The issue reached Japan too One employee, JPY 10 million in monthly AI usage
The US cases were not a distant problem. Lessons from a domestic case reported by Nikkei
Key takeaway
On June 9, 2026, Nikkei reported a case in which one employee at a major Japanese company incurred JPY 10 million in AI usage in May alone. The CIO reportedly reacted with disbelief at an annualized pace above JPY 100 million. The US cases in articles 12–15 were not a distant concern: unexpected metered AI bills had also been reported in Japan.
Unexpected AI bills have reached Japan
JPY 10 million/monthReported usage for one employee in one month (Nikkei, June 9, 2026)
According to the report, repeated long-running activity included having AI agents converse throughout the night. At the same monthly pace, annual usage would exceed JPY 100 million.
What was reported: a figure that surprised the CIO
According to Nikkei's article by senior staff writer Toyonori Nakanishi, a CIO at a major Japanese company was surprised by one employee's AI costs: JPY 10 million in May alone, over JPY 100 million if simply annualized. The employee reportedly ran prolonged tasks, including overnight conversations between agents. The company was said to be considering a response because of the expense.
The article framed the issue as a pitfall of appearing to adopt AI while allowing usage itself to become the goal. This resembles the tokenmaxxing pattern in articles 12 and 13. Blaming the employee alone would miss the point: in a company encouraging AI adoption, enthusiastic users tend to consume more, a pattern seen across countries.
Why costs can be hard to notice in Japan
| Reason it goes unnoticed | What can happen |
|---|---|
| Dollar-denominated billing in arrears | A bill arrives the following month, potentially increased by exchange-rate movements. A weaker yen can make it higher than users expected. |
| Small-looking unit prices | A few tens of dollars per million tokens may look cheap, but an agent can consume millions or tens of millions in one job. |
| No individual usage breakdown | A shared departmental API key makes it difficult to determine who used the service and for what purpose. |
| Reluctance to dampen the adoption push | People may hesitate to raise costs while the company is promoting AI, delaying detection. |
What can happen
- Dollar-denominated billing in arrears
- A bill arrives the following month, potentially increased by exchange-rate movements. A weaker yen can make it higher than users expected.
- Small-looking unit prices
- A few tens of dollars per million tokens may look cheap, but an agent can consume millions or tens of millions in one job.
- No individual usage breakdown
- A shared departmental API key makes it difficult to determine who used the service and for what purpose.
- Reluctance to dampen the adoption push
- People may hesitate to raise costs while the company is promoting AI, delaying detection.
Another factor we see in Japan is the tension with annual budgeting practices. Many companies fix yearly budgets by cost category at the start of the fiscal year, with extra approval needed for increases. Metered AI spending can change severalfold from month to month, so initial estimates and actual spending may diverge structurally. Late discovery can leave an unwelcome choice between reallocating budgets and stopping usage. This is our interpretation of a risk associated with Japanese budgeting practices, rather than a fact established by the US example in article 15.
Reviewing the pattern: five cases, one cost structure
Across the cases from article 12 onward, a common tension emerges: organizations want people to use AI, while the pricing model charges for every use. Enthusiastic, productive use can grow the bill.
- 12
Reported case: Amazon: a consumption leaderboard was discontinued
- 13
Reported case: Meta: an unofficial consumption leaderboard drew intense interest and was closed
- 14
Reported case: Microsoft: internal use of a popular tool was reconsidered
- 15
Reported case: Uber: the annual budget was reportedly used in four months
- 16
Reported case: Japan: JPY 10 million in a month for one employee
The US and Japanese examples share the same underlying pricing dynamic, despite differences in scale and timing. Nikkei also pointed to rising costs as use shifted from chat interfaces to agents.
Changing the structure behind a surprise JPY 10 million token bill
Monitoring, limits, and usage breakdowns all help. But each is ultimately a way to have people watch an open-ended tap. If that monitoring weakens, the same pattern can recur.
An upfront-purchase on-premises system such as Sovereign GaiXer changes the tap itself. Overnight local agents do not create additional per-token charges, although electricity and applicable operating costs remain. Nighttime may also offer spare processing capacity. Useful overnight work can become better use of an existing asset, instead of a source of metered charges. The same activity can have very different cost consequences under a different pricing model; it still needs a useful purpose and appropriate controls.
Imagine a whole family using phones on metered international roaming. One email may cost little, but unrestricted video changes the scale, and the bill arrives later. Small unit prices, open-ended volume, and delayed discovery combine to create an unexpected bill.
Frequently asked questions
Is JPY 10 million per month an extreme case?
It was reported as an exceptional case. But article 15 also described a USD 1,200 two-hour session at Uber. Powerful models, agents, and continuous operation can make costs accumulate rapidly. Do not assume that an extreme example is impossible in your environment.
What kind of usage makes the cost grow that much?
The report cited overnight conversations between AI agents. An agent can make hundreds of model calls from one instruction and repeatedly include conversation history, causing token usage to grow rapidly.
Should we tell employees not to use AI?
A blanket prohibition can undermine a developing culture of useful adoption. Address the cost structure through limits, appropriate model selection, and, where suitable, fixed-cost local inference, while preserving productive use.
What is an appropriate AI cost per employee?
There is no universal figure. The reported coding examples in article 15 ranged from USD 150–250 monthly on average to USD 500–2,000 for heavy users. The more important question is whether outcomes justify the spending.
What is a simple way to detect an overrun sooner?
Arrange daily usage-cost reporting to the responsible team's email or chat where the provider supports it. Monthly invoices can reveal a problem weeks late; daily visibility can bring it to attention the next day.
Can an on-premises system really run all night?
Overnight local inference does not incur additional per-token charges, though electricity, cooling, hardware capacity, and operational controls still matter. Useful batch work during otherwise idle hours can be well suited to a fixed hardware investment.
Summary
- Nikkei reported on June 9, 2026 that one employee at a major Japanese company incurred JPY 10 million in AI usage in May.
- Long-running AI agents were reported as a factor; moving from chat to agents can sharply increase usage.
- The underlying dynamic matches the US examples in articles 12–15: unexpected metered AI bills are relevant to Japan too.
- Individual usage records, effective budget controls, and agent stopping conditions help, but require continued monitoring.
- Local upfront-purchase AI avoids surprise per-token bills and can put spare nighttime capacity to useful work, while operating costs remain.
Related articles
This article is based on public reporting, without independent interviews. The company and employee were not identified in the reporting, and this article does not speculate about their identity. Amounts reflect the reporting date. Please contact us with factual corrections. The media, researchers, and companies discussed do not endorse Sovereign GaiXer.
Company, product, and service names mentioned are trademarks or registered trademarks of their respective owners.