‘Get me out’: Traders dump software stocks as AI fears erupt
有时候处在大的技术浪潮下,对认知的要求陡然上升,不同时期不同时代背景下,影响事物发展的主因是不同的,也越发让我理解,真的没有什么是一成不变的,假如有,可能是时间周期不够经历的不够,比如放到30年60年周期
过往二十年“轻资产高毛利高增速”的 SaaS 是所有投资者趋之若鹜的领域,谁曾想去年开始 AI 加速深入发展之后,一边是 Claude Google OpenAI 每次震惊四座的产品发布,另一边却是将 SaaS 软件推向墓冢。
泥沙俱下不可怕,可怕的是结构分化。说明一定有你没有看到没有搞清楚的地方,在悄然发生变化
$INTC "Just to push on that a little bit, you guys have your own factories. Why are you in the inventory situation that you’re in? And even if you look at it, I get the whole idea of finished goods versus other components, but you have $11.6 billion of inventory and it’s still not in the right place at the right time. How does that happen?"
英特尔这个管理层还真是一如既往的积重难返,直接被贴脸开大😂,所以懂得周期资本配置的 CEO 真是难求,还是多开除些人继续改造吧
整体看下来还是业绩预期太高,不过更关心的 CPU 需求旺盛得到确认,供给端又不是一家没法放量,大家都无法短期扩产。
「看之前一个华泰的报告:“海外各家大厂进展来看,Meta收购manus,Anthropic发布Cowork,OpenAI的ChatGPT Agent,Google的Gemini CLI、Chrome Auto Browse;国内来看,阿里千问Agent、豆包手机助手,由此可见大厂主流Agent均已基本打磨完成,PC端、手机端个人生活工作助理向的Agent成为主流产品形态。26年C端Agent渗透节奏或超预期。
算力Claude Code工作流来看,启动阶段(只打开一个项目,几个小文件)智能体思考一次消耗约1-2w token;深度工作阶段,用户打开20个以上文件,Agent积累大量尝试记录(编码、debug、日志、比较代码差异),每一步消耗20w token。Claude Code官方建议用户每分钟的Token限制设置在20-30w token,足以见得算力需求的巨大(对比Chatbot,一个用1min交互2次左右,单次消耗2k token左右)。总体量化来看,完成一个小项目消耗百万级别token、与Chatbot单次相比提升3个数量级。从过去一周Openrouter数据来看,Claude 4.5调用量提升明显,下一次token加速点很快出现。
Agent带来的两大结构性影响:存储与CPU。CPU角度,从Agent的工作执行原理来看CPU重要性提升:
1)服务器拿到众多推理任务,第一步打batch(批处理),此任务在CPU完成;
2)数据进⼊GPU之前,需要进⾏解压缩、解密和向量化,对于不需要并行计算能⼒的⻓尾推理任务,CPU将成为⾸选;
3)虚拟机技术路线下,Agent需要CPU创建大量沙箱环境、奖励与验证器,需要CPU构建各种工具环境;
4)Agent任务步骤的分配与多线程工具调用在CPU执行,上下文存储也优先外溢到CPU。」
算力-运力-存力-CPU-电力,真是轮番上阵,既然CPU 需求得到确认,其实反而感觉更有利于 INTC 这种晶圆厂自己调整产线,有点像之前美光:田忌赛马,砍掉下等马,把粮草供给到上等马,从而放量到高毛利产线
继续持有了
Brad Gerstner @altcap discussing why CoreWeave $CRWV is the only neocloud position he likes.. Interesting to hear from someone who knows Jensen, Altman, Satya etc. very well..
"We were investors in CoreWeave when it was a private company. We've been investors, you know, as they entered the public market and it's pulled back here, we've added to our position. We're not adding to other neoclouds. We don't think other neoclouds are created equal. We think CoreWeave stands alone in terms of its performance, the software that it's build on top of the cloud, its execution capability, its strategic importance to Nvidia in terms of deploying Rubin, etcetera. So we think it makes a really interesting opportunity."
Hyperliquid supports permissionless perps on anything. As all of finance moves onchain, there is a billion dollar opportunity to build a mobile app for non-crypto users.
The two keys are:
1. seamless fiat onboarding
2. a mobile UX that non-crypto users love
Hyperliquid and HIP-3 perps offer the full backend liquidity infrastructure. Solve 1 and 2, and you have a product that markets itself: global, permissionless finance at your fingertips. Builder codes allow monetization proportional to volume flowing through the app, fully configurable per-trade.
Hyperliquid’s ethos is to let talented, hungry teams reinvent the pillars of finance. For those who specialize in UX and building magical user experiences, nothing stands between you and the empires to be built.