New Business Models

In eighteen months, Salesforce changed how it charged for its flagship AI product three separate times. When Agentforce launched, the company billed two dollars per conversation. By May 2026 it had switched to "Flex Credits," ten cents per action. Then came per-user licenses at $125 a month—and, for its customer-service Help Agent, a fourth option: two dollars every time the agent resolved a ticket on its own, with no charge if the customer asked for a human instead. This is not a startup fumbling in the dark. Salesforce is a company with three decades of enterprise pricing expertise and a market capitalization in the hundreds of billions. If it cannot decide how to charge for AI, that tells you something about how unsettled the question is.

Contrast that with Midjourney, which turns text prompts into images. In 2025 it crossed $500 million in revenue, up two-thirds from the year before, serving nearly twenty million registered users (Sacra; SQ Magazine, 2026). It did this with almost comic simplicity: pay a monthly fee, generate images, and roughly 85 percent of the money comes straight from those subscriptions. No ads. No usage meter. Not even a free tier—the company killed it in 2023 after abuse and never looked back. One company has cracked its pricing so cleanly it barely thinks about it; another, far larger and more experienced, keeps rewriting the rules every quarter.

That gap is the subject of this chapter. AI has attracted one of the largest waves of capital in the history of technology, and yet the industry still has not agreed on how to turn that technology into durable revenue. The money pouring in and the clarity about how to make money back are wildly out of sync. Understanding why—and where the pricing models are converging, where they are fracturing, and who wins under each—is the task ahead.

The investment is enormous; the monetization is not settled

Start with the scale, because it is genuinely hard to overstate. Gartner projected worldwide AI spending at $2.59 trillion in 2026, a 47 percent jump over the prior year, most of it flowing into the data centers and chips that hyperscalers are building at a furious pace (Gartner, 2025). Generative AI specifically—the chatbots, coding assistants, and image tools that most people mean when they say "AI"—accounts for a smaller but faster-growing slice: enterprise spending on it rose from $1.7 billion in 2023 to $11.5 billion in 2024 to $37 billion in 2025, and analysts expect it to keep compounding at roughly 60 percent a year (Menlo Ventures, 2025).

Now set that against a more sobering number. In Menlo Ventures' 2025 enterprise survey, only 39 percent of companies could attribute any measurable operating-profit impact to their AI spending, and a small group of "high performers"—around six percent of firms—captured a disproportionate share of whatever value was created. Enormous investment, uneven return. The money has arrived; the proof that it pays for itself, at scale, largely has not.

This is the central tension of AI's business models, and it explains almost everything downstream. When customers are still testing whether AI earns its keep, they resist committing to prices. When vendors cannot predict their own costs, they resist committing to models. The result is an industry spending trillions on infrastructure while arguing, deal by deal, over how to charge a few dollars for the output.

Why the software playbook breaks

To see why AI pricing is so contested, you have to understand the model it inherited—and why that model no longer fits.

Traditional software runs on a beautiful economic trick. You spend heavily to build the product once, and then serving the millionth customer costs almost nothing more than serving the thousandth. This is why healthy software companies have historically posted gross margins of 70 to 90 percent: the marginal cost of one more user rounds to zero. Scale is pure profit. It is the single fact that made the cloud-software industry one of the most lucrative in modern capitalism.

AI does not work this way. Every time a model answers a question, it performs a fresh round of computation—an operation called inference—that consumes real GPU time, real electricity, real cloud capacity. The cost does not vanish at scale; it grows with every query. A traditional software company that doubles its users barely notices on the cost side. An AI company that doubles its users doubles a very real bill.

There is a subtlety here that trips up even careful observers. Inference is getting dramatically cheaper: the per-token price of frontier models fell somewhere between 60 and 75 percent across 2025 alone, driven by better chips, smarter model architectures, and brutal price competition (SFAI Labs, 2026). You might think that would rescue margins. It has not, and the reason is instructive. As models got cheaper per call, product teams simply made more calls—adding retrieval steps, self-checking passes, and intent-classification layers to make features smarter. The number of calls per task grew faster than the price per call fell. Across late-2025 and early-2026 earnings reports, public software companies began describing a new normal: gross margins of 60 to 70 percent, down from the old 70-to-90 range, and they were clear that the compression is structural, not a passing phase (SFAI Labs, 2026). Making AI central to a product permanently changes its cost shape.

The token economy

The unit of account in this new economy is the token—roughly a fragment of a word, four characters or so. Modern AI interfaces charge by the token, and at first glance that seems simple. It is not, because not all tokens cost the same to produce.

