1.3.3 New Business Models

In 2025, Midjourney—a company that generates images from text prompts—crossed $500 million in annual revenue. They did it without ads. Without a marketplace. Without venture capital. Just subscriptions: pay a monthly fee, generate unlimited images.

Runway, a similar company for AI video generation, hit $70 million in annualized revenue the same year. Same model: subscribe, create, repeat.

These are not the business models we're used to. No freemium funnel with conversion optimization. No ad-supported platform harvesting user data. No complex enterprise sales cycle. Just a product people want, priced simply, distributed globally through the internet.

AI is spawning new ways to make money. Some look familiar—subscriptions, usage-based pricing, marketplaces. Others are genuinely novel, built around the strange economics of models that get cheaper to run as they scale, tokens that cost different amounts depending on how they're used, and agents that operate autonomously in markets we're still figuring out how to regulate.

The Scale of Investment

Companies spent $37 billion on generative AI in 2025, up from $11.5 billion in 2024—a 3.2x year-over-year increase. Projections suggest global AI spending will reach $630 billion by 2028, with generative AI growing at roughly 60% compound annual growth rate. Ninety-two percent of executives plan to boost AI spending over the next three years.

The paradox underlying this boom is striking: while spending is soaring, AI still represents less than 1% of total software application spending. The money is there, but the commitment is not. Most companies are still in the pilot phase—experimenting, testing, and waiting to see which models and vendors survive before going all-in. For AI companies trying to monetize, this hesitancy creates a difficult environment: enormous interest, but sluggish conversion into reliable revenue.

The Monetization Mess

A survey of AI builders in 2025 found that 29% bundle AI features for free with existing products, 24% charge a premium, and 11% have no monetization strategy at all. That last figure is remarkable given the billions being invested and the valuations in the tens of billions attached to some of these companies.

The confusion is rooted in a fundamental tension between AI's cost structure and the pricing norms it inherited from traditional software. Conventional SaaS operates at 70–90% gross margins because once software is built, serving an additional customer costs almost nothing. The economics reward scale: more customers, same infrastructure, more profit.

AI breaks this model. Every customer interaction requires real computing resources—running the model to generate output (known as inference) consumes GPU time, electricity, and cloud capacity. Costs grow with usage, not trivially but materially. Forty-one percent of AI builders report struggling with cost-effective scaling. The challenge is finding prices that customers will accept without losing money on every transaction—a balance that has proven surprisingly elusive.

The Token Economy

Modern AI APIs charge by the token, a unit roughly corresponding to a word fragment. But the pricing is more nuanced than a simple per-unit rate, because not all tokens impose the same computational burden.

Input tokens—the text sent to the model—are relatively cheap, since the model can process them in parallel. Output tokens—the text the model generates in response—are significantly more expensive, often four to eight times the input rate, because they must be produced sequentially, one at a time. Then there are reasoning tokens, used by models that deliberate internally before responding, which carry their own pricing tier.

Providers layer additional complexity on top of this baseline. Cached inputs—repeated prompts sent to the same model—may qualify for steep discounts after the first use. Some providers charge by action rather than token, billing based on outcomes like successful classifications or detected intents. Others offer subscription bundles with overage tiers, or enterprise volume discounts negotiated case by case.

The cumulative effect is a pricing landscape that is genuinely difficult to navigate, especially for companies building products on top of these APIs. When your own costs are tied to usage patterns you cannot fully control, pricing your product becomes an exercise in probabilistic guesswork.

The Marketplace Moment

While individual companies wrestle with pricing, a new structural model is taking shape: the AI agent marketplace. Google, AWS, Microsoft, and Salesforce have all launched or announced platforms where developers list AI agents—autonomous systems capable of tasks like writing code, handling customer inquiries, or processing data—and businesses purchase and deploy them on demand.

The economics resemble traditional app stores. Marketplace operators take a platform cut, typically 10–30% of transaction value, while developers retain the remainder. The model is familiar enough to be legible to buyers and sellers alike, which gives it an advantage over more experimental approaches.

The harder challenges are structural. Pricing individual agents remains unsettled territory: should an agent that automates document review be sold by subscription, by document processed, or by hours of analyst time saved? Each model makes sense under different assumptions, and buyers and sellers rarely agree on which assumptions apply. Platforms that rely exclusively on transaction fees tend to see lower user retention than those offering subscriptions—studies suggest subscription-based platforms retain users at rates roughly 40% higher—but subscriptions are not a natural fit for every agent type.

