--- title: "The modern AI pricing dilemma" description: "How to price a generative AI product when inference costs vary with use and the underlying models keep getting cheaper every few months." url: "https://dmassociates.xyz/journal/the-modern-ai-pricing-dilemma/" --- Deep Tech # The modern AI pricing dilemma 24 September 2025 · Last checked 18 October 2026 Building a generative AI company is a high-wire act. On one side, you have unprecedented power to create products that feel like magic. On the other, you face an economic reality that can sink the venture before it finds its footing. This is the modern AI pricing dilemma. The core of the challenge is a tension between two forces. First, your **cost of goods sold (COGS)** is unlike anything in traditional SaaS. It is high, variable and often unpredictable, driven by the computational expense of model inference. Every user action can have a direct, non-trivial impact on your cloud bill. Second, you are building on rapidly evolving technology where features can be replicated overnight. The foundation models that power your product are in a relentless race to become cheaper and more capable, which puts constant downward pressure on prices and threatens to **commoditise** your core offering. Getting pricing wrong in this environment is not a minor misstep. A flawed model can either bankrupt you on costs or render you irrelevant in the market. That is why your pricing strategy is more than a line on a webpage: it is a component of your competitive moat. What follows is a set of frameworks for building a pricing model that is profitable, sustainable and defensible. ## Chapter 1: the foundation, mastering your AI COGS Before you can price your product, you need an unflinching, granular understanding of what it costs to deliver. In traditional SaaS, COGS is often dominated by predictable expenses like hosting and support staff. For an AI company the equation is different and far more dynamic. Your primary cost driver is **model inference** : the act of running a model to generate a response. Mastering your AI COGS means breaking it into three components. ### 1. Direct model costs This is the most obvious expense. It is what you pay your model provider, or the cost to run an open-weights model on your own infrastructure. To calculate it accurately, go beyond headline figures and understand: - Tokenomics. What are the precise costs for input tokens , the data you send to the model, versus output tokens , the data the model generates? For many use cases the input context is significantly larger than the output, which is a crucial detail in cost forecasting. - Multimodality costs. If you are processing images, audio or video, those models have entirely different pricing structures, for example per image or per second of audio, and must be modelled separately from text. ### 2. Infrastructure overhead Running a production-grade AI application involves more than API calls. This layer is often overlooked in early-stage forecasts but becomes significant at scale. It includes: - Data processing and pipelines. The cost of running services for your retrieval-augmented generation (RAG) pipeline: vector databases, embedding models and data chunking processes. - Hosting and orchestration. The servers and services that manage your application logic, handle user requests, and orchestrate calls to various models and databases. ### 3. Hidden and ancillary costs These do not fit neatly into the first two buckets but are needed for a complete picture: - Fine-tuning expenses. The one-time or ongoing cost of training a custom version of a model on your proprietary data. - Monitoring and logging. The services required to track model performance, identify errors, and log inputs and outputs for compliance and debugging. Failing to model these accurately is a common pitfall. Before proceeding, you should be able to answer one question with confidence: what is the fully loaded cost of a single, typical user interaction on my platform? Without that number, any pricing strategy is guesswork. ## Chapter 2: the north star, aligning with value-based pricing With a firm grip on your costs, the next step is to determine your price. Most founders instinctively gravitate towards one of two simple models: **cost-plus** , calculating your COGS and adding a margin, or **competitor-based** , looking at what others charge and setting a similar price. In the generative AI market, both tend to fail. **Cost-plus pricing** anchors your value to your expenses rather than to your customer’s success. It is a race to the bottom, forcing you to compete on efficiency rather than on what you build. **Competitor-based pricing** assumes your rivals have a viable strategy and that your product delivers exactly the same value. Both are dangerous assumptions in a fast-moving market. There is a better way. The more sustainable strategy is **value-based pricing** . Value-based pricing flips the equation. Instead of looking inward at your costs or sideways at your competitors, it looks outward at your customer. It anchors your price to the quantifiable economic value your product delivers. That value could be: - Increased revenue. Helping a sales team close more deals. - Reduced costs. Automating a manual workflow and saving labour hours. - Mitigated risk. Detecting fraud or security threats more effectively. - Enhanced productivity. Allowing a researcher or analyst to accomplish in minutes what used to take hours. Adopting a value-based mindset may be the single most useful strategic shift an AI founder can make. It forces you to stop thinking of yourself as a seller of technology, of API calls and tokens, and start thinking as a partner who sells business outcomes. It requires a deeper understanding of your customer, and in exchange it lets you capture a fair share of the value you create. ## Chapter 3: the keystone, finding your unit of value Value-based pricing is the destination. The path there runs through a single question: **what is the specific unit of value your customers are actually buying?