What are the main business models in generative AI?
In our generative AI market deck, you will find everything you need to understand the market
SUMMARY
The main business models working in generative AI today are consumer subscriptions, enterprise seats, API usage, generation credits, subscription-plus-usage plans, outcome pricing, managed AI infrastructure, and monetization inside larger software or consumer platforms.
The clearest market-wide pattern is not a single pricing model but a hybrid: recurring access plus a meter for expensive usage. AI companies want SaaS-like predictability without pretending that a casual user and a power user cost the same to serve.
Seat pricing is proving much more durable than early AI commentary suggested. Enterprises still like predictable per-user budgets, but the seat is increasingly becoming the entry ticket while credits, capacity tiers or usage charges absorb the cost of heavy AI consumption.
Token pricing remains the natural model at the foundation-model layer because customers are buying variable computation. Its weakness is that comparable intelligence is getting cheaper, which makes raw, undifferentiated model usage a difficult place to defend margins for long.
Credits are emerging as the practical pricing language for multimodal AI. Video, image, voice and dubbing products can hide very different internal compute costs behind one balance that customers can actually understand.
Outcome pricing is the biggest structural break from traditional SaaS. When an agent can resolve a support issue, qualify a lead or complete another tightly defined job, the customer can pay for the result rather than the employee seat or the tokens consumed along the way.
The strongest application businesses are moving closer to workflows, not just models. Coding, customer support, corporate video and voice agents can charge above raw model cost because they own context, integrations, controls and the place where the work gets done.
Falling inference costs do not help every AI company equally. They squeeze businesses that resell computation with little differentiation, but they can expand margins dramatically for companies that keep charging for a finished workflow or business outcome while their model costs fall underneath.
AI infrastructure is already a standalone business rather than a temporary layer between model builders and applications. The value is shifting from access to GPUs alone toward inference optimization, specialized models, reliability, routing, fine-tuning and reserved capacity.
Large incumbents have a separate advantage: they can monetize AI through higher tiers, additional seats, credits, retention, customer expansion, ads or transactions without needing every AI feature to become a standalone product. Distribution is part of the business model.
The direction to watch is how far revenue can move away from raw computation and toward useful work. The closer pricing gets to the value of the task completed, while compute costs continue to fall, the more attractive the economics can become.
This market map, featured in our generative AI market deck, highlights top companies and startups in the generative AI market
What does a generative AI company actually sell?
A generative AI company can sell access to intelligence, the amount of intelligence used, the finished work produced, or the infrastructure needed to run it.
Those differences are more important than the price printed on the website. ChatGPT Plus and Claude Pro mostly sell monthly access. OpenAI and Anthropic APIs charge according to model consumption. Runway and Synthesia translate expensive generation into credits. Fin charges when its AI agent produces a defined outcome. Fireworks and Together AI sell inference and dedicated computing capacity.
A $20 subscription and $20 of API usage may generate the same revenue today, but the economics underneath are completely different. With a subscription, the vendor takes the risk that a heavy user consumes much more compute than an average one. With token pricing, that risk largely stays with the customer. With outcome pricing, the AI company accepts even more execution risk because it gets paid only when the software accomplishes something useful.
We also need to separate the model from the business built around it. A company can use Claude or GPT underneath and still have a very different business from Anthropic or OpenAI. Cursor sells a coding environment and workflow. Synthesia sells corporate video production. Fin sells automated customer-service work. The underlying models are only part of what customers are buying.
A useful way to map generative AI today is simply to ask what triggers the customer's bill. The answer usually falls into recurring access, usage, credits, outcomes, infrastructure, or some combination of them.
If you want more recent data on this point, please see our latest generative AI market report.
Is generative AI settling on one business model?
Yes, but only at a higher level: generative AI is currently converging on recurring access plus some way to charge heavy users more.
The details vary, yet the pattern keeps appearing. OpenAI's current ChatGPT Business structure combines fixed-price Standard seats with workspace credits for additional usage, while its new Premium seats offer five times more usage for a higher monthly price. Cursor sells individual and team subscriptions but allows customers to keep using expensive models through on-demand billing. Adobe combines Creative Cloud subscriptions with generative credits. Synthesia gives users one common credit balance across video, dubbing and other AI features. ElevenLabs mixes monthly plans with usage and pay-as-you-go charges.