Then there is the classic two-sided market problem. Developers will not build for a marketplace without a critical mass of potential buyers, and buyers will not visit a marketplace without a useful catalog of agents. The major platforms are addressing this by subsidizing early developers, seeding catalogs with first-party agents, and offering free tiers to reduce friction on both sides. Whether this is enough to overcome the bootstrapping problem before market consolidation sets in remains to be seen.

The AI-Native Advantage

One of the more striking findings of recent years is the degree to which AI-native startups are outcompeting established incumbents on their own turf. In 2025, AI-native companies captured 63% of the generative AI market, up from 36% the year before, and generated nearly two dollars in revenue for every dollar earned by legacy players.

The reasons are structural, not merely a matter of energy or focus. AI-native companies build their entire product, pricing, and distribution model around AI's specific economics from day one. They have no installed base to protect, no existing pricing tiers to harmonize, and no legacy sales motions to retrain. When inference costs fall, they can pass savings on immediately or reinvest in product. When a new model capability opens up a new use case, they can pivot without navigating organizational politics. And because they attract talent drawn specifically to AI-first environments, they tend to iterate faster on the technical and business problems that matter most.

Incumbents face the opposite problem. Offering AI for free risks devaluing existing paid features. Charging a premium risks customer backlash and competitive exposure. Creating a separate AI product line requires investment and risks cannibalizing core revenue. The result, for many large companies, is the paralysis visible in the survey data: 29% bundle AI in, 24% charge extra, and the rest are still deciding. That indecision compounds over time, ceding ground to startups that are shipping and learning while legacy players deliberate.

The Value Alignment Problem

The theoretical ideal for AI pricing is outcome-based billing: charge for the value delivered, not the resources consumed. An agent that saves a company 100 hours of analyst work might bill at a fraction of the avoided labor cost—say $50 per hour—yielding $5,000. An agent that delivers no improvement bills nothing. Cost tracks benefit directly, and both parties share the risk of underperformance.

The appeal is clear. The implementation is not. Verifying outcomes requires agreeing on counterfactuals—how long would the work have taken without the AI?—and those counterfactuals are genuinely hard to establish. If an agent drafts a contract and a lawyer still spends two hours reviewing it, how much of the value does the agent deserve credit for? If the same task would have taken a senior associate 50 hours but a junior associate 150, which baseline applies?

Without clear, auditable metrics, outcome-based pricing becomes a source of ongoing negotiation and dispute. In practice, most companies default to simpler structures: usage-based pricing tied to tokens or API calls, flat subscriptions, or per-action transaction fees. These models are predictable and easy to administer, even if they do not align cost with value in any principled way. Outcome-based pricing remains an aspiration—one that will likely become more viable as measurement tools improve and as buyers accumulate enough historical data to set meaningful baselines.

The Margin Squeeze

The implications of AI's cost structure ripple through into company financials in ways that sharply distinguish this industry from prior generations of software. Even companies with substantial revenue often operate on thin margins because inference costs are so high. Major foundation model providers reportedly spend a significant fraction of revenue on compute alone, leaving limited room for profit even at scale.

This creates a strategic fork. Companies can invest in proprietary infrastructure—building or leasing data centers, negotiating chip supply agreements, hiring infrastructure engineers—to reduce per-unit inference costs over time. This path requires hundreds of millions in capital and years to pay off, and it is accessible only to a small number of well-funded players. Alternatively, companies can remain on third-party cloud APIs, accepting structurally lower margins but preserving capital for product development. Most startups choose the latter initially, which is rational in the short term but increasingly constraining as they scale. The companies that figure out how to achieve competitive margins without a custom infrastructure stack—through model efficiency, smarter caching, or novel architectures—will hold an advantage that capital-light competitors cannot easily replicate.

What Is Actually Working

Amid the experimentation, a clearer picture of effective monetization is beginning to emerge. Different models work better in different contexts, and the choice of pricing structure often matters as much as the quality of the underlying product.

Pricing model Best fit Representative examples Core advantage
Subscription Consumer creative tools Midjourney, Runway, Jasper Predictable revenue; users prefer flat-rate access for frequent use
Usage-based Developer tools and APIs OpenAI, Anthropic, Cohere Scales naturally with customer activity; appeals to technically literate buyers
Enterprise custom deals High-value B2B applications AI in CRM, HR, customer service platforms Aligns with procurement norms; accommodates complex requirements
Marketplace / platform Ecosystem plays Google Cloud AI Marketplace, AWS Network effects between builders and buyers create compounding value

What is clearly not working is ad-supported AI. There is no meaningful evidence that users will accept advertising in exchange for free access to generative AI tools—the interaction model is too intimate, the friction too high, and ad-free paid alternatives readily exist. Similarly, purely free offerings fail as a standalone strategy. Free tiers are useful for acquisition and can convert into paid relationships, but absent that conversion they generate compute costs with no offsetting revenue.

