Friday, June 26, 2026

Looking at these signals today, there is only one word in my mind: fragmentation.

On the one hand, the AI industry is madly blowing up Agents, and the Doubao Professional version of the online office task model, and an AI that can work 200 million a day. It sounds like the work of all mankind will be automated tomorrow. On the other hand, the post on Reddit made me happy. The senior engineer complained bitterly about AI cutting off the lifeline of the database with a knife. The title was "Don't let AI touch the production environment." When you look at these two together, it's too fucking real. Agent is a thing that is smooth and smooth during the demo presentation. It is actually launched in the production environment. Once the authority is released, it can clear your data assets for several years with one click in minutes. It's not that AI is not good, but that current Agents are still far from being able to achieve reliability and security boundaries. It is commendable for Doubao to dare to push this ability to 200 million users, but I advise all developers not to rush to give it the database root password, let it run in the sandbox for half a year first.

Another thing worth talking about is the wave of price increases for storage and chips. Apple's entire product price increase by 20%, Micron's financial report burst, and HBM4 delivery exceeded US$1 billion. Behind this is the crazy expansion of AI data centers that has sucked up all memory chip production capacity. Interestingly, at this juncture, OpenAI also released its own chips, and SK Hynix is going to go public in the United States to raise US$29.4 billion. What is the concept? This shows that in the next two years, the arms race in AI infrastructure will only become more fierce. Not only will the cost of computing power not drop, ordinary consumers will have to pay for electronic products. The reason for Apple's price increase is very straightforward. Cost pressure, which translates to: If you pay an extra $200 for an iPad Pro, it will be a contribution to the AI industry.

One more thing must be said. Anthropic has sued four China AI companies in the past four months, and Ali's latest turn is. This is no longer technological competition, but a naked industrial encirclement and suppression. OpenAI and Anthropic are both preparing to go public. Talent from Google is still moving abroad. The entire AI industry is changing from a scientific research competition to a capital game. In the future, whoever has chips, computing power, and legal team will be qualified to go to the table.

The most interesting signal today is the resonance between Apple's 20% price increase due to memory shortages and Anthropic's serial lawsuit against the China model team. This may seem to be an isolated incident between consumer electronics and legal rights protection, but in fact it jointly outlines two high walls in the second half of the AI industry: the extreme run on physical resources and the comprehensive tightening of intellectual property rights.

The siphon effect of AI computing power centers on high-bandwidth memory has spread from the cloud to the end side. Apple's rare sharp price adjustment due to hardware costs shows that the computing power arms race is overdrawing the global semiconductor supply chain. Micron's surge in financial reports and locking in hundreds of billions of long-term agreements means that the storage giant is transforming from a cyclical supplier to a tax collector for AI infrastructure. When memory and computing power became hard currency, OpenAI launched its self-developed chips and Apple reconstructed its M-series chip roadmap. In essence, it was the giant's attempt to get rid of the upstream bottleneck and internally absorb the cost of computing power. Within six months, we will see consumer electronics forced to compete with AI servers for production capacity, hardware inflation will become the norm, and the profit margins of terminal manufacturers without the ability to develop their own chips will be extremely compressed.

Beyond the physical wall, Anthropic's distillation lawsuit against Ali erected a data wall. As high-quality training data dries up and computing power costs soar, head model companies no longer tolerate latecomers using shortcuts to generate fine-tuning for nothing. The lawsuit marks the complete end of the free lunch for the big model open source ecosystem, forcing the China team to follow a completely independent data and pre-training route. In the short term, shell teams that rely on closed-source model distillation will face life and death, while companies with computing power reserves and private high-quality data will face a value revaluation.

At the same time, the bean bag professional version pushed the Agent to 200 million daily activities, marking that the large model has officially moved from chatting to working track. But the bloody and tearful post on Reddit complaining about AI bombing is a dangerous omen. When AI moves from generating text to performing database operations, trust becomes the most expensive cost. The winners of the future will no longer be the models with the largest parameters, but the engineered middle layers that provide secure sandboxes, cross-warehouse memory, and context compression. The outbreak of Agent is inevitable, but whoever can put the reins on this beast can truly get tickets to the enterprise-level market.

Apple's all-round price increases and Micron's record-breaking financial report reveal a cruel reality. The boom in AI infrastructure is draining consumer electronics supply chain resources. Storage giants rely on high-bandwidth memory to earn ultra-high gross profits, and Hynix is also eager to go public in the United States to cash in, but the cost of the computing power arms race was eventually passed on to ordinary consumers who buy computers and tablets. This is not technology universal benefit, but an implicit taxation of upstream on downstream. In history, every hardware shortage has created a bubble. When the growth rate of capital expenditure in giant data centers slows down, the prices of these high chips will inevitably face reassessment.

Compared with the huge profits at the hardware layer, the prosperity of the application layer is more like an illusion that cannot withstand scrutiny. Doubao launched a high-profile intelligent office, and the scale of specific intelligence has been blown to a trillion-dollar scale. However, overseas engineers complained that AI blew up the database, which is a true reflection of the production environment. The big model is omnipotent in demonstrations. Once it takes over the real business, its uncontrollable destructive power is enough to make the company pay a heavy price. What's even more ironic is that the leading domestic company is still being sued by its overseas counterparts for model distillation. Before the moat at the ground level is built, it is rushing to race for land on the application level.

In this carnival, the ones who make money will always be the computing power giants selling shovels and the capital that drives up valuations. Companies that blindly connect AI to core businesses and consumers forced to pay for hardware premiums have become the most silent payers. When the tide recedes, those agents that claim to subvert everything often leave behind error codes and shrinking financial statements that need to be remedied manually.