Can a single adjustment demonstrate the true capability of Moore Threads?

On September 3, at the performance briefing of Moore Threads, the company confirmed that it had onboarded leading internet enterprises and telecom operator clients in the first half of the year. Its semi-annual report also showed that its revenue reached 1.736 billion yuan, a year-on-year increase of 147%. 

On September 7, 25.7745 million restricted shares were lifted from the trading ban, and the share price of Moore Threads saw its largest adjustment since its listing. 

In the same September, the same company faced two completely different pricing outcomes. So the question arises: what exactly was being sold off in this adjustment? 

A Liquidity Event

Let's start with the shareholding structure. Before the lifting of the ban, the total tradable shares of Moore Threads were only 30.2255 million, accounting for about 6.4% of the total share capital — an extremely tiny free float. On September 7 alone, the free float expanded by 85%, and this batch of shares were the offline allotment shares allotted at the issue price of 114.28 yuan 9 months ago, with floating profits of multiple times, and the demand to realize returns was clearly recorded on the books. 

An ultra-small trading pool faced a supply of shares worth tens of billions of yuan, and price fluctuations were almost inevitable. On that day, the trading volume reached 3.467 billion yuan with a turnover rate of 14.72%, and the profit-taking orders and supporting orders completed full handover. 

So who exactly was selling? The answer is hidden in the issuance structure. All of these lifted shares came from offline allotment, and the allotment results of the company's IPO show that Class A investors — public funds, social security funds, pension funds, annuities, insurance funds, QFII — accounted for as high as 98.44% of the allotted shares, a considerable proportion of which are new share subscription strategy products under public funds. 

The profit model of such funds determines their behavior pattern: they get shares at the issue price, realize returns after the lock-up period expires, and roll over to participate in the next round of new share subscription. "Selling" has been written into the product contract from the first day of position building. The 9-month lock-up period is exactly the standard cycle for new share subscription strategies. 

In other words, the main sellers on September 7 are most likely not long-term funds that cast a no-confidence vote on the company's fundamentals, but new share subscription funds that must exit when the period expires — they sell not because they are not optimistic, but because "selling" is inherently the last link of this strategy. 

This is the most essential difference between the share handover after the ban lift and general reduction: reduction is the pricing of the company's future, while the realization of new share subscription is only the implementation of rules. The 14.72% turnover rate also shows that there is an equal amount of supporting funds on the other side of the selling pressure — the transfer of shares from new share subscription accounts to allocation accounts is exactly the inherent meaning of the expansion of the free float. 

The company's response on that day was also very straightforward: it called on investors to view the lifting of restricted shares rationally, stated that production and operation continued to improve, and the new product of the "Huagang" architecture will be launched within the year. 

From this perspective, the nature of this adjustment is closer to a "liquidity event": it tests the shareholding structure, not the operating quality. 

Don't Only Focus on One Single Stock

Pull the camera back, what is happening in the domestic computing power track? 

Tencent's capital expenditure in the second quarter was 52.78 billion yuan, a year-on-year increase of 176%, of which about 51.4 billion yuan was the advance payment for computing power procurement; Alibaba's capital expenditure in the latest quarter was 67.678 billion yuan, a year-on-year increase of 75%, and it has invested a total of 190 billion yuan in its 380 billion-yuan AI investment plan; Baidu's single-quarter capital expenditure was 11.39 billion yuan, a year-on-year increase of nearly 200%, accounting for as high as 36% of its revenue; according to media reports, ByteDance even raised its 2026 capital expenditure plan to the level of 200 billion yuan. 

The high degree of consistency of these moves and the firmness of their strength are extremely rare in the internet industry over the past two decades. Large tech enterprises have all aimed at the same target: computing power. According to statistics from Caixin, the capital expenditure of internet manufacturers in 2025 has accounted for more than 65% of the national total, and it is expected to rise to 75% in 2027. 

This is the first layer of dividend for domestic GPUs: explosive demand. 

