Samsung Sets Token Quotas for Three Major Models

Overseas AI Models Enter the Enterprise—But Token Budgets Are Running Out

July 27 — As major South Korean companies expand generative AI from limited employee trials into formal business workflows, model usage costs are rapidly emerging as a new management challenge.

South Korean companies, including Samsung and SK Group, have recently begun providing employees with access to Anthropic Claude, Google Gemini, and OpenAI ChatGPT through enterprise contracts. Samsung has also introduced tiered token quotas: different employees receive different allowances, with the basic tier limited to relatively simple tasks. Employees who want larger quotas must do more than submit a request—they must demonstrate that AI has actually improved their work efficiency.

One Samsung employee said that quotas are not increased automatically. Users must present tangible results showing that AI has been used for work. The system has also triggered internal dissatisfaction: some employees believe the current limits are too restrictive, as the model may consume most of their quota merely by reading the project context, leaving too little to complete the full task.

This development deserves attention not because of how many fewer tokens Samsung is allocating, but because it shows that the central challenge of enterprise LLM adoption is changing. The question used to be, “Can we use it?” Now it is, “Who can use it, which model should they use, how much should it cost, and is the output worth it?”

Why Tokens Have Suddenly Become an Enterprise Burden

When individuals use chat products, costs are typically bundled into a monthly subscription fee. What users see are message limits, rolling time windows, or vaguely defined “fair use” restrictions. Once enterprises integrate models into software development, documentation, customer service, and data analysis workflows, however, the billing logic becomes entirely different.

A single enterprise task consumes more than just the sentence entered by the user. It usually also includes:

• System prompts, access-control rules, and enterprise security policies;
• Retrieved internal documents, code files, and conversation history;
• Responses, code, and structured data generated by the model;
• Multi-step agent planning, tool calls, and automatic retries after failures;
• Code repository indexing, long-context analysis, and information passed among multiple subtasks.

As a result, “Help me modify a function” and “Assess the risks of upgrading dependencies across the entire repository” may each appear as a single instruction in the employee interface, yet their actual token consumption may differ by several orders of magnitude.

This is particularly true of coding agents such as Claude Code. Rather than merely completing the next line of code, they read files, search for references, inspect tests, modify multiple modules, and continue iterating based on execution results. With every additional step, both the input context and output may continue to grow. The more capable a model becomes at handling complex tasks, the more likely an enterprise is to assign it larger work packages. Costs therefore do not remain at chatbot levels.

Three concepts that are often conflated must also be distinguished: the model’s context window, the usage limits imposed by the vendor on users, and the token budget allocated internally by the enterprise. A large context window means that a single request can theoretically accommodate more information; it does not mean that this information is processed for free. Purchasing an enterprise plan on a per-seat basis also does not mean that every API, agent, and intensive coding task can be used without limits. Samsung’s quota system is essentially an internal cost-governance policy and may not map one-to-one to the raw token counts on model vendors’ bills.

Claude Is Drawing Attention Because It Is Being Used in the Most Cost-Intensive Scenarios

The fact that South Korean companies are paying particular attention to Claude’s costs does not necessarily mean that Claude is more expensive than every competitor. A more accurate explanation is that it is used more extensively in demanding scenarios such as software development.

According to the Anthropic Economic Index cited in reports, South Korea ranks 14th in Claude usage among 121 countries and regions, with per-capita usage exceeding the global average by more than 3.5 times. Software development is the field in which Claude has achieved the highest penetration in South Korea—and it is also a scenario characterized by long contexts, numerous interaction rounds, and large outputs.

South Korean internet company Naver has deployed Claude Code at scale across its engineering R&D teams, describing it as one of the largest deployments of its kind in Asia. Gaming company Nexon also has thousands of developers using related AI coding capabilities. For these companies, AI is no longer a webpage that employees occasionally open. It is becoming embedded in code review, refactoring, testing, and troubleshooting workflows.

This produces a counterintuitive result: models that are genuinely useful may be more likely to create budget pressure.

If a model performs poorly, employees will abandon it after a few attempts, and the bill will remain low. If the model can actually complete work, however, usage frequency, context length, and the number of automated execution steps will all increase. The enterprise gains productivity—but it also acquires a usage-metered “digital machine” that thousands of employees can operate concurrently.

From this perspective, token quotas are not a sign that AI has failed. They are a byproduct of AI entering production environments.

Samsung’s Approach Is Reasonable, but Static Limits Can Easily Penalize Heavy Users

Samsung’s requirement that employees earn higher quotas by demonstrating results is not misguided. Enterprises cannot absorb unlimited model usage costs simply because employees are willing to use AI. In the semiconductor industry in particular, companies must consider not only costs but also whether source code, design documents, process parameters, and supply-chain data may be sent to external models.

The problem is that allocating fixed token quotas based solely on employee level is easy to administer but may not be economically efficient.

First, reasonable consumption varies enormously across roles. A legal professional summarizing a contract, a marketer generating a few lines of copy, and a developer asking an agent to read a large code repository should not be measured by the same standard. If the basic quota is sufficient only for simple questions and answers, the enterprise may appear to have costs under control while actually restricting the most valuable, in-depth tasks.

Second, token consumption does not directly indicate waste. A token-intensive code migration may save an engineer several days of work, while dozens of inexpensive copy-editing requests may produce no measurable benefit. Judging solely by usage volume is like assessing a project’s value based on how much electricity employees consumed: it is useful for identifying anomalies, but it should not determine budgets on its own.

