The day Knowledge Research Group (KRG) released its report tallying 2023–2025 results for 80 AI-related companies listed on Korea's KOSPI and KOSDAQ exchanges, one figure stood out above all the rest. It wasn't the headline news that combined net profit at these AI-related public companies had more than tripled in just two years, from 25 trillion won to 81 trillion won (roughly $18 billion to $59 billion). It was the fact that 76 trillion won of that 81 trillion came from a single sector: semiconductors and the materials, components, and equipment — known in Korea as so-bu-jang — that go into them.

Add up the net profit every other sector generated over those two years and the total still falls short of 5 trillion won. Medical AI posted a 62.6 billion won loss in 2025 even as its revenue grew, while physical AI and robotics watched 2024 net profit fall nearly 90% from 2023 before clawing partway back. "Where the money is in the AI industry" and "where the money is in building AI services" are not, at this moment, the same thing.

Where the AI Boom's Profits Are Actually Made

What the KRG analysis shows most sharply is the concentration of profit. Net profit in semiconductors and so-bu-jang rose from 21.9 trillion won in 2023 to 76.1 trillion won in 2025. Over the same stretch, the sector's net margin climbed from 10.5% to 22.6%. That isn't simply a matter of higher revenue — it means the business shifted into one that wrings more profit out of the same sales.

Defense AI was the second-healthiest grower in the report. Revenue rose from 9 trillion won to 13.5 trillion won, and net profit from 880.8 billion won to 1.68 trillion won — an annual average growth rate of 38.4%. Three years ago, if you'd asked people to name the places generative AI would take hold, not many would have led with the military. The numbers now point in exactly that direction.

The LLM and agent segment grew its revenue and held net margins in the 15–20% range. But KRG sees the segment still anchored in public-sector projects, build-to-order contracts, and proof-of-concept (PoC) work. That's a different shape from the API-based, recurring-revenue model in the United States. A business that delivers a project and is then finished, versus one where subscription fees arrive every month, may show the same net margin — yet the two are valued in completely different ways.

The Two Sides of "Making Money From AI"

There are two ways to read this data.

One camp sees it as the natural shape of an early market. In the internet's early days, too, the telecom-infrastructure and hardware companies turned a profit first; software and platform profits followed years later. On this reading, semiconductors take the lion's share of today's profits because AI is still in its infrastructure-building phase. The losses medical AI and robotics are posting now read, in this context, as money spent ahead of the curve.

There's an opposing view. It holds that software-centric AI service companies have already, structurally, fallen behind in the global race. While OpenAI, Anthropic, and Google DeepMind lock in the world's developer ecosystem through their APIs, Korean LLM companies keeping the lights on with public-sector projects is, on this view, not a mere time lag but a divergence in path. The argument draws support from the data-center, cloud, and software segment, whose 2025 net profit actually slipped from its 2024 peak of 1.65 trillion won. GPU procurement and infrastructure spending keep expanding, yet profit shrank.

Which side is right can't be settled now. But going by the current numbers alone, the surest money in the AI boom is being made not by the companies building the software, but by the companies building the hardware that runs it.

What These Numbers Mean for Solo Operators and Solo PMs

Reading the report, I kept circling back to a question that financial thinking emphasizes: where is the money made, and where does it flow? This isn't only a matter of grand investment decisions. The same question works just as well for deciding where to spend your own time and skills.

First, if you're planning to build and sell an AI service yourself, it's worth taking another look at the revenue structure of Korea's LLM and agent segment. KRG's diagnosis — that it "hasn't reached a recurring-revenue model" — is more than a market assessment. Deliver an AI service as a project and you book one-time revenue, but that service won't generate revenue again next month. The thing to weigh is not the margin but the recurrence of the profit.

Next, when you spend on AI tools, building the habit of watching where that money flows sharpens your judgment. If a solo operator now spends 100,000 to 500,000 won a month on AI subscriptions, a sizable share of it passes through to the companies that own the GPUs. The figure that 94% of AI profit is concentrated in semiconductors and so-bu-jang explains where the subscription fee you pay gets converted into profit. Knowing that flow versus not knowing it makes a subtle difference in how you choose a tool and decide how much to spend.

Finally, the fact that medical AI and robotics are running losses is both a warning sign and an opening. There's room in fields that haven't been monetized yet. But entering one of them expecting quick profit is an entirely different choice from entering it without a capital plan that can absorb three to five years of losses.

The news that 80 companies more than tripled their net profit in two years is genuinely encouraging. But look into how that profit is distributed and you can see that there is more than one way to make money in the AI boom — and that the place where the surest profit appears is not the place where the software gets built. This may not be a signal that you should change course right now. Still, knowing which layer you're standing on as you make your next move is different from making it without knowing.