The AI Bubble Won't Burst—The 'True Winners' and the Next Main Battlefield Seen in 2025
Will the AI bubble burst? Regarding this question, the on-the-ground sentiment in 2025 has moved in a somewhat unexpected direction. Y Combinator (YC) insiders speak of the realization that 'the AI economy has stabilized, and the division of roles between the model layer, application layer, and infrastructure layer has become clear.' In other words, it is not merely a frenzy (a bubble), but a perspective that investment and competition are preparing for the 'next phase of adoption.'
1. The AI economy moves from 'unstable' to 'stable'—Why the tectonic shifts of 2025 have ceased
From 2023 to 2024, if you waited a few months, the 'major announcements would overturn all premises.' However, in 2025, while models are improving incrementally, we are no longer in a phase where world-shaking updates are released one after another. As a result, it is said that a 'relative playbook' for how to build AI-native companies has emerged.
The side effects of this are also important. Previously, there was an atmosphere where 'if you wait, new ideas will fall into your lap,' but now 'finding ideas has returned to normal difficulty.' This is a sign of maturity; excessive dream-like stories have decreased, and one could say the real competition as a business has begun.
2. 'Model hegemony' is not fixed—A 'changing of the guard' from OpenAI to Anthropic at YC
Symbolic of this is the change in 'preferred LLMs' in the latest YC batch. While OpenAI was overwhelming until now, there is talk that Anthropic has surpassed them in the most recent batch. Moreover, this is not a short-term fluctuation, but a 'guard change' that has progressed over the last 3 to 6 months.
The background cited for this is the rise of so-called 'vibe coding' and coding agents. The nuance of the sentiment is this: 'Models that are strong in the coding domain end up becoming the founder's default.' Even if the product is not for coding purposes, a 'seepage effect' occurs where the model the founder is accustomed to using daily is more likely to be chosen.
3. The essence of Gemini's rise is not just 'reasoning'—Connection to search and real-world data
Gemini is also growing. There was mention that while it was in the single digits last year, it has risen to the 20% range recently. What is important here is that performance comparisons are not determined solely by 'pure intelligence.'
In one statement, the reason for using Gemini as a 'replacement for search' is attributed to trust in its grounding and real-time capabilities using Google's index. By contrast, people choose the tool that is 'more likely to be accurate even if a bit slower' over one that is 'fast but sometimes misses.' This is a sign that consumer-facing AI is moving from the 'convenience' phase to the 'trust' phase.
4. 'Switching models' is becoming normal—Orchestration and in-house eval become a moat
What is even more interesting is the observation that companies at the Series B level are abandoning 'loyalty to a specific model.' Founders are building an orchestration layer that abstracts models, moving toward operations where they 'swap in the optimal model for each task.'
The key here is their own proprietary evaluation (eval). In sectors like regulated industries, datasets and evaluation criteria become unique to the company, and the path to victory shifts from 'which model to use' to 'how to measure and how to switch.' In other words, as model commoditization progresses, the application layer can become more advantageous.
5. The 'answer' to the AI bubble theory—Some will collapse, but the next YouTube will benefit
Whether it is a bubble or not changes meaning depending on your position. While it may be a life-or-death issue for those holding GPUs and infrastructure (e.g., suppliers like NVIDIA), it can actually be a tailwind for student entrepreneurs and the application layer. The statement is explained using the metaphor of the dot-com bubble. It was precisely because of the excessive investment (dark fiber) in the 90s that bandwidth became cheap and abundant, allowing new services like YouTube to be established.
If the same thing happens with AI, the accumulation of excessive computing resources will become the foundation for creating the 'next giant app.' In conclusion, the AI bubble theory is an 'infrastructure investor's question,' and for those building apps, it is organized as a 'timeline of opportunities.'
6. Summary: The focus of 2026 is not 'who is smarter' but 'who is better at switching'
The core of 2025 was that the AI economy has 'stabilized.' The power map of models will shift, but the way to win has become visible. That is why what will matter next is not flashy hegemony battles, but
reliable grounding
optimization through in-house eval
Design based on the premise of switching (orchestration)
—these are the three points.
The truth about the AI bubble is not whether it will burst, but rather 'who will seize the "next standard" created by the surplus.'
