An AI startup CEO said something unusual to a reporter: "It comes down to how many tokens our model produces, and how many of those tokens get consumed." This was a revenue formula, offered in a context where most executives would be talking about product vision or technical capabilities. The phrase was "AI tokenomics," and it came from Sung-hoon Kim, CEO of Upstage, in a June 2026 interview.

It is rare for a Korean AI company to explain its revenue model this directly. In most interviews, AI executives enumerate technical superiority, the scale of partnerships, and the breadth of use cases. Asked—in that setting—how the company makes money, and answering "the product of tokens produced and tokens consumed," there is a useful lens embedded in that answer for anyone trying to understand how AI businesses actually work.

How Tokens Become Revenue

Upstage released its own language model, Solar, in 2023 and climbed toward the top of the Hugging Face Open LLM Leaderboard, after which it secured API supply agreements with Korean conglomerates including Samsung Electronics and SK. The company serves enterprise customers in healthcare, finance, and manufacturing, and has recently been linked to potential acquisitions of domestic platform companies.

The tokenomics structure is straightforward. Language models process text in units called tokens. In English, one word is roughly 1.3 tokens; Korean, with its particle-heavy grammar, typically requires more tokens to express the same meaning. An AI company's revenue is simply token price multiplied by volume consumed. The more users you have, using the product more frequently and for longer sessions, the higher the revenue.

The cost structure runs in the opposite direction. Training a language model costs anywhere from several billion to tens of billions of Korean won—but it is a one-time expenditure. After that, the core ongoing cost is inference: the GPU cycles consumed every time a user sends a query. As the user base grows, per-inference cost falls, and once consumption crosses a threshold, the initial fixed costs are more than recovered. In this structure, what every AI company wants is clear: more people, using it longer.

Platform acquisitions follow the same logic. No matter how capable a model is, it generates no revenue without users consuming tokens. A platform with existing traffic and an established user base is a way to capture token-consumption touchpoints at scale. It is not a strategy for building technical capability—it is a strategy for securing distribution.

Network Effects at AI Speed—and What They Hide

The classic network-effect cycle that business school textbooks describe plays out at a distinctly faster pace in AI. More users generate more training data; more data improves the model; a better model attracts more users. Because the company that completes this cycle first tends to lock in market leadership, AI companies accept short-term losses in a race to build user bases. Upstage's rumored platform acquisition reads, through this lens, less as a move toward technical sophistication and more as an attempt to secure the entry point for the cycle.

Before accepting this frame uncritically, though, it is worth examining the other side. Tokenomics is a supplier's language. "Increasing consumption" as a corporate goal translates, from the user's perspective, to "getting people to use more." The way OpenAI embeds GPT into an ever-broader range of everyday tasks, or the way Google integrates Gemini into search and Workspace tools, echoes how social media platforms were designed to maximize time-on-site. Whether consuming more tokens actually creates more value for users—or simply locks them into habit and convenience—is a question suppliers are poorly positioned to answer. Some corners of the Korean startup ecosystem have taken to saying: "The true unit price of AI services is the user's attention."

There are also more structural concerns. Tokenomics only works at scale. OpenAI exceeded $4 billion in annual revenue in 2024; Anthropic is targeting $3 billion for 2025. Without a consumption base at that level, competing on token pricing is treacherous—the economics do not hold. A framework can be theoretically sound while not applying equally to all participants. Whether Korean AI companies with proprietary models can survive long-term in a global consumption race remains an open question.

The Buyer's Math

What does this structure mean, practically, for solo operators and small organizations buying AI tools? There are a few things worth actually verifying.

Knowing exactly how your current AI tools charge you is the starting point. Many services look like flat monthly subscriptions but have token-based pricing running underneath; many that appear unlimited have daily caps or context-length limits built in. Claude Pro, ChatGPT Plus, and Gemini Advanced all sit at roughly $20 per month on the surface, but they differ meaningfully in how much complexity and volume they can handle. If your work is short and repetitive, you will get better unit economics from tools optimized for that. If you are regularly processing long documents or running complex reasoning tasks, your effective cost will differ. Which model you use for which task is itself a set of choices that plays out within that monthly subscription.

If you are working with AI APIs directly, the numbers get more precise. Building automation pipelines, customer service bots, or document-processing workflows means model selection directly determines operating costs. From Anthropic's lineup: Claude Haiku runs about $0.80 per million input tokens; Claude Sonnet is around $3; Claude Opus is considerably more expensive. Matching model tier to task complexity can shift costs by tens of percent for the same workload. And more expensive does not always mean better results—simple classification or templated summarization tends to favor lower-cost models, while complex judgment calls or creative generation tends to benefit from higher-end ones.

This lens also applies when deciding which AI company to trust as a long-term partner. A company whose profitability depends on growing token consumption will design its service differently from one whose contracts are tied to per-user outcomes. The former has every incentive to nudge you toward using more; the latter has reason to optimize for better results with less usage. That difference does not show up in product UI or marketing copy. It shows up when you examine the pricing structure and contract terms closely—that is where the design intent becomes visible.

Tokenomics was a concept AI companies coined to describe their own revenue model. Flip it around, and you get a clear picture of the economic relationship you enter every time you choose an AI tool. Two buyers paying the same monthly subscription—one who understands the structure, one who does not—will end up in meaningfully different places, both in terms of cost efficiency and actual results.