Token tiers: Samsung rations AI access as costs add up
As Samsung, SK and others now rush to adopt U.S. AI models, rising token costs are forcing strict limits on workplace use, sometimes blunting effectiveness.
A hand holds a spread of cards bearing the logos of major AI models — Claude, Gemini and ChatGPT — against a stack of coins, illustrating the growing cost of enterprise AI adoption.GETTY IMAGES PRO
[NEWS ANAYSIS]
Korean firms are grappling with the steep costs of AI token usage as they race to integrate foreign models into their operations — a move that, ironically, risks undermining the very efficiency gains they are chasing.
Nearly four years after ChatGPT kicked off the global AI boom, Korean companies have largely lagged in adoption, with many either moving slowly or banning its use over security concerns.
But recent moves show that hesitation is now giving way to action. Korean conglomerates including Samsung and SK are opening up access to U.S. models — Claude, Gemini and ChatGPT — through enterprise contracts.
Still, the transition is proving costly and still delicate for security-sensitive sectors like semiconductors. Companies including Samsung are now wrestling with how to ration token usage among employees, trying to strike a balance between AI-driven productivity and the expense of running these models at scale.
Mixed signals and Catch-22s
Enterprise use tends to be token-intensive by nature, as corporate workflows typically involve complex, high-volume tasks that consume more compute than casual consumers use.
Claude’s token usage have surfaced as a new cost factor domestically. A relative newcomer to the mainstream, Claude has quickly become the hottest AI commodity in Korea, particularly among developers and other heavy users who favor it for its coding and task execution capabilities.
To manage costs, Samsung has imposed strict limits on AI token usage as it opened up company access to the three U.S. models internally. Employees are divided into tiers, with the base level restricted to simple tasks. For an employee to move up, it requires demonstrating that AI is actively improving their work output.
“The level-up is not automatic,” said a Samsung employee who requested anonymity. “If you want your limit raised, you need to show results. You have to prove you’re putting it to good use.”
This has raised a small pool of complaints internally. Some employees are saying the token limits are too excessive to be able to put them to good use.
Korea ranks 14th out of 121 countries for Claude usage, according to the Anthropic Economic Index, with usage per person running more than 3.5 times higher than the global average.
The deepest penetration is in software development. Naver has rolled out Claude Code across its entire engineering organization, a deployment the company describes as one of the largest of its kind in Asia. At Nexon, thousands of developers are using Claude Code across the full game development cycle, from design and coding to review and deployment on live-service titles.
For large conglomerates, adoption is broadening into task automation and system integration. Samsung SDS has introduced Claude Cowork and Claude Code for Samsung Electronics employees, applying them to everyday workflows, agent-based task automation and software development. LG CNS is rolling out Claude to thousands of its employees in phases for software development and client technology solutions, with plans to expand deployment across the wider LG Group.
Many smaller companies, however, lack such a structured approach, relying instead on subtle signals to discourage excessive use.
“One day, the company shared a ranking of the 10 biggest AI token users. The No. 1 employee had spent around 6 million won [$4,100],” said an employee at a major biotech manufacturer. “Management wants us to use AI, but if you use too much, you start getting the feeling they're watching the costs.”
At some companies, entire teams share a single token pool, forcing dozens of employees to divide the allocation among themselves and limiting access to heavier models like Fable.
“GPT and Claude are the most popular options right now, but the usage caps are tight enough that most of us gravitate toward Codex, which is less token-heavy for coding tasks,” said another employee at a manufacturing company.
Chip sector still holding out… for now
AI adoption is far from uniform across industries. Semiconductors, in particular, have been especially cautious about deploying AI across their systems.
Even within Samsung Electronics, the pace varies by division. Some units use AI freely, but usage at the semiconductor side remains largely off limits, according to a source familiar with the matter.
“Employees at the chip unit are waiting for approval to come through, but there has been no sign of it anytime soon,” the source said.
The gap is even starker within SK Group. While some affiliates have already deployed all three U.S. models through enterprise contracts and are actively working to build habitual AI use among employees, SK hynix remains limited to in-house models.
Days after Samsung's announcement, SK hynix CEO Kwak Noh-jung told employees at a town hall in Icheon, Gyeonggi, that the company is “reviewing the adoption of Microsoft 365 and Copilot, and also exploring the possibility of using ChatGPT Enterprise.”
For now, however, that remains a possibility rather than a reality.
