In June of this year, the CEO of the hiring platform Mercor disclosed that his company's AI token costs had, for the first time, exceeded its total payroll. Back in April, Nvidia executives said their internal math showed AI costs surpassing salary expenses. What sounded like isolated remarks got external backing in June, when Ramp — a company that aggregates corporate card spending data — published its AI Index: the top 1% of U.S. companies most aggressively adopting AI now spend more than $7,500 per employee, per month (roughly ₩10 million) on AI-related costs. The question this piece asks is what those numbers offer, as a yardstick for decision-making, to solo founders and hands-on managers in Korea.
The Two Different Worlds Behind $11.38 and $7,500
The Ramp AI Index reports figures at three tiers: $7,500 per employee per month for the top 1% of companies, $611 for the top 10%, and a median of $11.38 across the board.
To make that $11.38 median concrete: a single ChatGPT Plus subscription runs $20 a month, so the median company is spending about half of that. This bracket looks like occasional use of free plans, or maybe one Copilot license bundled into a work email account. These companies are using AI — but it isn't wired deeply into how the work gets done.
The $7,500 tier contains a different kind of spending altogether. At this level, you find automation pipelines built on direct API calls, multimodal processing, and operational systems that compare multiple models to optimize for cost and performance. Ramp notes that these companies don't depend on a single service; they mix open-source models with commercial APIs to manage costs. Under the same label of "AI-adopting company," wildly different realities coexist.
The speed matters, too. AI spending among the top tier is growing 14.1% every month. Compounded, that puts per-employee spending at roughly $36,000 a year from now. The average U.S. software engineer currently earns about $16,000 a month. Today's $7,500 is less than half an engineer's salary — but Ramp attaches a pointed qualifier to that fact: "yet." The message is to watch the trajectory, not the current number.
Bigger Spending Doesn't Mean Proportional Value
Treating these figures as benchmarks calls for caution. Criticism that the link between AI spending and actual results remains murky is coming from several directions.
In a 2025 McKinsey survey, more than half of companies that had adopted AI tools said they couldn't quantitatively measure any productivity gains. Gartner projects that roughly 30% of enterprise AI projects will be scaled back or shut down without delivering the expected results. That spending is rising, and that the spending produces real work output, are still two separate stories.
There's also the fact that subscription counts don't equal usage. It's hard to deny that the signal of "we're using AI" creates a favorable impression on its own — internally with employees, externally with investors and customers. That's why some organizations subscribe to more tools than they actually need, or treat adoption itself as the goal. When spending functions as a signal rather than a strategy, chasing the number ends in wasted money.
And the more the top 1%'s spending figure gets repeated in the press, the greater the risk that organizations rush into adoption without sufficient scrutiny. A big number is its own form of pressure.
When AI Costs Land on the Same Line as Payroll
Even so, there's a distinct reason this data matters to solo operators and small teams in Korea: AI spending has begun migrating from the "software subscription" category to the "labor replacement" category.
Until now, AI tools were budgeted as add-ons that made existing work a little easier — a SaaS fee of a few dollars a month, kept on a separate line from payroll. But as API-based automation spreads, the character of the expense changes. When recurring research, draft writing, data cleanup, and customer responses run on AI, that cost isn't a software cost. It's the cost of replacing a specific role's labor.
Long-range analyses tracking how the job landscape will shift toward 2030 see human–AI collaboration moving past simple tool use into a stage that reshapes role structures themselves. The core question of organizational design becomes not which tasks AI handles, but which judgments and responsibilities humans retain. Ramp's numbers show that this shift has already started registering as a line item.
I'd argue this calculation cuts even sharper in the Korean market. Few people have actually compared the cost of hiring experienced staff, freelance outsourcing rates, or the internal payroll consumed by repetitive work against today's cost of running AI. $7,500 is still a lot of money for a small Korean business — but the cost of starting at $11 and experimenting with which tasks can be automated, and how fast, is far lower than that.
Two things are worth checking. First: among the tasks you currently outsource repeatedly or burn internal staff hours on, are there any an AI workflow could handle — and if so, what fraction of the outsourcing rate would it cost? Put a number on it. Without that calculation, judging AI spending as expensive or cheap isn't even possible. Second: as your AI spending grows, are you tracking whether output quality or turnaround speed actually rises in proportion? Growing your subscription count and growing your results are different problems.
There's no reason to chase the $7,500 figure, and no reason to run from it. What the number says is that certain companies have started treating AI not as a software line item but as part of their workforce structure. Whether that judgment is right is still being proven — but that the judgment is already showing up as real cost is visible in the data. The people who'll use that number well are the ones who first decide where, and how, that judgment applies inside their own organization.
