Recently, a long post went viral on X (Twitter). Although it was a personal claim, the tech industry took notice to the extent that various global IT media outlets covered it. The content is as follows.

Have Corporate AI Costs Really Increased?

There is a good reason why this article spread so quickly among people in the IT industry. Not only did it bring up real-life stories from prominent companies like Microsoft, Uber, and Nvidia, but it also clearly addressed the concerns of those who have recently introduced AI into the workplace. The anxiety that "using AI to work better actually costs more money" is a reality already felt by many. Why does your wallet get lighter the more you use it? It is precisely because of the billing method of AI services. People approach AI assuming it is a fixed subscription model where you pay a fixed amount per person, just like Netflix or office software. This is because paid services we use in our daily lives, such as 'ChatGPT Plus' or 'Claude Pro,' are actually flat-rate plans that allow unlimited use for a fixed monthly fee. However, the world of 'enterprise APIs,' which companies use when integrating AI into their systems or deploying large-scale staff for full-scale utilization, is a completely different story. From that point on, AI transforms from a monthly gym subscription into a taxi service where the meter goes up every time you ride. Technically, this is called 'token-based billing.' The unit of data and text exchanged with the AI ​​is called a token, and costs are charged in real-time based on the amount of these tokens consumed. Data charges accumulate steadily, like a meter, every time a line of code is reviewed or a bug is found. This is where a paradox arises. The better the tool, and the more enthusiastically employees use it, the steeper the costs rise. "Wouldn't this cost issue resolve itself if the unit price of AI tokens drops over time?" The answer is no. The actual market trend is moving in the exact opposite direction. The unit price of the tokens themselves, which serve as the raw material for AI, has been steadily declining since its launch. As recently as March 2023, when 'GPT-4' was first released, the price per million tokens reached $30 (approximately 45,000 won) based on input. However, the latest models with the same performance level are currently priced at only $2.50 (approximately 3,750 won). In just three years, the cost of using AI has plummeted to one-twelfth of its previous level. As costs dropped, companies increased their spending. With lower prices, companies sought to use AI for more tasks and much more frequently. No matter how much the unit price drops by half, if you use it in large quantities out of excitement, the total cost is bound to eventually rise. The emergence of 'AI agents' that make decisions and act on their own has added fuel to the fire. 'AI agents' that perform hundreds of tasks with a single command. Unlike existing AI that simply answers user questions, an AI agent refers to an autonomous AI system that makes decisions on its own and completes goals through complex steps. General AI usage is simple. If you input "Refine this sentence," the process ends with a single text exchange, and the cost is minimal. On the other hand, AI agents operate differently. When given the instruction "Fix the payment error in our service," the agent performs the entire process—reading the entire code file, analyzing logs, fixing the code, and running tests—all on its own. It repeats this process infinitely until the problem is resolved. This process results in tremendous token consumption. This is because, every time the agent decides on its next action, it must re-feed in the entire history of all work performed so far and the source code to learn. According to enterprise deployment data released by Anthropic, the average cost for an engineer using Claude Code is $13 per day (approx. 19,500 KRW) and $150 to $250 per month (approx. 225,000 to 375,000 KRW) based on active usage days. This is the average. As a result of one developer actively utilizing the agent and using Claude Code every day for eight months, the tokens consumed reached up to 10 billion. Converted into API prices, this amounts to a scale exceeding $15,000 (approximately 22.5 million KRW). Furthermore, companies have failed to properly grasp this cost explosion and have encouraged it by claiming that "using a lot of AI is innovation and capability." A prime example is Meta. Meta operated an internal dashboard called "Claudeonomics" that ranked its more than 85,000 employees in real-time based on who used AI the most. AI usage itself began to be managed as an indicator to evaluate employee performance. The result was obvious. As companies encouraged usage and fostered competition, employees spent tokens more aggressively, which resulted in unbearable bills for the companies. Only 3 out of 10 companies utilized AI properly in proportion to the cost. Statistics showing "the gap in return on investment costs for AI is widening" ©mavvrik, Lindsey Tishgart Are these uncontrollably escalating costs truly worth the price? While the numbers on the monthly bills are clear, whether that money actually returned an equivalent amount of profit to the