Input tokens—the text you send the model—are comparatively cheap, because the model can chew through them in parallel. Output tokens—the text it writes back—cost several times more, because they are generated one at a time, each new word depending on the last. On Anthropic's mid-2026 lineup, its Opus model cost five dollars per million input tokens but twenty-five dollars per million output; Google's Gemini Pro ran two dollars in, twelve dollars out (BenchLM, 2026). A rough four-to-five-times premium on output is the industry norm.

Then reasoning models added a third category. These systems "think" before they answer, generating a hidden internal monologue that the user never sees but nonetheless pays for, billed at the output rate. Because that invisible deliberation can run to thousands of tokens, a reasoning model's real cost can land three to ten times above its advertised price (BenchLM, 2026). Layer on cached-input discounts for repeated prompts, volume deals for big buyers, and the different tiers across a dozen model families, and you arrive at a pricing surface that is genuinely difficult to reason about.

For a company building a product on top of these interfaces, this is not an academic problem—it is an existential one. Your own costs move with usage patterns you cannot fully see or control. A single user who leans on a reasoning-heavy feature can cost you more than a hundred casual users combined. Pricing your product becomes a wager on the behavior of customers you have not met yet, using a cost structure that shifts every time your upstream provider ships a new model. It is monetization built on quicksand.

Three segments, three answers

For all the confusion, the market has not stayed formless. Distinct segments have settled on distinct dominant models, and the logic of each is worth understanding, because it reveals where AI pricing is actually converging.

Segment	Dominant model	Representative players	Why it fits
Consumer	Flat monthly subscription	Midjourney, ChatGPT Plus	Frequent, casual use; buyers value predictable cost over precision billing
Developer	Usage-based, per token	OpenAI, Anthropic, Google APIs	Technical buyers accept metered cost; price scales with their own activity
Enterprise	Per-seat licenses and negotiated contracts, increasingly per-outcome	Salesforce, Microsoft, Zendesk	Matches procurement norms; large deals absorb complexity

Consumers overwhelmingly want a flat fee. Midjourney's twenty-dollar-a-month simplicity, ChatGPT Plus at the same price—these work because a person using an image or chat tool a few times a day does not want to watch a meter run. Predictability is the product feature. Developers, by contrast, live comfortably with metered pricing: they understand tokens, they build the cost into their own margins, and usage-based billing lets them start small and scale spending only as their own product grows. Enterprises want something that fits how they already buy software—annual contracts, per-seat licenses, procurement cycles—which is why the big incumbents default to those structures even for AI.

The interesting movement is at the enterprise edge, where a genuinely new model is being tested in public.

The marketplace and the outcome-pricing experiment

The theoretical ideal of AI pricing has always been obvious: charge for the value delivered, not the resources consumed. If an agent does the work of an analyst, bill a fraction of what the analyst would have cost, and bill nothing when it fails. Cost tracks benefit; both sides share the risk. Economists have wanted this for decades across all of software. AI, because its agents can autonomously complete whole tasks, is the first place it looks genuinely achievable.

In 2026, several companies stopped theorizing and started shipping it. Salesforce's Agentforce Help Agent charges two dollars for each customer issue it resolves entirely on its own—and crucially, nothing when the customer gives negative feedback or asks for a human (Constellation Research, 2026). Zendesk bills $1.50 per automated resolution; Intercom's Fin agent charges ninety-nine cents. This is outcome-based pricing in the wild, and its arrival is not an accident. It is a direct response to enterprise anxiety about unpredictable token costs: buyers would rather pay a fixed price for a definite result than an open-ended meter for uncertain effort. The shift is being felt in adoption, too—one Salesforce survey found AI agents handling customer service jumped from 39 percent of respondents in 2025 to 66 percent in 2026.

Yet notice how narrow these examples are. Every one of them involves a task with a clean, bright-line definition of success: the support ticket was resolved, or it was not. That is precisely why outcome pricing works here and almost nowhere else. The moment the outcome becomes fuzzy, the model collapses. Consider an agent that drafts a contract. If a lawyer still spends two hours reviewing the draft, how much of the value belongs to the agent? If the same document would have taken a senior associate fifty hours but a junior one a hundred and fifty, which baseline sets the price? Outcome pricing requires agreeing on a counterfactual—what would have happened without the AI—and counterfactuals are, for most real work, unknowable and endlessly disputable. Absent a metric both parties trust and can audit, outcome pricing turns every invoice into a negotiation. So it flourishes in customer support, where "resolved" is measurable, and stalls everywhere the value is real but diffuse.