More troublingly, strict limits may encourage “shadow AI.” When official quotas are insufficient, employees may turn to personal accounts, browser extensions, or unapproved third-party services. The company’s recorded costs decline, but the risks of data leakage and compliance violations are shifted out of sight. For a company such as Samsung, which holds extensive core technology assets, this is clearly not a worthwhile trade-off.

Our assessment, therefore, is that Samsung needs to manage token quotas, but if the system remains limited to “allocate quotas by job level and require a written request for overages,” it will quickly become a new approval bottleneck. What enterprises truly need is not a wall of uniform height, but an observable, routable, and accountable AI cost-management system.

Model Routing Matters More Than Usage Limits

The simultaneous introduction of Claude, Gemini, and ChatGPT into the enterprise does not mean employees should always be free to choose the most powerful—and often most expensive—model. A more mature approach is to route requests dynamically according to task value, complexity, and data classification.

For example:

1. Route simple tasks to lightweight models. Text classification, format conversion, short summaries, and basic Q&A do not need to invoke flagship models by default.
2. Upgrade long-context tasks as needed. Long-context capabilities and larger budgets should be enabled only when analyzing large documents or code repositories is genuinely necessary.
3. Allow higher consumption for high-value tasks. Core code reviews, complex incident diagnosis, and analysis of important contracts can receive dedicated budget pools instead of consuming employees’ routine allowances.
4. Use caching and retrieval for repeated content. Enterprise policies, shared code, and fixed system prompts should not be retransmitted indiscriminately with every request. Prompt caching, retrieval trimming, and context compression should be used to reduce costs.
5. Set execution boundaries for agents. Limit the maximum number of steps, retries, tool calls, and the budget per task to prevent automated loops from continuously burning money in the background.
6. Incorporate sensitivity into routing. Tasks involving core designs, customer privacy, or nonpublic financial data should be routed to controlled environments or prohibited from being sent to external services that do not comply with company policies.

Enterprises should also track both costs and outcomes rather than tokens alone. In software development, they can monitor pull-request acceptance rates, defect rates, code-review time, and task completion cycles. In customer service, they can track first-contact resolution, escalation-to-human rates, and customer satisfaction. For documentation tasks, they can measure the amount of manual revision required and delivery speed.

Only by evaluating these metrics alongside model consumption can enterprises answer the most important question: Are these tokens creating value, or merely generating more content that looks productive?

AI FinOps Will Become New Infrastructure

As cloud computing became widespread, enterprises gradually established FinOps practices: allocating cloud-resource costs among departments, setting budget alerts, identifying idle instances, and choosing different configurations based on business workloads. LLMs are following the same path, except that the units of measurement are shifting from CPU, memory, and storage to input and output tokens, cache hits, tool calls, and agent execution time.

Large enterprises are likely to establish governance structures resembling cloud cost centers:

• Allocate model costs by employee, team, project, and application;
• Set daily budgets, monthly budgets, and per-task limits for different models;
• Automatically switch to cheaper models when usage approaches a threshold;
• Detect abnormally long contexts, repeated requests, and endless retries;
• Reserve budgets for high-value projects instead of using a first-come, first-served model;
• Bring model quality, latency, cost, and data compliance together in a single decision dashboard.

This will also change the significance of multi-model strategies. Enterprises procure Claude, Gemini, and ChatGPT not only to avoid dependence on a single vendor, but also to control costs through competitive routing: assigning coding tasks to the best-performing model, prioritizing better-integrated services for tasks within office ecosystems, and directing low-risk batch processing to less expensive models.

In other words, integrating multiple models is only the first step. The enterprise’s ability to select models in real time based on the task will ultimately determine how much it pays.

As South Korean Companies Catch Up, They Are Encountering the Real Cost of Scaling

South Korean companies had previously been relatively cautious in promoting generative AI, partly because of slow implementation and partly because of concerns about data security. Now, as Samsung, SK, Naver, Nexon, and others accelerate the deployment of overseas models, they are effectively moving beyond small-scale demonstrations and confronting the practical challenges of large-scale use.

These challenges will not be unique to South Korea. Any company with thousands of knowledge workers will eventually encounter the same cost curve once it broadly enables coding agents, long-document analysis, and enterprise search. Falling model prices may not solve the problem, because lower prices often lead to more calls, longer contexts, and more complex automated workflows. This is similar to what happened in the cloud-computing era: unit computing costs kept falling, yet enterprises’ total cloud spending continued to rise.

Token quotas are therefore more than a cost-cutting measure. They signal that enterprise AI is entering its next stage. Model capabilities have largely crossed the threshold of usability; the next competitive challenge is determining who can bring cost, access control, security, and business value together within a single system.

Samsung’s current approach may seem somewhat rigid, but it raises a question that every enterprise must address: When AI truly becomes a productivity tool for every employee, how much should a company pay for a single answer, and how should it determine whether that answer was worth the cost?

The answer will not be a uniform token-limit table. It is more likely to involve granular routing, departmental cost accounting, task-level budgets, and continuous evaluation of final business outcomes.

References

• ITHome: South Korean Companies Face Cost Challenges After Adopting Overseas AI Models; Samsung and Other Giants Introduce Token Quotas—Describes how Samsung, SK, and other South Korean companies are deploying overseas large language models, as well as Samsung’s tiered token quota system and Claude usage in South Korea.