The reason is simple. AI companies want the predictable recurring revenue that made SaaS attractive, but they cannot ignore the cost difference between someone who occasionally asks a chatbot a question and someone running agents for hours every day.
Cursor's recent pricing changes make the problem unusually visible. The company says a relatively small group of power users drives much of a team's unpredictable AI spend. Its response was to introduce a Premium team seat with five times the included usage for three times the price of a Standard seat. OpenAI is now moving in almost exactly the same direction with different seat types and shared credits inside ChatGPT Business.
We would describe the emerging default as "subscription with a meter." The customer gets predictable baseline access, while the AI company keeps a way to monetize unusually expensive usage.
| Product | What customers commit to | What increases the bill |
|---|---|---|
| ChatGPT Business | Paid seats | Higher-capacity seats and shared credits |
| Cursor | Individual or team seats | Premium seats and on-demand model usage |
| Adobe Firefly | Software plans | Generative credits and enterprise operations |
| Synthesia | Monthly or enterprise plan | Shared AI credits |
| Fin | Helpdesk seats where relevant | Successful AI outcomes |
| Fireworks AI | Platform relationship | Tokens, training and dedicated compute |
As this chart shows, and as featured in our generative AI market deck, search interest in LLMs has surged
Are people really paying monthly for generative AI?
Yes. Consumer subscriptions are still one of the biggest proven generative AI business models, even though enterprise revenue is catching up quickly.
ChatGPT showed that millions of people will pay every month for better access to a general-purpose AI assistant. Anthropic followed the same basic structure with Claude Pro and much more expensive Max tiers for people who need substantially more capacity. Cursor has also shown that individual professionals will pay directly for AI when it sits inside a workflow they use every day.
What has changed lately is the importance of business customers. OpenAI said earlier this year that enterprise customers had already grown beyond 40% of company revenue and were on track to reach parity with consumer revenue. More recently, OpenAI said its products were being used by more than two million businesses, twice as many as a year earlier.
That does not make consumer subscriptions unimportant. It shows how quickly the revenue mix is broadening.
Subscription pricing works especially well for products people use frequently but unevenly. A subscriber who pays every month and uses the product lightly helps subsidize someone who uses more compute. The company gets recurring revenue without issuing a small invoice after every conversation.
The limit becomes obvious with power users. Generative AI has a real marginal cost, particularly when users choose expensive reasoning models, long context windows, image generation, deep research or autonomous coding agents. That is why "unlimited" plans increasingly come with fair-use rules, different capacity levels, slower modes or additional credits.
Consumer AI subscriptions are likely to remain a huge business. They are simply becoming less pure than the $20 flat-fee model first suggested.
Why are companies still paying per seat for AI?
Enterprise AI is still mostly bought like software: one paid seat per employee, with usage controls increasingly layered on top.
Microsoft gives us the clearest evidence. Microsoft 365 Copilot had around 15 million paid seats at the end of one quarter, passed 20 million the following quarter and then exceeded 30 million in its latest fiscal-year results. That means paid seats grew roughly 33% in one quarter and then another 50% in the next. Microsoft also said the net number of seats added in the latest quarter doubled compared with the previous quarter.
Those numbers are hard to reconcile with the idea that per-seat pricing is already obsolete.
A seat is easy for a company to understand. If 5,000 employees receive a tool that costs $25 per employee per month, the buyer can budget $1.5 million a year without estimating billions of tokens. Procurement teams already know how to buy software this way, finance teams know how to forecast it, and managers know which employees have access.
OpenAI uses the same logic with ChatGPT Business. Cursor currently charges $40 per month for a Standard Teams seat and $120 for a Premium seat before annual discounts. Anthropic also sells Claude for teams and enterprises around user access rather than forcing every employee to think about tokens.
The seat starts to break down when AI does much more than assist a human. Two employees with the same license can now create radically different costs if one occasionally drafts emails while the other runs several autonomous agents all day.
