本文来源:https://mp.weixin.qq.com/s/liZK2Uys3xgwcUV_eiaM0A
若有版权问题,请来信告知:13951431913#139.com 或在右方提交评论。

原创 | 作者:罗超
上周五Kimi K3发布,官方宣称,虽然其整体表现仍落后于最强的闭源模型 Claude Fable 5 和 GPT-5.6 Sol,但它在整套评测中展现出前沿水平的能力,并稳定超过了其他所有模型。
港股AI板块随即遭遇“黑色星期五”,智谱暴跌28.49%、MiniMax跌了15.63%,此外壁仞科技、广合科技、建滔、胜宏科技、长飞光纤、中芯国际等等AI基建股均下跌超过10%,有媒体迅速出来认领——这是Kimi K3干的。

逻辑链条大概是这样的:Kimi K3太强了,2.8万亿参数,全球首个开源3万亿级别模型,在Arena.ai编程排名上冲到了全球第,投资者担心模型这么厉害会导致算力需求减弱,于是抛售半导体和AI概念股,这套说辞似曾相似,看上去有点合理,我看到不少人都信了。
这完全是扯犊子。
只需要看一眼周五晚上的美股就知道了:Kimi K3发布的同一天,费城半导体指数(SOX)盘中一度大跌5.7%,收盘收窄至1.6%,英伟达只跌了2.21%,所以根本不存在什么“全球半导体板块大跌”。跌的是A股和港股,不是美股。非要说是K3导致全球半导体板块大跌,那就是先打枪再画靶,还乱画。
因为就算Kimi K3真的强大到让人焦虑,真的像有些媒体人吹嘘的创造了“第二个deepseek时刻”,就算周五晚上美股半导体板块真的暴跌了,K3导致全球半导体板块暴跌的逻辑也不成立。
2025年1月DeepSeek R1横空出世,用远低于行业平均的算力成本展现出逼近OpenAI o1的性能。市场当时的反应是这样的:英伟达单日暴跌17%,市值蒸发近6000亿美元。当天费城半导体指数(SOX)跌9.19%。这才是真的“暴跌”好吧?
不过接下来一个月,英伟达、SOX指数都很快收服了失地,因为市场完全搞反了DeepSeek的影响。黄仁勋当时是这样说的:投资者误将DeepSeek的进步视为AI公司不再需要高强度算力的信号,而“事实恰恰相反”。接下来一年半,半导体板块延续疯涨趋势。
所以说,更强大的开源模型从来不会抑制算力需求,它只会推动AI应用普及,进而创造更多算力需求。DeepSeek不是英伟达的天敌,Kimi K3也不是。
况且这波半导体板块的调整,早在Kimi K3面世之前就已经开始了。
截至7月17日收盘,费城半导体指数较6月高点累计回撤20%,正式进入技术性熊市。中证芯片产业指数回撤约30%,台湾半导体指数回撤28%。这些下跌发生在K3发布之前,不是之后。
所以稍微关注半导体行业的人就能看出“K3导致半导体板块下跌”的说法是瞎扯淡。
这波半导体调整真正的导火索是什么?至少有这两方面的显著影响:
一是Meta这颗“搅屎棍”掀桌子来了。7月初,在AI这波掉队后,Meta出来掀桌子,先是宣布对外出租闲置AI算力,市场的第一反应是“算力买多了”,Meta一直是算力的大买家,现在要将GPU拿出来出租,是不是说算力过剩了?紧接着扎克伯格在内部员工大会上承认,过去四个月AI智能体的开发“没有像我们预期的那样加速”,公司在新组织架构上的押注“尚未结出果实”。
所以市场担心AI商业化应用的验证路线跑不通,AI基础设施投资存在过热,Meta这样的大买家可能会缩减资本开支,基建类(包括芯片、存储、数据中心)都跟着跌了一波。

二是巴菲特给当前AI市场投资定性了:过热,甚至赌场化。7月16日,巴菲特在接受CNBC专访时直言,当前市场被短线交易主导,长期价值投资被边缘化:
“当贪婪与赌博情绪盖过理性,泡沫一定会诞生。我从不预判短期市场涨跌,只寻找能合理估算内在价值的企业。如果一家公司仅靠 AI 概念炒作,没有稳定持续盈利,无论股价涨得多猛,我都不会碰。赌场永远会催生浮华泡沫,但耐心的价值投资者只会彻底无视这类噪音。”
95岁的巴老爷子说AI市场投资在赌博,这是对估值逻辑的根本性质疑,影响比一个模型要大得多。
最后,智谱和MiniMax的暴跌,根本不需要任何外力因素,这个是我一直以来的观点:靠“市概率”撑起的市值迟早会现原形。
(图源:ChatGPT生成 by 罗超Pro)
就算Kimi不发布K3,或者K3表现拉胯,这两家公司也很难不跌。7月8日,智谱迎来首批限售股解禁,涉及股份约2568万股。解禁当天股价不跌反涨,收涨13.35%。但紧接着智谱以1588港元配售近2000万股新H股,募资314亿港元——参与配售的机构七个交易日后浮亏30%。2025年全年营收7.24亿元,经调整净亏损31.82亿元,市销率一度超千倍。
MiniMax更惨。3月股价最高触及1330港元,市值突破4100亿港元,超过百度等等深耕AI多年的科技大厂。到7月17日收盘,216港元,市值754亿港元,缩水超80%。而这一切发生在Kimi K3发布之前,其中解禁首日单日暴跌近18%。所以之前这80%的下跌找谁背锅?