The agent marketplace is the structural bet layered on top of all this. Google, AWS, Microsoft, and Salesforce have all built platforms where developers list autonomous agents and businesses buy them on demand, with the operator taking a cut of ten to thirty percent, much like an app store. The model is legible, which is its main virtue. But it inherits the classic two-sided-market problem: developers will not build for an empty marketplace, and buyers will not shop an empty catalog. The platforms are papering over the gap by seeding their stores with first-party agents and subsidizing early developers. Whether that bootstraps a genuine ecosystem before the market consolidates around a few winners is still an open question—and how to price any given agent (by subscription, by task, by outcome) remains unresolved inside these marketplaces just as it is outside them.

Why the newcomers are beating the giants

One of the most striking patterns of the past two years is that AI-native startups are outcompeting far larger incumbents on the incumbents' own ground. At the generative-AI application layer, startups captured 63 percent of the market in 2025, up from 36 percent the year before, earning nearly two dollars in revenue for every dollar the established players took (Menlo Ventures, 2025). The best AI startups now generate revenue per employee five to six times higher than a typical leading software firm, which hovers around $610,000 per head (Menlo Ventures, 2025).

The reasons are structural, not a matter of hustle. A startup building around AI from day one designs its product, its pricing, and its cost model to fit inference economics natively. It has no installed base to protect, no legacy price tiers to keep consistent, no old sales process to unlearn. When inference costs drop, it passes the savings on or reinvests them immediately. When a new model capability opens a use case, it pivots without a committee. The coding-assistant market shows the pattern cleanly: Cursor beat Microsoft's Copilot to features like repo-wide context and multi-file editing simply by shipping faster, and coding became the single largest category of departmental AI spending at four billion dollars in 2025 (Menlo Ventures, 2025).

Incumbents face a genuine bind, not merely inertia. Give AI away free, and you cannibalize the paid features customers already buy. Charge extra, and you invite backlash and hand competitors an opening. Spin up a separate AI product, and you risk devouring your own core revenue. Every option threatens an existing revenue stream, so the rational corporate response is to hesitate—and hesitation, in a market moving this fast, is how you lose. This is why the survey data shows so many large firms bundling AI in quietly or charging a cautious premium while startups ship and learn.

Whether this advantage lasts is a live debate, and honesty requires flagging it. The current startup dominance could be structural—small, focused firms may simply be better suited to a fast-moving technology—or it could be an artifact of the hype cycle, the kind of froth that recedes when incumbents finish retooling and bring their distribution, data, and balance sheets to bear. Two facts cut against overconfidence. Incumbents still hold 56 percent of the infrastructure layer, where switching costs are high and trust matters most. And analysts widely expect that somewhere between 70 and 90 percent of today's AI startups will fail or be absorbed at depressed valuations within eighteen months as the market corrects for over-investment and thin moats (FindNStart, 2026). Out-executing a giant this quarter is not the same as building something the giant cannot eventually copy.

The margin squeeze and the capital wall

Underneath the revenue race sits an uncomfortable question: are these companies actually making money? For the frontier model builders, the answer is a moving target, and the direction of travel is the most important story in the industry.

The trend is genuinely encouraging. OpenAI's compute margin—the profit left after paying to run the models, before the enormous costs of training them—climbed from roughly 35 percent in early 2024 to around 70 percent by late 2025 (SaaStr, 2026). Anthropic's trajectory is even more dramatic: from a gross margin of negative 94 percent in 2024, meaning it spent nearly two dollars serving every dollar of revenue, to an expected positive 50 percent in 2026 and a projected 77 percent by 2028 (SaaStr, 2026). Efficiency gains, better hardware, and scale are pulling the unit economics from catastrophic toward genuinely healthy.

But compute margin is not the whole ledger, and the top-line losses remain staggering. OpenAI reached roughly $25 billion in annualized revenue by early 2026 while running an operating margin of around negative 122 percent—spending more than two dollars for every dollar it booked, once the cost of training the next generation of models is counted (ValueAdd VC, 2026). The path to profit runs through inference margins that are improving faster than almost anyone predicted, colliding with training and infrastructure costs that are also climbing faster than almost anyone predicted. Which curve wins is the multi-hundred-billion-dollar question, and no one honestly knows the answer yet.