Seat pricing is evolving rather than disappearing. The seat usually buys the relationship with the customer; usage determines how expensive that relationship becomes.
If you want more recent data on this point, please see our latest generative AI market report.
This chart, featured in our generative AI market deck, shows annual VC investment in generative AI startups
Why do OpenAI, Anthropic and other model companies charge per token?
Model companies charge per token because API customers are buying variable computation, and the amount of work can change enormously from one request to another.
A short classification request may contain a few hundred tokens. An agent reading a repository, calling tools repeatedly and generating a long answer can process millions. Charging both customers the same amount would make little economic sense.
OpenAI's current rate cards separate input tokens, cached input and output. Anthropic does the same across Claude models. Pricing also changes by model because a smaller model designed for high-volume tasks can cost a fraction of a frontier model doing heavier reasoning.
The scale is already enormous. OpenAI said earlier this year that its APIs were processing more than 15 billion tokens per minute. Anthropic has reported an even more striking commercial trajectory: company run-rate revenue rose from about $9 billion at the end of 2025 to more than $30 billion during 2026, while the number of customers spending at least $1 million annually rose from more than 500 to more than 1,000 in less than two months.
Those figures include more than raw API usage, but they show how quickly businesses can spend when intelligence becomes a metered input inside their own products and workflows.
The weakness of token pricing is equally clear. The price of a given level of AI capability keeps falling. Epoch AI's historical inference-price work has found declines of more than an order of magnitude for comparable levels of model performance over relatively short periods. Competition between proprietary models, open models and inference providers keeps pushing the same way.
A model company can build a huge business on tokens, but selling undifferentiated tokens at a markup is fragile. Customers can route requests toward cheaper models, cache prompts, use smaller models for easier tasks or move workloads to another provider.
The durable model businesses need something customers cannot replace as easily: better models, better reliability, lower latency, proprietary capabilities, fine-tuning, tooling, distribution or a broader enterprise relationship.
Why do image, video and voice AI use credits?
Credits have become the cleanest way to price multimodal generative AI because different creations can have radically different computing costs.
Runway is a good example. Its plans provide monthly credit balances, while different models consume those credits at different rates. Gen-4.5 currently uses 12 credits for each second of generated video, while other video models can use much more. Runway's developer API gives the credit a simple monetary value of one cent, so the same accounting system can cover many models and media formats.
Synthesia has gone in the same direction. Credits are now one shared currency across its usage-based AI features. A Starter subscription includes 1,200 credits, while its Creator tier includes 3,600. Enterprise customers can allocate a central credit pool across teams and workspaces.
Adobe has built credits into a much larger software ecosystem. Standard Firefly generations can consume very little, while premium video and audio operations consume more. Adobe also allows enterprise customers to use shared credits after individual allocations run out. In its latest reported quarter, Firefly ARR across the Firefly app, credit packs and Firefly Enterprise was approaching $300 million, while Adobe said generated enterprise assets had increased more than fourfold year over year.
ElevenLabs stretches the same idea across speech, transcription, dubbing, music and conversational agents. The company recently moved further toward pay-as-you-go pricing as its product range expanded. ElevenLabs had already passed $500 million in ARR after ending 2025 around $350 million, with management saying enterprise voice-agent deployments were driving the acceleration.
Credits let all of these companies hide a complicated cost structure behind one understandable number. Customers do not need to know which GPU ran a video or how many internal model calls produced a dubbed presentation. They just see their balance falling.
For multimodal AI, that simplicity is hard to beat.
This chart, featured in our generative AI market deck, looks at OpenAI’s strategy in generative AI
Why are subscriptions and usage charges getting mixed together?
The subscription-plus-usage model is becoming generative AI's commercial default because fixed monthly pricing cannot absorb the difference between ordinary users and AI power users.
Cursor has essentially documented the transition in public. Its Teams plan now combines paid seats with included model allowances, separate pools for its own models and third-party models, Premium seats for heavy users, and on-demand billing after included usage runs out. Cursor says daily agent users often consume $60 to $100 of model usage per month, while users running multiple agents or automations can exceed $200.
