Thursday, July 30, 2026

The most interesting thing today is not which model has hit the list again, but that several signals put together to tell a story about the AI industry's "end of the first half and start of the second half."

Let's talk about Microsoft's financial report first. Cloud business exceeded expectations, but quietly cut capital expenditures in fiscal year 2027 from 190 billion to 175 billion, which rose by 8% after hours. The meaning of Wall Street's voting with your feet is clear: you people have been burning cards long enough, it's time for us to see the payoff. On the same day, Hynix's profit rose by 557%, but its stock price fell instead, and the storage sector collapsed collectively. This is not a contradiction, this is the market saying that "growth has peaked." Coupled with the fact that optical communications and storage leaders such as Micron and Coherent have fallen together, the entire AI infrastructure chain is sending the same signal: the arms race that has been running for two years has begun to slow down. It's not that I stopped throwing, it's that the pace of throwing is changing.

Then you look at the application layer. Tencent's WorkBuddy achieved 12 million DAUs in three to four months, and Ali launched Qianwen Office hard overnight. What does it mean that two major manufacturers are fighting each other on the track of office agents? This shows that the dividends of infrastructure are shifting to the application layer. It's not that there are not enough cards, but that "cards alone are not enough." Whoever can truly put AI into the daily workflow of migrant workers will get the next ticket. But to be honest, these office intelligences are still doing the icing on the cake of "helping you organize meeting minutes" and "helping you write weekly reports." They are still far from truly changing the productivity structure. 20 million months of work sounds scary, but the Retention rate and paid conversion are the real questions.

Another thing worth talking about is security. Claude's shared chats are included by Google, and users 'sensitive conversations and Artifacts are directly exposed in search results. This is either a profound technical loophole, or the difference between "sharing" and "disclosure" was not clearly thought through when designing the product. ChatGPT has made the same mistake before. You are a company that claims to be a "security super intelligence" company, but you can't even understand the most basic link rights management. Ilya took $5 billion from Lao Huang. Should you divide the budget to the product security team?

At the same time, more than 1100 employees of AI companies jointly petitioned the government to control the pace of AI development. This is quite ironic. Anthropic's CEO took the lead in signing, but his own product is revealing user privacy. The matter of security has never been solved by petition letters, but by specific authority control, data isolation, and index blocking in engineering practice.

Meituan's batch of stolen swipes in the early morning is also worthy of vigilance. Secret-free payments and off-site write-offs make the attack link too slippery. While platforms pursue conversion rates and convenience, security design is always the final hole to fill. This problem will only become more serious in the AI era, because AI agents naturally need to proxy users to perform operations. Once the authority boundary is improperly designed, stealing group purchase coupons is the smallest loss.

My judgment is that starting from the second half of 2026, what is really valuable is not the models that can write poetry, but the people who can handle security, permissions, and data flow cleanly in engineering. The scuffle at the application level has just begun, but the one that survives will not be the one with the most functions, but the one with the most reassurance.

The most noteworthy thing in today's signal is that the AI office track has suddenly entered a white-hot stage. WorkBuddy achieved 12 million daily activities in three months after its launch. Ali immediately launched Thousand Questions Office Hard Firm, and Tencent even began to discuss whether it could enter the "Famous Products Hall". There is an underlying trend behind this: competition for AI products is shifting from universal dialogue to deep binding of vertical scenarios. It is not uncommon for universal models to have tens of millions of monthly activity, such as bean buns, Qianwen, and Yuanbao. However, WorkBuddy, a product that focuses on office scenarios, can reach the same level in three months, which means that users are interested in "AI helps me complete specific tasks" The need for "work" is far greater than that for "AI chatting with me." Office scenarios are extremely sticky. Once users are accustomed to writing documents, doing PPTs, and organizing data in WorkBuddy, the migration cost will be high. This is another enterprise service portal that Tencent has obtained in the AI era after Enterprise WeChat and Tencent Meeting. Ali was forced to rush to fight, which means that the first window period of this track is closing. The next six months will be a battle for corporate office agents. Whoever first occupies the user's workflow will get the next round of ecological admission tickets.

Another signal worth digging into is the ban on autonomous driving "little blue lights". On the surface, it is a compliance issue, but on the deeper side, smart driving is shifting from "showing off technology" to "taking responsibility." The essence of the small blue light is that car companies want users and passers-by to know that "this car is driving automatically", but the attitude of the regulatory authorities is very clear: you don't need to tell others that the car is driving automatically, you just need to ensure that no accidents occur when driving automatically. This is directly related to the previous background of frequent assisted driving accidents and vague definition of responsibilities. When a car company uses lights to imply that "I am driving automatically", it is essentially transferring safety responsibility-if there is an accident, the user may say "I thought the car could drive itself." Banning the small blue light is pushing "technical show-off" back to the "safety bottom line" and forcing car companies to invest more in functional safety, rather than making a fuss about marketing gimmicks. This signal is superimposed on the "secret payment" loophole in the Meituan theft incident, and we can see a more macro trend: whether it is AI driving or AI payment, the technology industry is experiencing a shift from "running first and then wiping the butt" to "wiping it first". The granularity of supervision is getting thinner, and the compliance costs of enterprises are rising, but in the long run, this is a good thing for users.