AI市场从“稀缺性溢价”向“基本面定价”切换,故事讲完开始算账了,仅此而已。
所以如果你依然搞不清楚有些AI公司为什么暴跌,多看看巴菲特最近的这段专访吧。
话说回来,Kimi K3是好模型,让中国在大模型生态位又多了一张王牌。kimi、deepseek、qwen这些开源玩家你追我赶,让中国AI有希望凭借开源跑出一条不一样的路,这比智谱和MiniMax的暴跌更值得关注。
巴菲特那句话放在这里倒是挺合适:“市场会出现两种极端阶段,一种是优质机会扑面而来,多到让人应接不暇;另一种是耗费两三年时间,能找到一个靠谱标的就算万幸。”
大多数时间里,市场都处于后一种状态,狂热的人类终将再度回归到“常识”这件事情上来。
主题为「智能伙伴·共创未来」的WAIC 2026正在火热开展中! AI叙事路线从模型参数堆叠,转变到Agent生产力落地;异构协同、光子计算持续提高计算上限;具身智能加速应用,机器人到家进厂让物理AI变为现实。 雷科技WAIC探展团已抵达上海,直击AI产业化落地年度巅峰时刻,敬请关注!
Source: https://mp.weixin.qq.com/s/liZK2Uys3xgwcUV_eiaM0A. For copyright issues, please email: 13951431913#139.com or submit a comment on the right.
Original | Author: Luo Chao
Last Friday, Kimi K3 was released. The official announcement claimed that although its overall performance still lags behind the strongest closed-source models, Claude Fable 5 and GPT-5.6 Sol, it demonstrated frontier-level capabilities in the entire evaluation suite and stably outperformed all other models.
The HK stock AI sector immediately encountered a "Black Friday," with Zhipu plummeting by 28.49% and MiniMax dropping by 15.63%. Additionally, AI infrastructure stocks such as Biren Technology, Guanghe Technology, Jian Tao, Shenghong Technology, YOFC, and SMIC all fell by more than 10%. Some media outlets quickly claimed responsibility—this was done by Kimi K3.
The logical chain is roughly this: Kimi K3 is too strong, with 2.8 trillion parameters, the world's first open-source model at the 3 trillion parameter level, and it surged to the top of the global programming rankings on Arena.ai. Investors worried that such a powerful model would lead to a weakening of computing power demand, so they sold off semiconductor and AI concept stocks. This narrative sounds familiar and seems somewhat reasonable, and I saw many people believing it.
This is completely nonsense.
Just look at the US stock market on Friday night: On the same day Kimi K3 was released, the Philadelphia Semiconductor Index (SOX) dropped as much as 5.7% during the session, narrowing to 1.6% at the close. Nvidia only fell by 2.21%. So there simply was no "global semiconductor sector crash." What fell were the A-share and HK stock markets, not the US stock market. If one insists that K3 caused a global semiconductor sector crash, that is shooting first and then drawing the target, and drawing it randomly.
Because even if Kimi K3 is truly powerful enough to cause anxiety, and truly created a "second DeepSeek moment" as some media people hype, and even if the US semiconductor sector really crashed on Friday night, the logic that K3 caused the global semiconductor sector crash still does not hold.
In January 2025, DeepSeek R1 appeared out of nowhere, demonstrating performance approaching OpenAI o1 with computing power costs far below the industry average. The market's reaction at the time was: Nvidia plummeted by 17% in a single day, with nearly $600 billion in market value evaporating. The Philadelphia Semiconductor Index (SOX) fell by 9.19% that day. That is a real "crash," right?
However, in the following month, Nvidia and the SOX index quickly regained lost ground because the market completely misunderstood DeepSeek's impact. Jensen Huang said at the time: Investors mistakenly viewed DeepSeek's progress as a signal that AI companies no longer need high-intensity computing power, while "the fact is exactly the opposite." Over the next year and a half, the semiconductor sector continued its crazy upward trend.
So, more powerful open-source models never suppress computing power demand; they only promote the popularization of AI applications, thereby creating more computing power demand. DeepSeek is not Nvidia's nemesis, and Kimi K3 is not either.