This dynamic sorts the entire industry by access to capital. A company can chase sustainable margins by owning its infrastructure—building or leasing data centers, striking chip-supply deals, hiring hardware engineers—to drive down per-query cost over time. That path costs hundreds of millions to billions up front and takes years to repay, and only a handful of players can even attempt it. Everyone else rents capacity from third-party clouds, accepting structurally thinner margins in exchange for keeping their capital for product. Most startups rationally choose renting, which works until scale makes the margin gap punishing. The uncomfortable implication is that the capital intensity of AI infrastructure may quietly determine the shape of the market: if durable margins require billions in fixed investment, the number of companies that can reach sustainable scale is small, and the industry drifts toward a concentrated structure regardless of how many startups bloom in the meantime. This is the distribution question the book returns to again and again—here it appears as a question of who can afford the compute to survive.

How much of this can we actually believe?

A note of epistemic caution is warranted, because the numbers in this chapter come with real caveats. Most of the leading AI companies are private. Their revenue figures are not audited financial statements; they are annualized run rates—a single strong month multiplied by twelve, or contracted commitments extrapolated forward. These measures can flatter reality considerably.

The gap is not small. The difference between contracted revenue and recognized revenue at AI startups can reach 70 percent (Lambda Finance, 2026). The definitional slack is wide enough to fuel open disputes: when Anthropic's run rate was reported at $30 billion in early 2026, OpenAI's own chief revenue officer circulated a memo arguing the figure was overstated by roughly eight billion dollars (The AI Corner, 2026). Both companies are now growing at rates—Anthropic reportedly reached $47 billion in annualized revenue by mid-2026, up from one billion eighteen months earlier—that would be extraordinary if the underlying accounting were solid, and merely impressive-sounding if it is not. Read every headline revenue figure in this space, including the ones in this chapter, as a directional signal rather than a settled fact. The growth is real; the precision is not.

We have been here before

If the current confusion feels unprecedented, it helps to remember that every major platform shift has passed through a period exactly like it. The early commercial internet had no idea how to make money. Through the dot-com years, advertising became the default business plan precisely because no one had a better one, and marketing spending reached absurd heights—between 90 and 133 percent of revenue in 1999, meaning companies spent more attracting customers than those customers ever paid (Profit over Privacy, Univ. of Minnesota Press). The models that eventually stabilized the web did not exist yet. Search advertising, the engine that would fund the entire consumer internet, only came into its own in the early 2000s when Google made pay-for-performance work.

Stabilization, when it came, came fast. By 2002 marketing's share of dot-com revenue had fallen to a sane 14-to-37 percent, and the industry settled toward a durable 20-to-25 percent within a few years. The lesson is double-edged. Pricing chaos in a young platform market is normal and temporary—but the model that resolves it may not yet be visible, and the resolution can wipe out most of the companies experimenting along the way. AI's monetization mess is not evidence that the technology is a mirage. It is evidence that AI is a genuine platform shift, running the same gauntlet the web and mobile ran before it, on a compressed timeline.

Who gets left out

There is a fairness dimension to all of this that deserves to be named plainly. Usage-based and infrastructure-heavy pricing models systematically favor the large over the small. A big enterprise negotiates volume discounts, absorbs unpredictable token bills as a rounding error, and can afford to build or reserve dedicated compute. An individual developer or a small business gets list prices, feels every spike in usage, and has no capital to buy its way to better margins. When the sustainable version of a technology requires billions in fixed investment, the economics tilt toward whoever already has billions.

The consumer subscription model is, in this light, the most egalitarian corner of the market: a flat twenty dollars a month is the same twenty dollars for everyone, and it has made frontier image and language tools genuinely accessible to hundreds of millions of people. But the deeper infrastructure—the custom silicon, the proprietary data centers, the negotiated enterprise deals—runs the other way, concentrating capability among those who can pay to own the means of computation. Whether AI's business models widen access or entrench advantage is not a technical question. It is a choice the industry and its regulators are making, mostly by default, right now.

The road ahead

What, then, is the most plausible trajectory for the next five years? The honest answer is segmentation rather than a single winning model. Consumers will keep paying flat subscriptions. Developers will keep paying by the token. Enterprises will keep signing negotiated contracts, with outcome-based pricing spreading into every corner where success can be cleanly measured and stalling everywhere it cannot. Expect the token meter and the per-outcome charge to coexist, often inside the same company, the way Salesforce already runs several models at once—not because anyone loves the complexity, but because different buyers genuinely need different things.

