise control, private deployment and sovereignty.
There probably will not be one winning revenue model for every foundation lab. Consumer scale favors subscriptions and ads. Deep enterprise adoption favors workspaces and agents. Open models push monetization toward deployment and services. Commercial-intent traffic opens commerce.
The direction is clear: foundation model companies want to own more of the work that happens after the model produces an answer.
That is where the stronger pricing power is starting to appear.
If you want more recent data on this point, please see our latest foundation model market report.
Our methodology
The question behind this analysis is easy to answer loosely and much harder to answer well. Foundation model companies are expanding into many different products at the same time, but simply listing those products does not tell us which business models are actually becoming meaningful beyond API access.
We therefore broke the question into the main commercial dimensions where value is starting to accumulate around foundation models: direct software products, enterprise adoption, specialized applications, agents, private deployment, licensing, distribution, advertising, commerce, marketplaces and implementation.
For each dimension, we looked for the freshest signals that could show what is actually happening commercially. We prioritized disclosed revenue and run-rate figures when available, then looked at customer adoption, pricing, deployment structures, product availability, usage growth, distribution and transaction infrastructure. We used the signal that most directly answered the question at hand rather than forcing very different businesses into the same metric.
We also distinguished between different levels of evidence. A product already generating material revenue carries more weight than one with strong adoption but no disclosed revenue; a live commercial product carries more weight than an announced direction; and functioning infrastructure for a future business model is evidence of progress, but not evidence that the revenue has already arrived.
Where companies report annualized or run-rate revenue, we use it as a signal of current commercial scale and momentum rather than treating it as completed fiscal-year revenue. Where product-level revenue is not disclosed, we do not manufacture an estimate: we use the strongest observable commercial evidence available instead.
We relied primarily on first-hand company disclosures, product documentation and pricing pages, and favored the newest relevant disclosure when a company had published several successive updates.
The final conclusions come from putting those signals together point by point. Rather than relying on intuition about where AI companies should make money, we looked at what customers are already paying for, where adoption is accelerating, where companies have built genuine commercial infrastructure, and where monetization is still mostly prospective. That structured aggregation of recent evidence is what gives us a clearer and more defensible answer to the main question.
Key sources used for this analysis include: OpenAI on overall product and business scale, OpenAI on enterprise revenue and Codex adoption, OpenAI's enterprise usage data, OpenAI on scaling Codex, OpenAI on ChatGPT Ads revenue and advertiser growth, OpenAI on advertising formats and measurement, OpenAI's Agentic Commerce Protocol documentation, OpenAI's commerce and checkout documentation, OpenAI on the Albertsons and Safeway integration, and OpenAI's enterprise Marketplace documentation.
For Anthropic, xAI, Cohere and Mistral, key sources include: Anthropic on Claude Code revenue and large-customer growth, Anthropic on company growth and cloud distribution, Anthropic's Claude Partner Network, xAI's current pricing, xAI on Grok Bot availability, Cohere's Model Vault pricing, Cohere on using Model Vault with North, Cohere on North 2 and private enterprise deployment, Mistral on Le Chat Enterprise, and Mistral's licensing guidance for open models.
A dollar invested early can be worth many times more today. See who made the best bets in our foundation model market report.
Written by the New Market Pitch teamWe track new markets so founders and investors can move faster. About us