At a press interview, the CEO of an AI startup said something unexpected. "How many tokens our model produces — and how many of those tokens get consumed." In a room where most executives talk up technical capability and strategic partnerships, he led with a revenue formula. It was Upstage CEO Kim Sung-hoon, speaking at a June 2026 interview, who coined the term "AI tokenomics."

It's rare for a Korean AI company to explain its revenue model so plainly. In most interviews, AI executives list technical superiority, partnership scale, and breadth of use cases. So when someone answers the question "how do you make money?" with "the product of token production and token consumption," there's a way of seeing the AI business embedded in that answer worth unpacking.

How Tokens Become Revenue

Upstage launched its in-house language model Solar​ in 2023 and quickly climbed the Hugging Face Open LLM Leaderboard, then signed API supply agreements with Korean conglomerates including Samsung and SK. The company serves enterprise customers in healthcare, finance, and manufacturing, and has recently been linked to potential domestic platform acquisitions.

The structure of tokenomics is straightforward. Language models process text in units called tokens. In English, a single word runs about 1.3 tokens; Korean, with its grammatical suffixes parsed separately, requires more tokens to convey the same meaning. An AI company's revenue is simply token price multiplied by volume consumed. The more users there are, using the product more often and for longer, the more revenue grows.

The cost structure runs in the opposite direction. Training a language model costs anywhere from millions to tens of millions of dollars — but that's a one-time expenditure. The ongoing cost driver is inference: spinning up GPUs every time a user sends a query. As usage scales, per-inference cost falls, and once consumption crosses a threshold, the initial fixed investment is fully recouped. In this structure, what a company wants is clear: more people using the product for longer.

Platform acquisitions make economic sense through this same lens. However capable a model is, no usage means no revenue. A platform that already has traffic and an established user base is a shortcut to acquiring token-consumption endpoints at scale. It's not a strategy for deepening technical capability — it's a strategy for securing distribution.

The Network Effect on Fast-Forward — and Its Blind Spots

The network-effect dynamics familiar from business-school textbooks operate at a faster tempo in AI. More users generate more training data; better data improves the model; a better model attracts more users. The company that completes this cycle first claims market leadership — which is why AI firms absorb near-term losses to lock in user bases early. Upstage's reported platform acquisition moves read, through this lens, less as a push for technical depth and more as an attempt to secure the front door of the cycle.

But before accepting this framing wholesale, it's worth looking at the other side. Tokenomics is the supplier's language. A company goal of "growing token consumption" translates, from the user's perspective, into "getting you to use it more." The way OpenAI embeds GPT ever deeper into everyday workflows, or Google integrates Gemini into Search and Docs, mirrors how social platforms once optimized for time-on-site. Whether consuming more tokens creates genuine value for users — or simply binds them through habit and convenience — is a question suppliers find difficult to answer honestly. In parts of the Korean startup ecosystem, the shorthand is: "the real unit price of an AI service is the user's attention."

There's a more concrete problem, too. Tokenomics requires scale to function. OpenAI cleared $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 becomes structurally difficult. A framework can be valid and still not apply equally to every player within it. Whether Korean AI companies running proprietary models can survive long-term in a global consumption race remains an open question.

The Buyer's Calculation

What this structure means for freelancers and small organizations buying AI tools is worth working through concretely.

The starting point is knowing exactly how the AI tools you use actually charge you. Some services appear to offer flat monthly rates but run on token-based pricing internally; others look unlimited but have daily caps or context-length limits tucked into the fine print. Claude Pro, ChatGPT Plus, and Gemini Advanced all cost roughly $20 per month on the surface, but differ meaningfully in the complexity and volume of work they can handle. If your work skews toward short, repetitive tasks, per-unit cost efficiency is high; if you frequently process long documents or run complex reasoning chains, your real operating cost shifts. Which model you reach for within a subscription is itself a meaningful variable.

If you're working directly with AI APIs, the numbers get sharper. Building an automation pipeline, a customer-service bot, or a document-processing workflow means model selection directly determines operating costs. Anthropic's Claude Haiku runs about $0.80 per million input tokens; Claude Sonnet is around $3; Claude Opus is substantially more. Matching model tier to task complexity can cut costs by tens of percent on the same workload. And higher-capability models don't automatically produce better results. Simple classification or templated summarization suits a cheaper model fine; complex judgment calls or creative generation warrant the higher tier.

The same lens applies when choosing a long-term AI partner. A company whose profit grows with token consumption is designed differently from one that contracts on per-user outcome metrics. The former has structural incentives to encourage more usage; the latter has reasons to optimize for better results with fewer tokens. This difference doesn't show up in product UIs or marketing copy. It surfaces when you examine pricing structures and contract terms — that's where the design logic reveals itself.

Tokenomics was coined by AI companies to explain their own revenue models. Turn it around, and it illuminates the economic relationship a buyer enters every time they choose an AI tool. Two organizations paying the same monthly subscription can see meaningfully different costs and outcomes — separated entirely by whether they understood the structure they were buying into.