In May 2026, Korea's semiconductor exports hit an all-time monthly high of $37.16 billion. Yet the same country ranks second in the world in ChatGPT payments. Part of the money earned selling hardware gets wired to U.S. Big Tech every month under the name of subscription fees. Here are the numbers behind a phrase you hear more and more these days: 'digital rent.'
Two bills are landing on the Korean economy at the same time. One is money coming in. In May 2026, semiconductor exports reached $37.16 billion, up 169.4% from a year earlier and a new all-time monthly record. DRAM exports alone came to $18.6 billion in a single month (+369.8%). This is a supercycle built on HBM for AI servers and surging memory prices. The Korea Institute for Industrial Economics and Trade (KIET) projects that full-year chip exports will reach $350.1 billion, more than double last year's $173.3 billion.
The other bill is money going out. The digital services deficit captured in Bank of Korea balance-of-payments data — app subscriptions, cloud usage fees and the like — peaked at $3.4 billion in 2022, then shrank to $1.7 billion by 2024, only to start swelling again in 2025: $2.2 billion in the first nine months alone. What reversed the direction was, of all things, generative AI subscriptions. Over the same period, digital accounted for roughly 10% of the total services deficit ($22.7 billion).
This piece puts both bills on the same table. The conclusion, up front, is this. By amount alone, the digital deficit is worth no more than a few days of semiconductor exports in a boom month. The problem is not the size but the structure. Chips ride a cycle; subscriptions go out every month like rent, keep getting more expensive, and are entering the stage where canceling them means work grinds to a halt.
Digital rent is a metaphor, but the bills arrive through four very real channels.
The digital deficit in Bank of Korea statistics captures only part of this. Because generative AI subscriptions are not broken out as a separate line item, experts believe the actual outflow is larger than the statistics show. As a Korea Economic Daily report put it, "include generative AI and the digital deficit snowballs."
Chips are cyclical windfalls; subscriptions are rent that ignores the cycle. The boom may fade, but the billing date still comes. — This article in one line
Start with the trajectory of the digital deficit. The key point: a deficit that had been shrinking changed direction the moment generative AI went mainstream.
| Year | Digital Services Deficit | Trend |
|---|---|---|
| 2022 | $3.4B | all-time high (peak) |
| 2023 | $2.0B | improving |
| 2024 | $1.7B | still improving |
| 2025 (Jan–Sep) | $2.2B | worsening again — past 2024's full year in nine months |
Based on Bank of Korea balance-of-payments data (app subscriptions, cloud usage fees, etc.), as reported by the Korea Economic Daily on 2025.12.07. The deficit had been shrinking since ChatGPT's arrival in late 2022, then turned back up in 2025 as paid generative AI took hold.
Next, the numbers behind that reversal: the generative AI side.
| Metric | Figure | Context |
|---|---|---|
| ChatGPT spend, Korea (Jan–Oct 2025) | $200M | 5.4% of global total · No. 2 after the U.S. (35.4%) |
| Korea's app-download rank | 21st | but 2nd in spend — "downloads little, pays a lot" |
| Revenue per download (RPD) | $8.7 | effectively tied with the U.S. at $8.8 — world's top tier |
| ChatGPT users in Korea (2025.6) | 18.44M | dominant No. 1 AI chatbot in Korea |
| Paid gen-AI subscription rate, Korea (2025.7) | 14.3% | one in seven already pays |
| ChatGPT Plus monthly fee | about ₩29,000 | heavier with the won in the mid-₩1,550s |
Sources: Sensor Tower (published 2025.11), WiseApp and Opensurvey-affiliated studies, as relayed by Korean media. The gap between 21st in downloads vs. 2nd in spend is the core dataset of this piece. A country of 52 million pays the second-most, behind only a country of 340 million. Per capita, Korea is effectively paying the world's most expensive AI rent.
Why Korea, of all places? The data points to three answers.
First, those who use it, use it deeply. Ranking 21st in downloads while ranking 2nd in payments means that, relative to casual installers, the share of people who pay and embed the tool into their work and study is overwhelmingly high. Revenue per download of $8.7 is effectively the same as world-leading America ($8.8). On willingness to pay alone, Korea is already the United States.