The second layer of dividend lies in the supply pattern. The research report of Soochow Securities shows that the share of domestic AI acceleration cards exceeded 40% for the first time in 2025; according to industry surveys, the demand for domestic AI chips in 2026 is about 4 million units, while the actual delivery is only about 3 million units, and the orders of downstream intelligent computing centers have been scheduled for three years later. This is a typical "seller's market" — the shortage itself is the best screening for qualified entrants. 

The third and most profound layer is the shift of verification scenarios. The competition for domestic computing power is moving from "sample delivery testing" to the stage of "real business load": whoever can be integrated into the production system of large tech enterprises can truly get the admission ticket. 

Moore Threads' Reputation Is Built Through Real Operation

Following the clue of "real load" to look at Moore Threads, you will find a timeline submerged by short-term fluctuations. 

The earliest connection with JD Cloud can be traced back to the unveiling of the first kilowatt-level intelligent computing center in December 2023, when the chief AI scientist of JD Cloud attended the roundtable; in July 2024, the Ku'e 10,000-card cluster was released, and JD Cloud took the stage to share; at this year's WAIC, Cao Peng gave a public speech, and the joint research of the two sides around the JoyAI large model, the enterprise-level reasoning platform and the full-link embodied intelligence service have been fully launched. 

Three years of contact, deepening step by step — this is not the rhythm of concept speculation, but the rhythm of engineering verification. 

The cooperation with Zhipu AI is more continuous: from the public service of GLM-4.6 through the Zhipu MaaS platform in September 2025, to four consecutive "Day-0" adaptations of GLM-5, 5.1, 5.2 and 5.3-Flash, the new product release rhythm of the leading large model manufacturer has been perfectly aligned with the adaptation rhythm of this GPU company. 

On the training side, ecological partners have completed the training of the MoE-236B basic large model with more than 25 trillion Tokens of corpus from scratch based on the S5000 cluster, and the loss difference on the 10,000-card cluster is only 0.62%; on the reasoning side, the PD heterogeneous solution with Converge Technology has been put into production operation, undertaking the real Token traffic of the official business of leading model manufacturers. 

For a GPU company, these assets are more valuable than any parameter sheet. Because trust in the chip industry has a unique attribute: it can only be obtained by running on other people's production lines, and cannot be proved in the laboratory. When large model manufacturers entrust you with the first adaptation of their new models, and internet giants entrust you with their real traffic, they essentially bet their business continuity on the stability of your products. 

Back to the industry coordinate system: according to Caixin's sorting out of the "five rising stars of GPU", Moore Threads ranked first with a revenue of 1.74 billion yuan in the first half of the year; it is also the only domestic manufacturer that has actual revenue from three product lines including AI computing, graphics rendering and consumer-grade products, and together with Huawei, it is one of the domestic implementers that have completed the verification of 10,000-card level training clusters. 

To be honest, some things have indeed changed. After the free float expanded by 85%, this stock has changed from an "emotional target with scarce shares" to an "allocation target that institutions can enter and exit normally", and the pricing logic will switch to the fundamental track accordingly — the fluctuation may not be smaller, but the anchor will be more solid. 

The unchanged things are equally clear: the capital expenditure of large tech enterprises will not shrink due to a one-day adjustment, the 4 million-unit demand gap will not narrow due to one ban lift, and the adaptation and training running on the production lines of large tech enterprises will not be interrupted due to share price fluctuations. 

The capital market votes on prices every day, while trust in the industry accumulates by the year. On September 7, the two time scales collided: the short-term one affected the market trend, and the long-term one is still on the way. 

The way to understand hard technology companies is not to count how many adjustments they have experienced, but to count how many production lines of other enterprises their products have been put into. The former is sentiment, while the latter is real value. 

This article is from the WeChat official account "florayang01" (ID: daily-case), the author is Tide Business Review, and it is released with authorization from 36Kr.