company is a completely different matter. The report card for the 'ROI' of AI, hidden behind the flashy rhetoric of technological innovation, is far harsher than expected. This phenomenon is numerically proven by data released by major global research institutions for 2025 and 2026. An analysis of a report by the RAND Corporation revealed that 80.3% of all corporate AI projects failed to realize the business value originally planned. Even in the results of an analysis of over 300 real-world deployment cases by the Massachusetts Institute of Technology (MIT), only about 5% of AI pilot programs achieved meaningful results. According to a survey conducted by Gartner targeting 782 global IT infrastructure leaders, 28% of AI projects achieved their ROI targets, while 20% were discontinued. These project results are also being reflected in the evaluations of corporate executives. According to a survey by global consulting group McKinsey, 88% of companies adopted AI, but only 39% responded that it had a substantial impact on overall corporate revenue. In a survey conducted by the U.S. business magazine Fortune in February 2026 targeting 6,000 C-Level executives at global companies, 90% of respondents answered that they "have not found clear evidence that AI has increased productivity or changed employment structures within their companies over the past three years." Macroeconomic analysis also shows a consistent trend. Investment bank Goldman Sachs released an analysis stating that "currently, no statistically significant correlation has been identified between the overall economic adoption rate of AI and productivity." However, productivity improvements of approximately 30% have been confirmed in certain job categories, such as customer service and software development. Torsten Slok, Chief Economist at Apollo, analyzed the situation regarding the time lag between the scale of investment and its reflection in indicators, stating, "AI is everywhere now, but it is not in economic indicators." Changes are being observed not only in terms of cost but also in workforce management. According to a study published in 2026 by a UC Berkeley research team, 67% of employees reported an increase in working hours following the introduction of AI. This is due to structural factors where the total volume of assigned work increases alongside faster processing speeds. A 2026 study by the Harvard Business Review also found that 88% of employees who use AI intensively reported experiencing increased work fatigue. A phenomenon is emerging where the time saved through AI is being replaced by additional work. This result is also confirmed in the quality data of deliverables. According to a report by the development analytics platform GitClear that tracked 211 million lines of source code over five years, the rate of code re-editing increased following the introduction of AI. This is because the percentage of code that needed to be re-edited due to errors occurring within two weeks of writing rose from 5.5% in 2020 to 7.9% in 2024. During the same period, the amount of simply duplicated code blocks was also recorded to have increased eightfold. The direction pointed to by the numerous statistical indicators and figures so far is ultimately the same. Among the companies that have invested massive amounts of money in AI, only about one in three (approximately 30%) have reaped tangible benefits. The other two are still searching for answers while holding only expensive bills. You cannot stop driving a car just because gas is expensive. The statement that "AI costs more than a person's salary" is actually happening in a few teams that use AI extensively or in places that have brought in the latest AI systems entirely. This is evidenced by remarks made directly by an NVIDIA vice president and the bills from Uber developers. However, just because the cost of using AI exceeds human labor costs does not necessarily mean it is a failure. This is because if AI creates more value than the money spent, it is ultimately profitable. In the past, when automobiles first appeared, the cost of fuel and maintenance was much higher than the cost of raising horses that pulled carriages. Nevertheless, the reason people eventually chose automobiles was that they could transport goods much faster, further, and carry more cargo than carriages. Ultimately, the current debate is closer to transitional growing pains caused by the excessive speed of change as the entire industry undergoes a transformation. The real reason for the current explosion in AI costs is not a flaw in the technology itself. Companies still do not fully understand how to use AI efficiently and without waste within a pricing structure that rises in real-time, much like a meter. What companies need now is not a simplistic choice of whether or not to use AI. It is to learn how to properly 'drive'—meticulously controlling wasted costs while determining which tasks within the company require the deployment of AI to ensure it delivers substantial value for money. Ultimately, the major concern of the current AI industry—"We are spending a fortune, but how can we break even and run the business properly?"—has surfaced through this comparative debate between AI and labor costs.