That cost distribution makes a single $20 or $40 price awkward. Set the subscription too low and the heaviest users become loss-making. Set it high enough to cover them and casual users leave.
OpenAI is now making the same segmentation more explicit inside ChatGPT Business. Standard seats remain fixed-price, Premium seats add five times the usage, and shared workspace credits can take over once included limits are reached. It looks much closer to cloud billing layered inside SaaS than to old unlimited software.
A very fresh example from Canva shows why companies care so much about the cost side. Canva founder Cliff Obrecht said the company had to slow the rollout of some ambitious AI features after compute costs made them uneconomical. Canva subsequently cut its cost per AI task by about 90% by building more technology internally, routing workloads more efficiently and avoiding expensive fresh generation when existing design assets could do the job.
Canva already had more than 31 million paying users and roughly $4 billion in annualized revenue before this adjustment. If a company of that size can still find an AI feature too expensive to roll out widely, marginal cost is clearly shaping product design and pricing.
The hybrid model handles the problem without making customers read an API invoice every month. Everyone pays for access; the minority consuming disproportionate compute pays more.
If you want more recent data on this point, please see our latest generative AI market report.
Will AI agents be sold by the task instead of by the seat?
For narrow, measurable jobs, yes: outcome pricing is already working for AI agents, and it is one of the clearest breaks from traditional SaaS.
Fin is the easiest example to understand. The company currently charges $0.99 when its AI produces outcomes such as resolving a support conversation, completing certain procedures or disqualifying an unsuitable sales prospect. When Fin successfully qualifies a prospect for sales, the price rises to $9.99.
Customers are paying for something they can compare with human work. A support manager knows roughly what a human-assisted ticket costs. A sales team knows what it pays to generate and qualify pipeline. Tokens are almost irrelevant to that purchasing decision.
Fin says more than 7,000 teams now use the product, and its average resolution rate has reached about 76%. The company has also expanded the model beyond customer support into sales outcomes. That is more interesting than it first sounds: outcome pricing gets much more flexible once the unit can change with the value of the task.
Sierra has pushed the idea even harder. It markets its agents around results rather than tokens and ended its second year with more than $150 million in ARR. Lately, Sierra has expanded from one-off conversations toward agents that can manage customer interactions across days or weeks while keeping the same outcome-based philosophy.
There are limits. A password reset has an obvious finish line. "Improve our marketing" does not. Companies also need rules for cases where an AI and a human jointly produce the result, where a supposedly resolved customer comes back later, or where two outcomes have completely different economic values.
We expect outcome pricing to spread first where success can be counted cleanly: support resolutions, qualified leads, collections, bookings, completed back-office tasks and similar workflows.
| AI product | What triggers payment | Current example |
|---|---|---|
| Fin for Service | Customer issue is resolved or another defined outcome occurs | $0.99 per qualifying outcome |
| Fin for Sales | Prospect meets the customer's qualification rules | $9.99 per qualification |
| Sierra | Agreed business result | Contracted outcome pricing |
| Traditional AI SaaS | Employee has access | Monthly or annual seat |
| Model API | Model performs computation | Tokens or related usage |
This chart, featured in our generative AI market deck, shows annual funding in generative AI startups
Can AI infrastructure companies make money without owning the best model?
Yes. Companies such as Fireworks, Together AI and Hugging Face show that running, specializing and serving models can be a large business even when the infrastructure company did not train the original foundation model.
Fireworks is the clearest proof today. The company recently said it had passed $1 billion in annualized revenue while serving more than 40 trillion tokens every day. More than 95% of those tokens came from models specialized on customers' proprietary data.
That last figure tells us more than the revenue number. Fireworks customers are increasingly paying to turn general models into something specific to their own business, then run those models cheaply and reliably at scale.
The charging model changes as customers get bigger. A developer can start with serverless inference priced by tokens. High-volume customers can move to dedicated deployments or reserved capacity. Training adds another usage stream. Together AI follows a similar path from serverless inference to dedicated endpoints and GPU clusters. Its current dedicated offerings are billed by hardware time, while serverless models are priced per million tokens.