The Road Ahead

The near-term trajectory of AI business models is genuinely uncertain. Subscriptions, usage-based pricing, marketplaces, and enterprise deals each have constituencies and use cases where they work well, and there is no reason to expect a single dominant model to emerge across the entire market. What seems more likely is continued segmentation: consumer tools converging on subscriptions, developer infrastructure on usage-based pricing, high-stakes enterprise applications on negotiated contracts, and the agent economy on some hybrid of marketplace and outcome-based billing as measurement tools mature.

What is less uncertain is the scale of the opportunity. The companies that capture durable shares of projected AI spending will likely be those that solve three interconnected problems: aligning pricing with the value actually delivered, achieving gross margins that can sustain long-term investment, and building enough integration depth or network effect that customers do not churn when a cheaper competitor appears. These problems are solvable, but none of them have clean answers yet. The period of experimentation is far from over.

Summary

This chapter has examined how AI is reshaping the economics of technology businesses, and why that reshaping is more complicated than it initially appears.

The core tension is this: AI creates enormous value but is genuinely expensive to operate. The pricing models inherited from traditional software—built on assumptions of near-zero marginal cost—do not transfer cleanly to a world where every inference run costs real money. That fundamental mismatch explains why a meaningful share of AI companies still lack coherent monetization strategies, even as the broader market swells toward hundreds of billions in projected spending.

Different pricing structures have taken hold in different segments of the market. Subscriptions work well for consumer-facing creative tools where users engage frequently and value predictability. Usage-based pricing suits developer tools and APIs, where costs naturally track activity. Enterprise custom contracts remain the vehicle of choice for high-value B2B deployments. Marketplaces for AI agents are gaining traction but face unsolved challenges around pricing, retention, and bootstrapping supply and demand simultaneously.

AI-native companies have significantly outpaced incumbents because they design their products and business models around AI's specific economics rather than retrofitting AI into structures built for a different era. Incumbents are responding slowly, hampered by legacy pricing commitments, existing customer expectations, and organizational indecision.

Outcome-based pricing—charging for value delivered rather than resources consumed—remains the theoretical ideal but is difficult to implement without verifiable, auditable metrics. The margin pressure created by inference costs forces a strategic choice between capital-intensive infrastructure investment and accepting structurally lower margins on third-party APIs, a tension that will define competitive dynamics for years to come.

The broader picture is one of rapid, often chaotic experimentation. Most current approaches will evolve significantly or fail outright. The business models that eventually dominate will be those that solve alignment, margin, and retention simultaneously—and the companies that get there first will shape the AI economy for the decade ahead.

Key Takeaways

  • AI's fundamental cost structure — every inference run consumes real compute, unlike traditional software's near-zero marginal cost — breaks the SaaS model and has left many AI companies without a coherent monetization strategy (11% of AI builders have none at all).
  • Token-based pricing is complex: output tokens cost 4–8x more than input tokens, reasoning tokens carry their own tier, and caching/volume discounts add further layers — making AI pricing a persistent source of uncertainty for companies building on top of APIs.
  • Different models work for different contexts: subscriptions suit consumer creative tools, usage-based pricing suits developer APIs, enterprise custom contracts fit high-value B2B deployments, and marketplaces for AI agents are emerging but face unsolved retention and bootstrapping challenges.
  • AI-native companies captured 63% of the generative AI market in 2025 and generated nearly 2x the revenue of legacy players — because they design business models around AI's specific economics from the start rather than retrofitting AI into structures built for a different era.
  • Outcome-based pricing remains the theoretical ideal but is practically difficult without verifiable, auditable metrics for measuring performance against counterfactuals — most companies default to simpler token- or subscription-based structures.
  • Infrastructure investment (owning or leasing data centers) is the path to sustainable margins, but requires hundreds of millions of upfront capital and years to pay off — accessible only to a small number of well-funded players, leaving most startups with structurally lower margins.
  • Ad-supported AI has not worked and purely free offerings fail as standalone strategies. The business models that endure will solve alignment, margin, and retention simultaneously — a problem that remains largely unsolved across the industry.

Sources:

Last updated: 2026-02-25