Tuesday, July 21, 2026

In today's technical news, there are actually two things worth talking about most: the server exploded 48 hours after Kimi K3 was released, and the Jacobi conjecture was falsified by AI. These two things may seem incompatible, but they tell the same story-when AI's capabilities exceed expectations, the entire ecosystem is caught off guard.

Let's talk about Kimi first. Dark Side of the Moon really played himself out this time. K3 's performance is approaching GPT-5.6, which is indeed a milestone, but the result is that the number of user requests has increased sharply, and the server has to suspend new user subscriptions. This exposes a cruel reality: the AI industry is not that the technology is weak, but that the infrastructure cannot keep up. No matter how awesome your model is, it will be an empty shell without computing power. Nvidia's share price plummeted accordingly. It is not unreasonable. The market realizes that computing power is not unlimited, and the bottleneck lies in cluster scheduling and network bandwidth, not simply insufficient cards. What does this mean for ordinary developers? Don't rush to use K3, think about whether your application can withstand API instability. In addition, Anthropic changed its subscription strategy overnight, which shows that everyone is scrambling for computing power, and this wave of price increases and restrictions is inevitable. Commercialization? K3 can make money in the short term, but in the long run, computing power costs will eat up all profits, unless the dark side of the moon can solve self-built clusters.

Besides, Jacobi's conjecture was falsified by AI. This matter is seriously underestimated by the technology community. Some people think this is the power of AI, but I think what is more worth discussing is: How was this counterexample discovered? It was not the AI itself that deduced it, but humans used AI to assist in searching and found a special case in the huge algebraic structure. This is actually a signal: AI's ability in mathematical reasoning is no longer just to "do problems", but to discover patterns that humans don't notice. But the risks are also obvious-if the counterexample found by AI cannot be understood by humans, then it is essentially a black box tool. For developers, this means one thing: in the future, when doing scientific computing, financial modeling, and encryption algorithms, you have to consider whether AI can take advantage of loopholes to find loopholes that you don't exist. The threshold for this industry will become higher and higher.

Connecting these two things shows a larger trend: AI is evolving from "being able to write code" to "being able to reason", but the price of reasoning is the exponential growth in computing power requirements. Lin Dahua of Shangtang said that multimodal is the next battlefield. I agree, but only if you can suppress calculation. Tianxian Intelligent's Tiangai 300 claims to be better than the Hopper solution, but don't forget that domestic GPUs still have a long way to go in terms of ecology and software stacks. Another signal is the article on Hacker News that "China's open source AI strategy is winning." This statement is a bit exaggerated, but it does reflect the reality: the open source model is close to closed source in terms of performance, but the cost is much lower. This is good for developers, but for startups, the barrier to entry is even higher-because open source models turn basic capabilities into free goods, and you can only differentiate them based on data, scenarios, and experiences. Those posts on V2EX that discuss stocks and fraud have nothing to do with technology, but they show that the developer community is also divided. Some people are still struggling at the bottom, while others are already thinking about how to monetize it.

A last heart-breaking word: The bubble of AI comics bursts as fast as the iteration of technology. It only took a year to go from hot to low tide. This profession has no feelings, only implementation. Whether the product you make can solve a real problem and whether it can run through even when computing power is unstable is the key. Don't be carried away by open source models, or by the numbers at the press conference. Let's first see how much computing power you can hold in your own hand.

Today's signals bring the two main lines of the AI industry onto the same canvas: breakthroughs at the model level are intertwined with bottlenecks at the computing power level. K3 caused a run on computing power within 48 hours with performance close to GPT‑5.6. The Dark Side of the Moon directly announced the suspension of new users. Shang Tang Lin Dahua positioned multimodal as the "battlefield after Coding." This shows that the speed of model innovation has far exceeded hardware supply, and computing power has become a scarce resource that determines the competitive landscape. The capital market's pursuit of computing power (such as Tianxian Zhixin Tiangai 300) is turning into an investment boom in the computing power supply chain. However, in the short term, the expansion of computing power still faces the dual pressures of high costs, uncertain supply and energy consumption supervision. If companies fail to deploy self-developed chips or rent reliable computing power pools in advance, they will be marginalized in the competition and even face the business risk of service interruption.

