Friday, July 3, 2026

One of the most worthy words to talk about today is actually: Is there too much computing power? Meta wants to sell computing power, SK Hynix plunged 14% in a single day, and DeepSeek collapsed. When these three things are put together, the taste is very strong. When Meta sells computing power, it says on the surface,"Oh, we have invested too much and want to make up for it." In fact, it tells you that they are no longer so anxious about the consumption of computing power internally. The big factory used to hoard cards like crazy, but now it finds that it can't use them all. Instead of leaving them to eat ash, it's better to take them out and cash them out. This directly slapped the narrative of "computing power will never run out". South Korean chip stocks responded more directly, and the capital market has already voted with its feet-the highlight moment of AI hardware may really be over. The collapse of DeepSeek is also a microcosm. The number of users increased, and the server couldn't withstand it. This shows that no matter how cheap and efficient the model is, the underlying computing infrastructure is still fragile when dealing with real concurrency. Therefore, for ordinary developers at present, stop using "unlimited computing power" to design products. Cost control and flexible expansion are the real skills.

Then another more realistic signal is that the model is beginning to compete for cost performance. Claude Sonnet 5 was hit by negative reviews as soon as it was launched. Domestic reviews couldn't beat Qianwen and Minimax, and the price was even more expensive. In fact, this is not that Sonnet 5 is too bad, but that the domestic models are improving too fast. The release of GLM-5.2 is open source, and the popularity of Hugging Face directly exploded, indicating that the community has a strong demand for open source, low-cost models. Anthropic's current position is very awkward. If the closed-source model cannot continue to open up generation gaps, why should users pay? Especially for Agent tasks, developers pay more attention to reproducibility and cost, and an open source alternative is enough. For ordinary developers, when playing AI applications now, they prefer open source models as a baseline. When things go smoothly, they will consider closed-source upgrades, and don't be a fool.

Finally, there is another undercurrent in today's signal: AI is moving from "generation" to "execution." Kelin AI raised US$3 billion to directly generate video;OpenMontage was released open source, claiming to turn the AI coding assistant into a complete video studio. These are no longer simply "writing a copy" or "drawing a picture", but automated workflows constructed into a system. Claude Sonnet 5 also emphasizes Agent capabilities. Behind GLM-5.2 is the Agent framework of the ZCode platform. Even Tesla limits employee AI spending to $200 a week, essentially preventing AI from being used as a toy. This marks a stage of change: before, everyone thought that AI was like a search engine and asked whenever they wanted; now it is closer to "task outsourcing"-you need to design inputs and processes well, and AI runs automatically. Engineering tools such as the Skill system and Codex transfer station have begun to become just needed. Put it in human terms: Stop writing prompt and start writing pipeline. Whoever has a stronger and more efficient automation chain will survive.

Overall, don't be fooled by the ups and downs of the market. AI has not receded, but it is not a golden rain. Now is the pragmatic stage of striving for implementation, cost, and engineering capabilities. For developers, either embracing open source and fighting costs, or building automated workflows to improve efficiency is the direction where they can truly reap dividends in the next two years.

The most deafening part of today's signals is actually the collision of two things: on the one hand, Meta announced that it was selling excess AI computing power, causing South Korea's KOSPI index to plummet by 8%, and SK Hynix to evaporate its market value of US$160 billion a day; on the other hand, DeepSeek "collapsed" and hit the hot list, users complained that it couldn't be opened-these two signals together outline the critical turning point that the AI industry is experiencing, what I call "supply-side differentiation."

Meta's sale of computing power is interpreted by the market as a signal that AI has excess computing power, but what I see is a deeper structural change. Meta is not the first to do this. Previously, Microsoft and Google have also successively adjusted their cloud computing power strategies. The core logic is that the dividend period for large-scale model training is narrowing, and reasoning requirements are beginning to diverge. When training giants found that they couldn't use up all the H100/Gaudi they had spent hundreds of billions of dollars hoarding, and small and medium-sized enterprises couldn't afford it, selling excess computing power through cloud services became an inevitable business choice. This just shows that AI infrastructure is moving from an "arms race" to a "commercial realization" stage-but this does not mean that AI is ebbing, but that the demand structure has shifted from extensive expansion to refined matching. SK Hynix's plunge is more of a collective panic in the market about semiconductor cycle expectations. AI storage demand is indeed slowing down, but storage demand for mobile phones, cars, and edge devices is filling this gap. This is a new variable that should be paid attention to in the second half of 2026.

