ChatGPT's third anniversary: The battle of large models comes to a close, but where is the real moat?

OpenAI's ChatGPTAI has been around for three years. The real battlefield has now shifted to computing power, cost, and the ecosystem. The ultimate outcome may not lie in the model itself, but in who can turn AI into a sustainable operating system. (Background: ChatGPT will support direct payments via PayPal in 2026, the final piece of OpenAI's e-commerce empire) (Context: Check out the three main features of OpenAI's native browser “ChatGPT Atlas”; can AI agents shake Chrome's dominance?) From the launch of ChatGPT on November 30, 2022, until today, it marks exactly three years, and the pace of evolution in the tech industry has been astonishingly fast. At that time, everyone thought the AI competition would be a pure “model battle”: whoever has larger model parameters, thicker data, and more violent computing power would reign supreme in the market. However, looking back three years later, we realize that those early imaginings were overly simplistic. The real competition has never been about “who can create the largest model,” but rather “who can turn the model into a complete system.” A system that can be implemented, commercially viable, bear costs, sustain computing power consumption, and survive within enterprise workflows. Over the past three years, the technological curve of large models has visibly advanced, but the commercialization speed of AI companies has not kept pace and has even been slower than many expected. The reason is not a lack of technology but rather that the entire industry has shifted from comparing model sizes to the battlefield of “who can withstand the pressure before dawn arrives.” The four curves of model capability, computing power supply, inference cost, and user expectations are all rising madly, like a bow string pulled tight. And every AI company is on that string; whoever can hold on longer, more steadily, and bear the costs is the true winner. From parameter arms race to efficiency race In the first year of AI's emergence, everyone only saw parameters. The larger the model, the more advanced it was; the more expensive, the more high-end it was. The mainstream narrative at that time even regarded the amount of parameters as a kind of “dignity,” as if super-large models themselves could represent technological leadership. However, after 2024, the situation began to change subtly. Companies realized only upon actual deployment that whether the model was large or not had become unimportant; what mattered was whether the model could complete tasks “stably, cheaply, and quickly.” The increase in the intelligence of the model itself has ceased to show a linear explosion as it did in previous years; instead, progress has become more like a gradual fine-tuning. The larger the model, the more astonishing the inference cost, the higher the deployment threshold, and the less willing companies are to pay. Conversely, smaller models that are trained more efficiently, can be compressed, and run on ordinary GPUs have become the most popular AI products among enterprises in 2025. Many companies have even started to use open-source models internally to replace closed-source APIs, not because open-source is stronger, but because open-source has shattered all expectations in terms of “cost-effectiveness.” Lower computing power requirements, faster iteration speeds, and more flexible deployment methods have led many companies that originally relied on closed-source models to start questioning: “Do we really need to pay so much?” “Isn't 80% of the capability of open-source models, plus internal tuning, enough?” Model competition has now shifted from a “power competition” to an “efficiency competition.” It’s not about who is stronger, but about who can make it affordable for enterprises. GPUs are no longer hardware; they are strategic resources If we say that models have transformed from myths to commodities within three years, then GPUs have directly upgraded to “strategic materials” in this period. What AI companies fear most is no longer being behind in models but rather not having enough GPUs. As models grow larger, inference tasks increase, and user expectations rise, every AI company seems to be hanging on NVIDIA's supply chain. If there are not enough chips, new models cannot be trained; if there are not enough chips, inference speeds cannot be improved; if there are not enough chips, user bases cannot be expanded; if there are not enough chips, even funding becomes difficult to raise, because investors clearly know: without computing power, there is no future. This creates a strange state of competition in AI: technology is indeed advancing, but bottlenecks are in power, chips, and supply chains. The entire market seems to be simultaneously stepping on the gas and brakes, moving forward at a speed that leaves people breathless, but any chip shortage can cause companies to stall instantly. This is the most realistic and fundamental pain point of the AI industry: you are not competing with your opponents; you are competing with the supply chain. Therefore, inference costs have become a matter of life and death for enterprises. The stronger the model, the more expensive the inference, and the more users there are, the more losses are incurred. AI companies have become a counterintuitive business model: the more popular, the more they lose; the more users, the riskier it becomes. This is also why the true AI moat begins to become clear from this point. The real moat is not in the model Three years later, the market has finally reached a nearly brutal consensus: the capability of the model itself is no longer the most important moat. Because models can be copied, compressed, fine-tuned, and chased by the open-source community. The only two things that can distinguish winners from losers are: The first is “distribution” Enterprises with system-level entry points do not need the strongest models to dominate the market. Google uses its search engine and entire ecosystem to ensure stable traffic for Gemini, while Microsoft uses Windows and Office to make Copilot a natural entry point; Meta goes even further, directly integrating open-source models into Instagram, WhatsApp, and Facebook, dominating distribution. Distribution is the most traditional and realistic competitive edge in the tech world. If you have the entry, you have the voice, and this is why emerging brands like OpenAI, Perplexity, and Manus are facing increasing pressure. The second is “can AI really get things done” Chat capabilities are no longer the highlight, and multimodal capabilities are no longer rare. What truly matters is whether the model can appropriately call tools, write programs, analyze documents, connect APIs, decompose tasks, and become a real executor within the enterprise. When the model evolves into an “intelligent agent,” capable of completing processes, making decisions, and executing tasks on its own, it truly generates productivity. Companies that can build a complete toolchain will become the irreplaceable underlying infrastructure of the future, much like today's cloud platforms. Three years later, the market's moat has finally become clear: it’s not about whose model is the strongest, but who can turn AI into a well-functioning operational system. The future landscape of the AI market is gradually diverging into three ecosystems As the gap in model capabilities narrows, the pressure on computing power increases, and costs become central, AI companies have quietly divided into three camps that will all exist in the future but have completely different fates. The first type is platform-level giants These companies may not have the strongest models initially, but they possess overwhelming ecosystem advantages and silver bullets that allow them to catch up later. Companies like Microsoft, Google, and Meta have global distribution entry points, their own clouds, GPU reserves, data pipelines, and integrated products. For them, models are not products but tools “attached to the ecosystem.” The second type is pure model companies Companies like OpenAI, Anthropic, and Mistral are purely technical players with leading model capabilities, but they lack operating systems, smartphones, search engines, social platforms, and most importantly, “distribution.” No matter how strong their models are, they need to rely on others' ecosystems to reach users widely. The next three years…