** Answering this is the keystone of the whole strategy, and many early-stage founders get it wrong. They default to pricing the thing that is easiest to measure, which is the underlying technology. They charge per token, per API call, or per minute of processing. That is a trap. It forces your customer to think about your costs rather than their benefits. To align your price with customer value, you first have to define what that value is. In practice it almost always falls into one of three categories. ### 1. Pricing outputs Here you charge for a discrete, tangible result generated by your product. This is the most direct translation of AI capability into a commercial unit. - Examples: a generated image, a completed marketing report, a translated document, a block of clean code. - When it works best: products where the value is transactional and the output is the primary reason for using the tool. - The risk: it can feel like a meter, causing customers to worry about every click. It also couples you directly to the model’s limitations. If an output is flawed, the customer feels they paid for nothing. ### 2. Pricing outcomes This is the aspiration for enterprise sales. Instead of charging for the tool’s activity, you charge for the business result it achieves. It requires a deep understanding of your customer’s P&L. - Examples: a percentage of fraud detected and prevented, a measurable increase in lead conversion rates, documented hours of manual labour saved. - When it works best: high-ACV, enterprise-focused products where a clear return calculation can be built into the sales process. - The challenge: it is the hardest to measure and usually requires a more consultative sale to establish the metrics of success. ### 3. Pricing access Here you are not selling individual results but the ongoing integration of an AI assistant into a key workflow. The value is in continuous productivity gains. - Examples: a co-pilot for a software developer, a research assistant for a financial analyst, an intelligent partner for a graphic designer. Usually sold as a flat per-seat, per-month fee. - When it works best: productivity and workflow tools where the value is ambient and consistent rather than tied to specific generations. - The risk: you have to manage the power-user problem, where one user generates costs far exceeding their seat price. That requires internal cost controls, covered below. Answer this for your business before going further. Are your customers paying for a specific **output** , a measurable **outcome** , or continuous **access** to a new capability? Your answer determines which pricing models are viable. ## Chapter 4: AI-native pricing models, trade-offs and traps Having identified your core unit of value, you can select the commercial model to deliver it. This is not just a packaging decision. It is a strategic choice that affects your sales motion, customer adoption and unit economics. Here are the most common models, with their hidden risks. ### Model 1: pure usage-based - The pitch: “It is the fairest model. Customers pay for exactly what they use, and it aligns our revenue with our costs.” - The hidden trap: this model creates budget unpredictability for your customers. For an enterprise, an unknown, variable bill is a significant barrier to adoption. It can also perversely incentivise your champions to use your product less in order to control their spending, which throttles your growth. - Strategic alignment: best suited to API-first products or developer tools where consumption is a clear proxy for value. It works well when your unit of value is a discrete output . ### Model 2: per-seat subscription - The pitch: “It is simple, predictable, and the standard for B2B SaaS. Customers understand it and our finance team can forecast revenue.” - The hidden trap: this model is often disconnected from the variable costs and value of an AI product. A single power user can generate orders of magnitude more cost and value than an average user, damaging your unit economics while their seat price stays the same. - Strategic alignment: works best for AI products that price access , especially collaborative tools where value genuinely scales with the number of users on a team. This model requires the cost-control guardrails in the next chapter. ### Model 3: tiered subscription with usage gates - The pitch: “It provides a clear upgrade path, letting us serve different segments from self-serve to enterprise.” - The hidden trap: if the tiers rest on arbitrary usage limits, for example 100 reports a month, customers will keep hitting a paywall that interrupts their workflow. The value difference between tiers has to come from features that unlock