Open-source models actually make this business model more useful. When a company can choose between dozens of competitive model families, it needs somewhere to run them, benchmark them, fine-tune them and move workloads between them. Hugging Face monetizes that problem through enterprise subscriptions, inference routing, pay-as-you-go compute and dedicated endpoints billed according to the infrastructure being used.
There is still plenty of commodity risk. A plain H100 hour can be compared across several providers in seconds. The real margin has to come from better utilization, faster inference, model optimization, tooling, reliability, enterprise controls or specialized models.
Fireworks' current scale suggests those layers can be worth a lot. AI infrastructure looks less like a temporary bridge to better models and more like the cloud layer underneath an increasingly fragmented model market.
Are vertical AI apps better businesses than general chatbots?
Often, yes. Vertical AI applications can charge for a workflow customers already value, while a generic chatbot has to keep proving why it deserves another subscription.
Coding gives us the strongest evidence so far. Claude Code reached more than $2.5 billion in run-rate revenue according to Anthropic, more than doubling from the beginning of the year. Enterprise customers now account for more than half of Claude Code revenue, and business subscriptions have quadrupled over the same period.
Cursor has travelled even further from the "thin wrapper" stereotype. The company currently says more than 50,000 enterprises use Cursor and 64% of the Fortune 500 use the product. Enterprise customers generate more than 100 million lines of code through Cursor every day. Its commercial product now includes codebase context, agents, code review, cloud agents, team controls, analytics and model routing.
Voice AI shows the same move toward workflows. ElevenLabs originally became famous for generating realistic speech, but its recent revenue growth has increasingly come from companies deploying voice agents in support, sales, hiring and other operations.
Synthesia found a similar niche in corporate training and internal communication. More than 50,000 teams now use the platform, and enterprise plans include collaboration, brand controls, custom credits, onboarding and implementation rather than simply selling avatar-generation minutes.
These companies gain room to charge above raw model cost because the product knows where the work happens. Cursor understands repositories and development workflows. Synthesia sits inside corporate video production. ElevenLabs connects speech models with real conversations and telephony. Fin sits inside customer-service systems.
A vertical product can still be copied, especially if most of the user experience comes from somebody else's model. Defensibility improves sharply once the application owns integrations, historical context, evaluations, company-specific data and the workflow itself.
Some of the fastest-growing generative AI businesses today are moving closer to the job rather than trying to build another general chatbot.
If you want more recent data on this point, please see our latest generative AI market report.
This chart, featured in our generative AI market deck, compares the main business model options for generative AI SaaS platforms
Can Microsoft, Adobe and other incumbents make more money by adding AI to existing software?
Yes. For incumbents with huge installed bases, generative AI can generate revenue through upgrades, seats, credits and higher customer spending without needing a separate standalone product to win.
Microsoft has the clearest distribution advantage. As discussed earlier, Microsoft 365 Copilot has already passed 30 million paid seats. Microsoft's broader AI business had also exceeded a $37 billion annual revenue run rate by its fiscal third quarter, up 123% year over year. That broader figure includes Azure AI and other products, but it shows the size of the commercial machine already sitting behind Microsoft's AI strategy.
Adobe uses its installed base differently. Firefly sits inside Photoshop, Premiere, Express and other Adobe products, while the company can also sell separate Firefly applications, enterprise services and additional credits. Adobe's overall subscription business is measured in tens of billions of dollars of ARR, so even modest changes in retention, upgrading or customer acquisition can create more value than many standalone AI startups generate in total.
Atlassian offers another useful angle. The company recently said customers using its Rovo AI products were increasing their spending at more than twice the rate of customers who were not using Rovo. That is exactly the kind of effect incumbents want: AI does not have to generate every dollar directly if it makes the existing customer relationship expand faster.
The risk sits in cost. Microsoft has repeatedly said its AI infrastructure investment is weighing on cloud gross-margin percentages. Canva's recent experience also shows that putting expensive AI into a product with hundreds of millions of users can become painful very quickly.