The one force that could override this segmentation is margin pressure. If inference costs keep falling faster than usage climbs, current models become sustainable and the experimentation slowly settles. If they do not—if the calls-per-task ratchet keeps outrunning the price-per-call decline—then something more fundamental has to give, and the resolution comes through consolidation, as the companies that cannot reach viable margins are bought or buried by the few that can afford the infrastructure to get there. That is the genuine unknown at the center of AI's business models. Not which pricing label wins, but whether the underlying economics can be made to work at all without collapsing the market down to a handful of players who own the compute. We will know within a few years. We do not know now, and anyone who tells you otherwise is selling something—possibly on an annualized run-rate basis.

Summary

AI has attracted one of the largest capital waves in the history of technology—worldwide AI spending was projected at $2.59 trillion for 2026—yet the industry has not agreed on how to turn that investment into durable profit. Only 39 percent of enterprises could attribute any measurable profit impact to their AI spending in 2025. The money and the proof that it pays for itself are badly out of sync, and that mismatch drives everything else.

The root cause is that AI breaks the economics of traditional software. Conventional software costs almost nothing to serve one more customer, which is why it enjoys 70-to-90 percent margins. AI performs fresh, costly computation on every query, so costs grow with usage. Even though per-token prices fell 60 to 75 percent in 2025, margins compressed anyway, because products added more model calls per task faster than prices dropped—pushing the new software-margin norm down to 60-to-70 percent as a structural, permanent shift.

The market has sorted into segments with distinct dominant models: flat subscriptions for consumers, per-token usage pricing for developers, and negotiated per-seat contracts for enterprises—with outcome-based pricing newly emerging where success is cleanly measurable, as in customer support, where agents now charge one to two dollars per resolved ticket. Outcome pricing remains confined to those bright-line cases because measuring AI's value against an unknowable counterfactual is, for most work, intractable.

AI-native startups captured 63 percent of the application-layer market in 2025, earning nearly two dollars for every dollar incumbents made, because they build around inference economics from the start while incumbents are paralyzed by the risk of cannibalizing existing revenue. Whether that lead is structural or a hype-cycle artifact is genuinely contested, especially as most current AI startups are expected to fail within eighteen months.

Frontier model margins are improving fast—OpenAI's compute margin roughly doubled to 70 percent, Anthropic swung from deeply negative toward profitability—but top-line losses remain enormous, and sustainable margins increasingly depend on billions in infrastructure investment that only a few players can afford. Combined with revenue figures that rest on unaudited run rates rather than audited accounts, the picture is one of a genuine platform shift running the same monetization gauntlet the early internet ran—normal, temporary, and likely to consolidate the market around whoever can afford the compute to survive.

Key Takeaways

Investment and monetization are severely out of sync: worldwide AI spending was projected near $2.59 trillion for 2026, yet only 39 percent of enterprises could attribute any profit impact to their AI spending in 2025.
AI breaks the SaaS model because every query triggers costly inference, so costs rise with usage; despite per-token prices falling 60–75 percent in 2025, margins compressed to a new structural norm of 60–70 percent as products added more model calls per task than price cuts could offset.
Token pricing is genuinely complex: output tokens cost roughly 4–5x more than input, and reasoning models' hidden "thinking" tokens can push real costs 3–10x above headline rates—making it hard for companies building on AI APIs to price their own products.
The market has segmented by dominant model: flat subscriptions for consumers (Midjourney, ChatGPT Plus), per-token usage pricing for developers, and negotiated contracts for enterprises—with outcome-based pricing (Salesforce, Zendesk, Intercom charging ~$1–2 per resolved ticket) emerging only where success is cleanly measurable.
Outcome-based pricing stays confined to bright-line tasks because measuring AI's value requires agreeing on an unknowable counterfactual; for most diffuse work, it turns every invoice into a dispute.
AI-native startups took 63 percent of the application-layer market in 2025, earning nearly 2x incumbents' revenue, because they build around inference economics natively while incumbents fear cannibalizing existing revenue—though 70–90 percent of AI startups are expected to fail or be acquired within 18 months.
Sustainable margins increasingly require billions in infrastructure investment (data centers, custom chips), which only a few players can afford—pushing the market toward consolidation regardless of near-term startup dynamism.
Revenue data is unreliable: most leading AI firms are private, and headline figures are annualized run rates, not audited financials—the gap between contracted and recognized revenue can reach 70 percent, and rivals publicly dispute each other's numbers.
AI's "monetization mess" mirrors the early internet, where advertising was a default plan and marketing consumed 90–133 percent of revenue in 1999 before stabilizing within a few years—signaling a real platform shift running a familiar, temporary gauntlet.

Sources

Last updated: 2026-07-12

V2 (in progress) Previous: V1