DeepSeek's collapse may seem to be an operation and maintenance accident, but in fact it exposes another trend: the traffic of C-side AI applications is exceeding expectations. If there is really an overall excess of computing power, why is a domestic model squeezed by users? This shows that models and products that are really used at high frequencies are facing insufficient computing power. It is also worth noting that Claude Sonnet5 was exposed by the Chinese community as "unable to beat Qianwen and Minimax" one day after its launch. Anthropic's proud agent capabilities have been overdone in front of domestic models. Behind this is a clear signal that model competition has shifted from "parameter competition" to "project implementation"-when basic capabilities converge, who can serve users at a lower cost and more stably can survive.

Put the third signal in series: Musk publicly stated that the early stages of mass production of the Optimus robot will be slow because all technology is new. This is exactly the same as Tesla set a cap on AI spending on employees last week ($200 a week), sending the same signal: in the process of hard technology commercialization, the "valley of death" from the laboratory to the factory is wider than imagined. Keling AI completed US$3 billion in financing, the exoskeleton robot built by Beihang's team received nearly 100 million yuan in angel rounds, and the Home Health Care Robot Company received strategic financing-these financing events are in sharp contrast to Musk's pessimistic statement. The reality is: Capital is still investing in AI and robots, but giants are starting to step on the brakes, while entrepreneurs are still accelerating. In the second half of 2026, we will see a large number of AI startups fall due to cash flow breaks, but those companies that truly integrate AI into specific scenarios and solve actual pain points will instead experience the double dividends of price reductions in computing power and open source models. Rise.

One easily ignored signal is the news on Hacker News that "for the first time a cell built from scratch can grow and divide" and the Spanish government blacklist banning Palantir. This reminds us that the boundaries of AI are extending from the digital world to life sciences and geopolitics. When AI infrastructure changes from scarce goods to commodities, the real barriers will be the teams that can master these tools and solve real human problems. Today's plunge is not the end, but a coming-of-age ceremony for the end of puberty in the AI industry.

Among today's signals, the most eye-catching thing is not that an AI model has been upgraded, but that South Korea's KOSPI index has plunged 8% in a single day, SK Hynix has lost US$160 billion in market value in one day, Meta's demand for computing power, and DeepSeek's collapse. I think these are not separate events, but different sides of the same story: the AI industry's computing power myth is undergoing its first real stress test.

Let's talk about Meta's computing power. Can you imagine that in 2024 and 2025, someone will publicly say that they have bought too many GPUs and need to sell them? At that time, the entire circle was just grabbing cards and tanking cards, for fear that if they were slow, they would miss the times. Meta is now saying that it will sell its remaining computing power to the outside world. It is not so much to expand its business, but rather to admit an embarrassing fact: their own AI needs cannot meet the consumption of hundreds of thousands of GPUs. This is not a brilliant foresight, but a huge mismatch between capital expenditures and actual applications. Some people in the industry say that excess computing power is a misunderstanding. Looking at the tragic situation of SK Hynix and the entire South Korean semiconductor sector-an 8% drop in a single day, and the market value of HBM memory's main supplier has been reduced in one day-the market has used its own way. The answer is given. This fall method is not a misunderstanding, but a real panic.

Combined with DeepSeek's collapse. On July 2, a large number of user feedback could not be accessed. The rapid growth of users makes the server unable to withstand it. This is not a big problem in itself, but it is very subtle when viewed in the same time window. On the one hand, top cloud service providers are releasing computing power, and on the other hand, services of star AI companies are intermittently offline. What does this mean? This shows that the matching of supply and demand for the entire AI infrastructure is simply unbalanced. There is local excess and local shortage. No one has truly mastered the accurate dispatch capabilities. This is just like during the Internet bubble, when optical fibers have been laid down, but the access and use of the last mile have not kept up.

There is also a signal that is easily ignored: Tesla has been exposed to limit employee AI spending to a $200 per week cap. If even a company like Tesla that relies on AI to achieve autonomous driving and robots relies on quotas to control the cost of AI use, how much do you think companies that buy huge amounts of GPUs can recoup based on selling power? This $200 per week limit shows more than any expert report how expensive enterprise-level AI is to use and how vague the benefits are.

Have you noticed that along the entire chain, shovels are still making money? The stock prices of SK Hynix and Nvidia have soared before, but those who actually use shovels to mine, whether they are Meta or DeepSeek, are either selling computing power at a discount or Carrying high user growth costs. Historically, Cisco encountered the same situation in 2000: companies frantically purchased network equipment, only to later discover that they had bought too much and could not digest it. The fuse of this storm may have been ignited by Meta's action.

As for Kelin AI that is still raising funds at a valuation of US$18 billion and preparing for an IPO, I can only say that investors 'final carnival often occurs when others start to leave.