In early 2023, almost every IR deck submitted to US venture capital firms contained one line, without exception: "We leverage GPT." At the time, that really was a signal. Few teams had access to the OpenAI API, and building AI features required a machine learning engineer. That alone was enough to set a company apart from its competitors.
How does that same sentence read in a pitch deck today?
AI inference costs have fallen sharply since 2022. Fine-tuning, agent orchestration, multimodal processing — all of these have become things a single developer can prototype in a matter of days, combining open-source tools with external APIs. The phrase "AI-powered" has stopped being a differentiator and become table stakes. Even so, many teams keep pouring their engineering effort into adding AI features, and the moment a competitor ships the same feature, they move on to building the next one. In this cycle, the first-mover advantage gets shorter with every round.
Why AI Features Don't Stay Differentiated
For an AI feature to create a competitive advantage, competitors have to be unable to copy it. But most AI features today are built on models or APIs released by a handful of companies — OpenAI, Anthropic, Google, Meta. Call the same API, and you get similar output. Building a proprietary model, by contrast, requires GPU infrastructure costing tens of billions of won (tens of millions of dollars) at a minimum just to get started.
The image-generation market illustrates this pattern well. In 2023, a wave of image-generation services launched one after another, built on Stable Diffusion and DALL-E. Features converged quickly, and the edge early movers retained came from brand recognition and users' workflow habits — not from the AI feature itself. Once the feature gap disappeared, price competition began.
The same pattern repeats in the Korean market. Legal document review, tax-filing assistance, customer service automation — in each of these areas, similar AI services have appeared just months apart. The underlying models are, for the most part, identical. Teams that got there first created some distance through UI polish or domain data, but that gap didn't hold for long.
Some teams do hold onto a performance gap through fine-tuning or proprietary data pipelines. But sustaining that gap requires a steady stream of new data, and generating data requires users. The reason users show up in the first place has to come from somewhere other than the AI feature. That's the basis for the argument that the sequence shouldn't be "build the AI feature, then get the advantage" — instead, you build the advantage first, and layer the AI feature on top of it.
When Counter-Positioning Actually Works
In business strategy, counter-positioning describes a move in which a new entrant stakes out a position that an incumbent finds hard to follow. The incumbent doesn't hold back because the technology is difficult — it holds back because following would cannibalize its own existing revenue.
Notion is the example that comes up most often. Microsoft is technically capable of building what Notion does, and it actually shipped Loop. But pivoting fully to Notion's approach would directly hit subscription revenue from Word, Excel, and Teams. Microsoft's hesitation to push Loop hard at the center of the Office ecosystem comes down to that revenue collision.
This dynamic sharpens in the AI era. There's a difference between areas where an incumbent can simply bolt AI onto its existing product to strengthen it, and areas where adding AI would erase the very reason that product exists. When a new entrant occupies the latter position, the incumbent is reluctant to follow at all.
There's one question that tells you whether counter-positioning holds: if a competitor copies your approach, does it directly damage their existing revenue? If it does, counter-positioning may hold. If it doesn't, then no matter how sophisticated the AI feature is, any first-mover advantage is temporary.
Even when counter-positioning does hold, it isn't permanent. It collapses if the incumbent decides to restructure its own business, or if an outside competitor claims the same position first. But that structure takes time to collapse — longer than it takes for a single AI feature to get copied by a competitor.
When Network Economics Kick In
Network effects describe a product that gains value as its user base grows. For this to hold true in an AI product, a rising user count alone isn't enough. There has to be a loop actually running: usage generates data, that data improves the AI's performance, and the improved performance delivers a better experience to the next user.
Regard AI, a clinical decision-support service, comes close to this pattern. As more hospitals join, more clinical case data feeds the training pipeline; as that data grows, diagnostic-support accuracy improves; and as accuracy rises, more hospitals adopt it. Even if a new competitor uses the same underlying model, the clinical patterns Regard has accumulated over years are hard to replicate quickly. It's a case where the more data piles up, the wider the gap with new entrants grows.
But AI products where this kind of data network effect actually works are rare. Many teams collect user data without any loop connecting it back to model improvement. If customer inquiry logs pile up without ever improving response quality, or usage-pattern data accumulates without ever feeding back into better recommendation logic, the data stacks up but no network effect ever materializes.
Whether you have 100 users or 100,000, if the 100th user gets the same value the first one did, that's not a network effect — it's just a service that scaled. There's one question that tells you whether a network effect is real: when the user base grows tenfold, does the value existing users get also go up?
What to Map Out Before You Add the Next AI Feature
Counter-positioning and network economics aren't new concepts. They're frameworks that competitive-strategy research has worked with for a long time — the kind of material you'd find in a mini-MBA strategy textbook. The reason both are getting renewed attention in the AI era is that AI features are leveling off so quickly that technology alone can no longer sustain an advantage.
For a solo founder or an early-stage startup, having both at once isn't easy. Counter-positioning requires an incumbent to exist in the first place, and network effects presuppose reaching an early critical mass of users. But if the alternative is just adding one AI feature after another, all that accumulates is development cost, while the first-mover window shrinks with every cycle.
A realistic starting point is to first map out what happens if the incumbent in your market follows you. If the incumbent has nothing to lose by shipping the same AI feature, then adding that feature is buying time, not building an advantage.
A feature is a piece of strategy, not the strategy itself. Teams that ask "What happens if a competitor copies this feature?" before they start building tend to last longer than teams that ask it afterward.
