Zyon Wang (@ZyonWang)

花 1450 亿美元建算力的小扎,今天说了句大实话:算力用不完,准备转卖。 这句话有点要命。 整个 AI 产业链现在的狂欢,就靠两根柱子撑着: 一是模型真能赚钱,OpenAI、Anthropic 的 ARR 一直创新高,证明生意跑得通; 二是大厂玩命砸钱,Google、Oracle、Meta 的 CAPEX 一路加码,半导体、PCB、光模块这些,赚的就是这份 CAPEX 的钱。 但这两根柱子底下,还压着一个共同的信仰——算力永远不够用。 现在砸钱最猛的那个,自己站出来说"我用不完、要转卖",这个信仰就开始晃了。 再叠加 Meta 广告和社媒本身的压力(消费降温、人跟 AI 聊得比跟人还热闹),万一哪天大厂真开始收手,CAPEX 一降温,上游那条链的估值,还撑得住吗? 你站哪边: A 第一道裂缝,链子要松了 B 想多了,能转卖=更会赚钱 评论区聊聊👇

Heard from someone in the space that Kling v4 ships in July — and Seedance 2.1 is reportedly landing in July too, so the two go head-to-head this round. v4 is basically aimed at Seedance, so going off the gaps between them, here's where I'd bet v4 puts its work: - Multimodal references: SD2.0's killer feature — one generation pulls in images, video and audio as references at once (up to 9 images). Kling will push here and build out multi-source referencing. - Physics: SD2.0 is genuinely ahead — flips, collisions, multi-body interactions top the benchmarks. v4 will improve on physical realism; water, hair and cloth should hold up better. - Speed: SD is fast and built for volume. Kling leans cinematic and slower, so v4 will optimize generation speed — probably a Turbo tier for volume. - Price: v4 likely costs more than v3 — new models almost never get cheaper. Even so it should still undercut Seedance, which is genuinely expensive. Two bolder bets on what it'll add: lip-synced audio, and single shots stretched to minute length. And with SD2.1 also due in July, the bar only goes up. This is the matchup to watch — I'll come back and check how wrong I was.