Third, Nvidia's $5 billion investment in Ilya creates an interesting contrast with Microsoft's downward revision of capital expenditure guidance. Ilya gets the money to build a "security super intelligence", and Nvidia is willing to spend US$5 billion not because of a shortage of computing power, but because AI security itself is becoming a huge market. However, Microsoft is lowering its capital expenditure forecast for the next year, and it has risen 8% after hours. This shows that Wall Street's attitude towards AI infrastructure is diverging: the marginal return on computing power investment is declining, and soft capabilities such as safety, compliance, and interpretability are becoming new valuation anchors. Hynix's profit rose 557% but its stock price fell. The same logic is true-the market no longer just looks at revenue growth, but at the quality and sustainability of growth. In the next six months, the narrative of the AI industry will shift from "who can build more computing power" to "who can make AI safer, more controllable, and more practical." In this process, security companies, compliance tools, and edge computing will be favored by capital, and pure computing providers may need to tell a story again.

In today's technology signals, there are three themes that are particularly worth pulling out and putting together: the AI bubble and real value, the trade-off game between technological security, and the office ecological war between giants. They may seem independent, but in fact they all point to the same question-who is really benefiting and who is taking risks behind technological progress?

Let's talk about the bubble and value of AI first. Ilya's SSI received a US$5 billion investment from Nvidia, and Nvidia also promised to increase its computing power tenfold within 12 months. This looks like another "Manhattan Project"-style gamble. But looking back at history, similar stories are not uncommon: the VR craze in the 2010s and the metaverse in the early 2020s were all products of capital and media hype, and eventually left behind feathers. SSI's core selling point is "security super intelligence", but "security" itself is a vague concept. Who will define it? Who will verify? More importantly, where is the return on investment of this $5 billion? The biggest problem in the AI field at present is not the lack of computing power, but the lack of a clear business model. OpenAI spends money in exchange for growth, and Anthropic's shared chat records are included by Google to expose privacy risks, all reminding us that AI's "super intelligence" may be just a tool for capital storytelling, while the real risks-data security, ethical boundaries, business sustainability-have been ignored intentionally or unintentionally.

Let's look at the trade-off game of technological security. The "little blue light" for assisted driving has been disabled, ostensibly because it does not comply with national standards, but behind it is the eternal tug of war between technical safety and user experience. Car companies use "small blue lights" to identify the assisted driving status, which is essentially a "trust agent"-making users mistakenly believe that the system is more reliable than it is. This is exactly the same as the incident where Meituan accounts were stolen in the early morning: behind the technical convenience is the lack of security mechanisms. Meituan's secret-free payment function makes it convenient for users, but it also allows illegal companies to take advantage of it; the "little blue light" makes users rely on assisted driving, but if something goes wrong, who will be responsible? Both incidents point to the same problem: there is always an irreconcilable contradiction between "user-friendliness" and "safety and reliability" of technology. The role of regulators is often to passively intervene after an accident, rather than participating in defining boundaries at the beginning of technical design.

Finally, the office ecological war between giants. WorkBuddy achieved 12 million DAUs in three months. Tencent even discussed whether it could impact the "Famous Products Hall", while Alibaba quickly launched Qianwen Office Hard Work. Behind this is the "arms race" of AI office tools-giants are no longer satisfied with general AI, but have to go deep into vertical scenarios and redefine productivity through "agents." But who are the real beneficiaries of this war? Users may just be forced to switch between different platforms, while the real value-data, user habits, business models-is monopolized by giants. WorkBuddy's success is more like a "victory" for Tencent's internal ecosystem than a contribution to the entire industry. Ali's Thousand Questions Office is more like a "defensive counterattack" against competitors. This kind of convoluted innovation may end up just making the giant's moat deeper, while users still pay for it in the "free" trap.

These three themes converge into one point: Behind technological progress is the complex game between capital, regulation and users. Capital pursues stories and rewards, supervision pursues safety and order, and users pursue convenience and experience. But in this chain, the real risks and costs are often passed on to the most vulnerable party-users. They may have nowhere to defend their rights after the Meituan account was stolen and swiped, may bear the risk of an accident under the misguidance of a "trusted agent" for assisted driving, or may be forced to become "data workers" for giants in the arms race for AI office tools. And those seemingly "revolutionary" technological breakthroughs, such as SSI's "Security Super Intelligence", may just be stories told by capital to capital, and their true value is far from here.