今天全球暴跌，煽动蝴蝶翅膀的，是美股一个消息 meta出售算力 本来这是一个很正常的行为，meta是全球头部厂商，拥有大量算力 外面有人高价买算力，自己没有充分用完，拿出来赚现金并不是坏事 马斯克自己有ai，也在出售算力，这是很常见的商业行为 但媒体解读新闻的时候，加了一句标题 “Meta Is Buildinga Cloud Business to Sell Excess AI Compute” 就这个Excess，意思从出售算力，解读成出售过剩算力 一字之差，天差地别。 更何况pb社是全球机构主要信息来源。 发现出了惊天乌龙后，pb马上修改了原文 改成 "Planning a Cloud Business Sell AI Computing Power" 出售都改成了计划出售，过剩删除了。 就这几个字差异，有巨大的影响力，改了之后，但原文到处传播 全球ai产业看头部，全球金融机构看pb，这下好了 pb说头部算力过剩 今天全球暴跌，韩国，日本，a股都暴跌 即使大家知道这是标题党，但又如何了 ai的繁荣本来就很脆弱，最重要的还是大家都赚肿了，回调只需要一个谣言。 本来就处于高波动环境，一个全球头部媒体的标题党，足够带来巨震了 光，存储，设备，都是重灾区，但存储也只是回踩ma5或者ma10而已。 从买方角度看，这是一个无法抵抗的波动 但这也是拥抱科技的代价之一 你知道ai很好，持仓过程会很艰辛，会遇到各种波折，但你还是要忍受这种波动

新版的 GLM 5.2 基本稳稳站住全球开源模型第一的位置了，作为开源模型可以排到全球前三。上一个这么牛逼的还是 DeepSeek，给 AI 圈带去一点震撼。Claude 4.6 之后提升就没那么大了，4.6 也是 AI 编程的可用线。前端时间巨头们开始对 AI 投入算总账，几十亿美金的 AI 账单，现在用开源的GLM 5.2自己部署，就能立即省下高额的 AI 账单，Claude 的收入端可能要大受影响了。 中美顶尖 AI 的差距大概也就是半年左右，另一方面也说明大模型的护城河没那么深，半年时间的差距并非不可逾越，Anthropic 的万亿估值显然是挺虚的。Claude 年初最辉煌的时候，总有人会说 Claude 多牛逼，迭代速度多强，我去看了一下 zcode 的迭代速度基本是一天一更，这次zcode3.0+GLM5.2 真的和 codex 使用起来没什么太大差距了。 还有中国为什么一直追不上，主要是因为算力资源的限制，假如算力限制放开怕是已经 4:6 开了。可怜的智谱算力一直都不足，难以想象要是有充足的算力且模型是闭源的情况下，难以想象收入会高到什么程度。 The new version of GLM 5.2 basically stand firmly in the first position of the global open source model, as an open source model can be ranked in the world's top three. The last one was DeepSeek, bringing a little shock to the AI circle. After Claude 4.6, the promotion was not so big, and 4.6 was also an available line for AI programming. The front-end time giants are starting to take a general ledger of AI investment, billions of dollars in AI bills, and now deploying themselves with the open source GLM 5.2 can immediately save the high AI bills, and Claude's revenue side may be greatly affected. The gap between the top AI of China and the United States is about half a year, on the other hand, it also shows that the moat of the big model is not so deep, the gap in half a year is not insurmountable, the Anthropic trillion valuation is obviously quite empty. When Claude was at its most brilliant at the beginning of the year, some people would always say how awesome Claude was and how fast the iteration speed was. I went to see that the iteration speed of zcode was basically one change every day. This time zcode3.0 GLM5.2 was really not much different from codex. There is also a reason why China has been unable to catch up, mainly because of the limitation of computing resources. If the limitation of computing power is released, it may have already been opened at 4:6. Poor intellectual spectrum computing power has been insufficient, it is difficult to imagine if there is sufficient computing power and the model is closed source, it is difficult to imagine how high the income will be.