new capabilities, not just more volume. - Strategic alignment: a flexible model that adapts to both outputs and access . The key is making each tier feel like a genuine step up in value, by unlocking advanced features, collaboration or security, rather than a higher meter limit. ### Model 4: hybrid - The pitch: “Best of both worlds. We get the predictability of a subscription with the upside of a usage-based component.” - The hidden trap: it can be complex to communicate to customers and difficult to instrument on the back end. Designed carelessly, you end up with the complexity of a usage model and none of the revenue upside. - Strategic alignment: often the end state for maturing AI companies selling to the enterprise. It typically involves a predictable platform fee granting access , with a usage-based overage component for high-volume outputs or quantifiable outcomes . The customer gets budget certainty, and you grow with your most successful accounts. ## Chapter 5: the playbook, taming cost volatility You have defined your value and chosen a model. Now comes the most pressing operational challenge in generative AI: how do you offer a predictable price to customers when your own costs are wildly variable? Ignoring this is how you end up selling a $100 a month subscription to a power user who costs you $2,000 a month in inference fees. Solving it takes two prongs: product architecture, and commercial guardrails. ### Architectural solutions, in-product controls These are controls you build into the product to manage costs without degrading the user experience. - Model tiering. One of the most powerful and user-friendly strategies. Instead of relying on a single expensive frontier model for every task, architect your application to use the right model for the right job, and give either the user or your own routing logic a choice: - Performance tier: a frontier model for high-stakes, complex tasks where quality is paramount. - Speed and cost tier: a smaller, cheaper model, or an open-weights model you host yourself, for routine, high-volume tasks like summarisation or data extraction. - Intelligent caching. Many user requests are repetitive. Rather than running a costly inference call every time for an identical query, add a caching layer. Storing the results of common requests reduces redundant model calls and lowers your COGS. ### Commercial guardrails, in-deal controls These are the terms and structures you build into your pricing plans to protect your unit economics, especially at the enterprise level. - Predictable tiers with fair-use caps. You can offer plans that feel abundant, even unlimited, while still protecting yourself from extreme outliers. A fair-use policy lets you define reasonable limits that affect only the most demanding couple of per cent of users, prompting a conversation to move them to a more appropriate enterprise plan. - Reserved capacity and credit packs. Instead of an open-ended usage bill, the customer pre-purchases a large volume of credits, which can correspond to tokens, reports or another unit of value, at a discount. They get budget predictability. You get cash upfront and a clear ceiling on your cost exposure for that account. No single solution covers everything. The more resilient AI companies combine architectural and commercial strategies into a system that can absorb cost volatility. ## Conclusion We have moved from the mechanics of **COGS** to the north star of **value-based pricing** , established the importance of defining your **unit of value** , worked through the trade-offs of the main **commercial models** , and set out a tactical approach to **taming cost volatility** . The most important point is this: pricing is not a static decision, it is a dynamic capability. In a market moving at the velocity of generative AI, your pricing model has to be as agile as your product roadmap. Your first attempt will not be right. The goal is a thoughtful starting point, then relentless iteration based on customer conversations, usage data and market feedback. Your pricing strategy is the connective tissue between the value you create and the venture you build. Written by Dineshwara Manideepu, founder of DM & Associates. [How we work with deep tech founders →](https://dmassociates.xyz/advisory/deep-tech/) [More from the journal](https://dmassociates.xyz/journal/) [Book a call](https://bookings.dmassociates.xyz/#/362591000000049559) ## Related reading - Deep Tech · Sept 2025 [Why trust is the essential moat for AI startups](https://dmassociates.xyz/journal/why-trust-is-the-essential-moat-for-ai-startups/) When model performance becomes a utility, defensibility shifts to reliability, transparency and fairness. How to build and measure that. - Deep Tech · Sept 2025 [Your first 100 users: a go-to-market playbook for vertical AI startups](https://dmassociates.xyz/journal/your-first-100-users-a-go-to-market-playbook-for-vertical-ai-startups/) Why SaaS growth playbooks fit vertical AI poorly, and what to do instead: one critical workflow, deep integration, and an enabled internal champion. - Deep Tech · Sept 2025 [How to acquire your minimum viable dataset efficiently](https://dmassociates.xyz/journal/how-to-acquire-your-minimum-viable-dataset-efficiently/) Eight capital-efficient plays for acquiring the smallest useful dataset, and how each one evolves into a compounding data flywheel.