Still, incumbents have a major advantage that AI-native startups would love to own: the customer is already there. They can place AI inside a workflow, expose millions of users to it, learn which features are valuable and then decide whether to charge through higher tiers, credits, consumption or a separate product.
For many established software companies, this could become the biggest generative AI business model in absolute dollars.
Are consulting and implementation becoming part of the generative AI business model?
Yes, especially in enterprise AI, where customers increasingly pay for the deployment work around the model as well as the model itself.
The gap between a good AI demo and a production system remains large. A bank, manufacturer or insurer needs permissions, internal data connections, evaluations, monitoring, security reviews, fallback rules and changes to the way employees actually work. Someone has to build all of that.
Anthropic recently made this direction unusually explicit by helping create a dedicated enterprise AI services company backed by Blackstone, Hellman & Friedman and Goldman Sachs. Anthropic engineers are expected to work with the organization on identifying use cases and deploying custom Claude systems for mid-sized companies.
The application layer has been doing a lighter version for some time. Synthesia's enterprise package includes tailored onboarding, implementation work and dedicated customer-success support. Hugging Face sells enterprise support and service guarantees around models and infrastructure. Infrastructure companies also help large customers optimize and move production workloads.
We would still treat services as a supporting business model rather than the main destination. A company that needs months of custom engineering for every new customer will struggle to produce software margins.
The attractive version is straightforward: implementation revenue gets the customer into production, then software subscriptions, usage or outcome fees keep flowing afterward.
That structure can be powerful because enterprise AI projects are becoming larger and more deeply embedded. OpenAI currently serves more than two million businesses, while Anthropic says more than 1,000 customers are already spending at least $1 million annually. At that level of commitment, customers will often pay significant amounts simply to make deployment work properly.
This chart, featured in our generative AI market deck, illustrates how revenue is distributed across customer segments in the generative AI market
Can ads and shopping become major generative AI business models?
They can, but today advertising and transaction fees are still second-layer monetization for AI platforms that already have huge consumer reach.
ChatGPT is currently the clearest test. OpenAI has moved beyond a small advertising experiment and now supports CPM, CPC and conversion-optimized campaigns through its Ads Manager. Advertisers can pay for impressions, clicks or campaigns optimized toward downstream conversions.
That puts generative AI closer to the economics of search and social platforms. A conversation about "which laptop should I buy?" or "where should I stay in Tokyo?" reveals commercial intent before the user necessarily reaches a merchant website.
OpenAI is also testing a different way to monetize that intent through commerce. Instant Checkout lets eligible users buy products directly inside ChatGPT, while merchants remain responsible for payment processing, fulfillment and the customer relationship. OpenAI charges merchants a fee on completed purchases.
Advertising and transaction fees can coexist because they monetize different points in the same journey. An advertiser can pay to appear while somebody is considering a product; a merchant can pay when the purchase actually happens.
The economics could eventually be huge, but we should keep the current scale in perspective. Subscription, enterprise and API revenue are already proven at billions of dollars. Generative AI advertising is much newer, and agentic commerce still needs broader merchant coverage, payments infrastructure and repeated purchasing behavior.
A new AI startup cannot realistically copy this model just because it has a chatbot. Ads become interesting after the product has millions of engaged users. Commerce becomes interesting after those users regularly arrive with purchasing intent.
For the largest consumer assistants, the question is no longer whether advertising and transactions are possible. The real unknown is how large they can become without making the assistant less useful or less trusted.
So what are the main business models in generative AI today?
Generative AI currently has eight main business models: consumer subscriptions, enterprise seats, API consumption, generation credits, subscription-plus-usage plans, outcome-priced agents, managed AI infrastructure, and AI monetization inside larger platforms through upgrades, ads or transactions.
The evidence points more strongly toward hybrids than toward any single winner.
At the model layer, usage pricing is hard to avoid because computation has a real cost. OpenAI, Anthropic and other model providers can charge directly for the intelligence consumed.
At the infrastructure layer, Fireworks, Together AI and Hugging Face make money by running and customizing models. Fireworks passing $1 billion in annualized revenue shows that this can already support a very large standalone company.