At the same time, the AI content ecosystem has been de-transformed: the bubble of AI comics burst within a year, and the industry has begun to warn of "commodity logic versus creative freedom." The threshold for content production has been lowered by the big model, but the commercialization path is unclear, resulting in the rapid withdrawal of capital. Combined with the reality of tight computing power, if content entrepreneurs cannot find a balance between computing power costs and creative value, it will be difficult for them to obtain sustained traffic and monetization. Overall, if computing power supply does not achieve large-scale breakthroughs after half a year, model competition will further tend to giants with self-developed hardware; one year later, the content ecosystem may enter a dual screening period of "high quality + low cost", with capital and supervision. Will jointly shape a new round of industry landscape. The beneficiaries are companies with complete closed-loop computing power and companies that can implement multimodal technology into production tools. The losses are small and medium-sized model providers who rely on external computing power and content creators who blindly chase traffic.

The dark side of the moon Kimi suspended new user subscriptions due to tight computing power. This is undoubtedly a microcosm of the current AI craze, but there are many aspects worth examining behind it. Claiming "support far beyond expectations" and "unexpected computing power challenges" may sound like surprises, but for a technology-centered company, especially at a time when AI computing power needs are so obvious, does this "unexpected" expose its insufficient prediction of market size and growth rate? Or is this a deliberately created sense of scarcity to pave the way for subsequent price increases or more refined user stratification? Historical experience tells us that when a technology or product rapidly inflows into a large number of users, but the underlying computing power or infrastructure cannot keep up, it will often be accompanied by a decline in service quality, a shrinking user experience, and even uneven allocation of resources on the platform side. Eventually lead to user loss. KimiK3's "performance close to GPT-5.6 Sol" is exciting, but if the subsequent service experience cannot keep up, how long can this leading position last?

Lin Dahua, chief scientist of Shangtang Technology, regards "multimodal" as "the next battlefield after Coding," which is in interesting contrast to Kimi's computing power crisis. On the one hand, everyone is competing for basic computing power, and on the other hand, some companies are drawing up more ambitious technical blueprints. Multimodal AI, which allows AI to understand and generate various forms of information such as text, images, audio, and video, sounds like a leap forward in AI's capabilities. But the saying,"Only when it can be delivered for commercial use is true productivity" also applies to multimodal. Many of the various AI applications currently emerging on the market remain in the concept demonstration or small-scale trial stage. There are still a few products that can truly be implemented on a large scale, solve practical business problems, and bring considerable returns. Is Shangtang's conclusion based on a profound insight into the future development trend of AI, or is it another pursuit of the "next outlet"? What data barriers multimodal technology will encounter in practical applications, model fusion problems, and how to ensure the authenticity and reliability of the content it generates are all potential risks concealed by the light of the "next battlefield".

Finally, the "big ebb tide" signal of AI comics is a loud wake-up call, reminding us that in the AI-driven innovation wave, we must not only look at the glamorous "new applications", but also pay attention to the sustainability of its business model. Sex and true value. The article bluntly stated that "the bubble has burst" and pointed out that "the production of affirmative action meets with distribution privileges" and that "the outlet created by big factories and ended with their own hands." The lesson here is that technological breakthroughs (such as AI-generated content) are not directly equated to commercial success. When capital and media attention quickly pushes an area to a climax and its commercial closed loop has not yet been truly established, the bursting of the bubble is almost inevitable. For investors and entrepreneurs, they cannot be carried away by the excitement. They need to more calmly evaluate the actual application scenarios of technology, the real needs of users, and the clarity of profit models. The case of AI comics may allow us to be more cautious and rational when looking at other AI applications, especially those AI products that claim to "subvert" a certain industry.