Second, the substitutes are weak. Naver held the home turf in search and Kakao held it in messaging, but in frontier-grade generative AI the gap between domestic models and the global top tier is still palpable. And as the English-centric ecosystem's models have grown fluent in Korean, "I use it because it's homegrown" has stopped being a workable choice.
Third, Korea adopts fast. In a global survey, Korea's growth in AI usage from the first half of 2025 to Q1 2026 was +43.2%, the largest jump in the world. As with the internet and the smartphone before it, Korea is an early-adopter nation. The difference from those eras: this time the platforms all sit outside the country.
① Unit prices are climbing. The price tags on frontier models only move in one direction. Claude Fable 5, Anthropic's latest flagship unveiled this month, costs $10 per million input tokens and $50 per million output tokens — exactly double its predecessor. Smarter models launch at higher prices, and more and more of the work becomes possible only on those models. And as agents spend more hours working in place of people, token consumption itself explodes.
② Personal rent is spreading into corporate fixed costs. If the deficit so far has been mostly personal subscriptions, the next stage belongs to the enterprise. Once in-house AI adoption becomes standard, headcount times seat price and call volume times API rates flow straight into recurring expenses. Even when Korean companies build their own models, training and inference run on overseas GPU clouds. While the government and industry buy up 260,000 NVIDIA GPUs, the bills for the cloud that runs those GPUs, and for the software on top, grow right alongside.
③ The exchange rate inflates the bill. Subscription fees priced in dollars get heavier every time they convert to won in the mid-₩1,550s. The same subscription that cost in the low ₩20,000s two years ago now runs close to ₩30,000. Even if the deficit stays flat in dollar terms, the won burden on households and companies has already grown.
④ The warning of structural dependence. The phrasing from consulting firm Luxent, quoted by the Korea Economic Daily, stings. If Korean companies keep supplying only chips and equipment in the AI value chain, they "risk being reduced to mere hardware suppliers — IT tenant farmers." Renting the land (computing) and the seed (models) from someone else, then dutifully handing over a share of the harvest: this is where the rent metaphor deepens into sharecropping.
The dollars earned selling memory pay the usage fees of the data centers that memory goes into. We build the hardware; software takes the margin. — The irony of the digital-rent structure
For balance, the opposing case deserves equal weight.
First, subscription fees may be producer goods, not consumption. A Bank of Korea issue note estimates that AI adoption could lift Korea's productivity by 1.1–3.2% and GDP by 4.2–12.6%. If a ₩30,000-a-month subscription buys back even a few hours of one employee's productivity, that money is not rent leaking away but a lease with a high rate of return. Being the country that adopts AI fastest and deepest also means, flipped around, world-class competitiveness in actually using AI.
Second, the hardware boom is itself bargaining power. The bottleneck of the AI era is memory. No frontier model runs without HBM. The billions of dollars flowing out in monthly subscriptions and the tens of billions flowing in for memory are two ends of the same value chain. Korea is a tenant that also happens to be the sole supplier of the building's critical materials. That is why the full tenant-farmer metaphor is only half right.
Third, the window for homegrown alternatives has not closed. The government's AI policy push, domestic companies' efforts on their own models and Korean AI chips, and domestic-cloud requirements in the public and financial sectors all continue. For now, though, the reality is that usage is growing faster than the gap is closing.
Korea in 2026 stands in a strange place. At the very moment it is earning record dollars from semiconductors, it has become, per capita, the country paying the world's most expensive AI rent. These are not separate news stories; they are two faces of the same structure. The AI boom created the memory supercycle, and the same AI boom keeps inflating the subscription bill.
If this were a question of amounts, there would be nothing to worry about. It is a question of structure, which is why the feature stories keep coming. The boom will end; the rent will not. And before the rent deepens into sharecropping dues, part of the money earned has to be converted into a home of one's own: homegrown models, homegrown infrastructure, talent. Where the supercycle's dollars flow will decide whether, in the next cycle, Korea remains a tenant or becomes the landlord.