At the application layer, recurring subscriptions and seats remain extremely strong, particularly when AI sits inside work people do every day. Coding, customer service, video creation and enterprise productivity have produced some of the clearest examples. Usage increasingly sits beside those subscriptions so the heaviest users pay more.
Agents introduce a genuinely different possibility. When software can complete a job rather than merely help a person do it, the natural pricing unit starts moving from seats and tokens toward outcomes. Fin and Sierra have already made that model real, although it currently works best for jobs with a clear finish line.
We are also seeing a major split in how cheaper inference affects companies. Falling model costs are dangerous for businesses whose selling price closely tracks raw tokens. They are much more helpful to a company charging for a finished workflow or business outcome. If the cost of resolving a support ticket falls by 70% while the customer still values the resolution at roughly the same amount, the AI vendor keeps much more of the difference.
That is the economic direction we would watch most closely now. The farther a generative AI company can move its revenue away from "how much computation did we use?" and toward "how much useful work did we create?", the better the business tends to become.
| Generative AI business model | What customers pay for | Where it is working now | Our view |
|---|---|---|---|
| Consumer subscription | Monthly access | General assistants, coding and creator products | Large and proven |
| Enterprise seat | Access for each employee | Copilots and workplace AI | Large and still growing fast |
| API consumption | Tokens or model usage | Foundation models and developer platforms | Core model-layer model |
| Generation credits | Shared units of AI creation | Image, video, voice and multimodal products | Becoming standard |
| Subscription + usage | Baseline access plus heavier consumption | Coding, enterprise AI and creator tools | The clearest cross-market pattern |
| Outcome pricing | Completed work or measurable result | Support, sales and narrow agents | Early but genuinely important |
| Managed AI infrastructure | Tokens, compute time or reserved capacity | Open models, inference and specialized models | Already a large standalone business |
| Platform monetization | AI upgrades, ads and transactions | Large incumbent and consumer platforms | Potentially huge, but unevenly mature |
If you want more recent data on this point, please see our latest generative AI market report.
This chart, featured in our generative AI market deck, shows how AI video generation technology has evolved over time
OUR METHODOLOGY
This analysis asks which generative AI business models are actually working today. We separate the market by what triggers the customer's bill: recurring access, seats, model usage, generation credits, additional consumption, completed outcomes, infrastructure capacity, or monetization inside a larger software or consumer platform.
We focused on recent commercial evidence rather than funding rounds or headline valuations. The main inputs were pricing changes, product packaging, paid-seat adoption, revenue and ARR disclosures, usage patterns, enterprise expansion, infrastructure economics, and examples of companies changing their pricing after seeing how real customers consume AI.
We gave the most weight to evidence showing how customers are actually charged and what they are adopting at scale. A pricing page can show that a model exists; paid seats, material ARR, large customer counts, usage growth or a company redesigning its pricing around real cost behavior tell us much more about whether the model is working.
We also looked for patterns across layers of the market. Foundation-model companies, AI applications, multimodal tools, agent platforms, infrastructure providers and large software incumbents have different economics, so we did not assume one pricing model should win everywhere. Where the same structure appeared independently across several categories, we treated that convergence as stronger evidence.
Key sources include OpenAI on ChatGPT Business Premium seats and credits, OpenAI's business pricing, OpenAI's API pricing, Cursor on Teams pricing and power-user consumption, Microsoft's FY2026 Q4 earnings materials on Copilot adoption, Intercom on Fin outcome pricing, Together AI's infrastructure pricing, Runway on generation credits, Synthesia on shared credits, Adobe Firefly pricing, ElevenLabs on API, agent and pay-as-you-go pricing, Hugging Face on dedicated inference pricing, and Epoch AI on historical inference-price declines.
The conclusion is a synthesis of those company-level observations. We use "working" to mean there is credible evidence of real paid adoption, repeatable revenue or meaningful commercial scale, while keeping a distinction between models that are already well established and models such as outcome pricing, advertising and transaction fees that are promising but less mature.
In our generative AI market deck, we identify pain points entrepreneurs should prioritize
Related blog posts