Moreover, this adjustment in the semiconductor sector had already begun before Kimi K3 was released.
As of the close on July 17, the Philadelphia Semiconductor Index had retreated cumulatively by 20% from its June high, officially entering a technical bear market. The CSI Chip Industry Index retreated by about 30%, and the Taiwan Semiconductor Index retreated by 28%. These declines happened before the K3 release, not after.
So anyone paying slight attention to the semiconductor industry can see that the claim "K3 caused the semiconductor sector to fall" is nonsense.
What is the real trigger for this semiconductor adjustment? There are at least these two significant influences:
First, Meta, this "troublemaker," flipped the table. In early July, after falling behind in the AI wave, Meta came out to flip the table. It first announced the leasing of idle AI computing power. The market's first reaction was "too much computing power was bought." Meta has always been a big buyer of computing power; now it is taking GPUs out to lease. Does this mean computing power is oversupplied? Shortly after, Zuckerberg admitted in an internal employee meeting that the development of AI agents in the past four months "had not accelerated as we expected," and the company's bet on the new organizational structure "has yet to bear fruit."
So the market worried that the validation route for AI commercial applications might not work, AI infrastructure investment might be overheated, and big buyers like Meta might reduce capital expenditures. Infrastructure-related stocks (including chips, storage, data centers) all fell accordingly.
Second, Buffett characterized the current AI market investment as overheated, even casino-like. On July 16, in an exclusive interview with CNBC, Buffett bluntly stated that the current market is dominated by short-term trading, while long-term value investment is marginalized:
"When greed and gambling emotions overshadow rationality, bubbles will inevitably be born. I never predict short-term market fluctuations; I only seek enterprises whose intrinsic value can be reasonably estimated. If a company relies solely on AI concept hype without stable and continuous profitability, no matter how fiercely the stock price rises, I will not touch it. Casinos will always breed flashy bubbles, but patient value investors will completely ignore such noise."
The 95-year-old Mr. Buffett saying AI market investment is gambling is a fundamental challenge to valuation logic, with an impact much greater than that of a single model.
Finally, the crash of Zhipu and MiniMax fundamentally does not require any external factors. This has always been my view: market capitalization propped up by "market probability" will sooner or later reveal its true form.
(Image source: Generated by ChatGPT by Luo Chao Pro)
Even if Kimi does not release K3, or if K3 performs poorly, these two companies would still hardly avoid falling. On July 8, Zhipu welcomed the unlocking of its first batch of restricted shares, involving about 25.68 million shares. On the unlocking day, the stock price rose instead of falling, closing up 13.35%. But immediately after, Zhipu placed nearly 20 million new H-shares at 1588 HKD, raising 31.4 billion HKD—institutions participating in the placement saw a floating loss of 30% seven trading days later. Full-year 2025 revenue was 724 million yuan, with an adjusted net loss of 3.182 billion yuan, and the price-to-sales ratio once exceeded 1000 times.
MiniMax is even worse. In March, the stock price reached a high of 1330 HKD, with market capitalization breaking 410 billion HKD, exceeding tech giants like Baidu that have been deeply engaged in AI for many years. By the close on July 17, it was 216 HKD, with market capitalization at 75.4 billion HKD, shrinking by over 80%. All this happened before the Kimi K3 release, including a single-day crash of nearly 18% on the first day of unlocking. So who should take the blame for this 80% drop before?
The AI market is switching from "scarcity premium" to "fundamental pricing." The story is over, and now the accounts are being settled, that's all.
So if you still don't understand why some AI companies crashed, read Buffett's recent exclusive interview.
That said, Kimi K3 is a good model, giving China another trump card in the large model ecosystem niche. Open-source players like Kimi, DeepSeek, and Qwen are chasing each other, giving Chinese AI hope to run a different path via open source, which is more worthy of attention than the crashes of Zhipu and MiniMax.
Buffett's words fit here quite well: "The market will have two extreme stages: one where high-quality opportunities come flooding in, too many to handle; the other where it takes two or three years, and finding a reliable target is already a huge relief."
Most of the time, the market is in the latter state. Fanatical humans will eventually return to "common sense."
#AI #MiniMax #Kimi #LargeModel #Zhipu
The WAIC 2026, themed "Intelligent Partners · Co-creating the Future," is now in full swing!
The AI narrative route shifts from stacking model parameters to Agent productivity landing; heterogeneous collaboration and photonic computing continue to raise the computing ceiling; embodied intelligence accelerates application, with robots entering homes and factories making physical AI a reality.
The Lei Tech WAIC exhibition tour group has arrived in Shanghai, directly witnessing the annual peak moment of AI industrialization landing. Stay tuned!
End
Scan the code to join for free
the "Together AI" community, and embrace the AI era together.


