AC.Huang

@adcock_brett @Figure_robot @hark_labs 太厉害了 羡慕❤️

Figure and Hark just took an entire data center of NVIDIA B200s - every rack in the building Figure will be using these to predict physics and Hark will train next generation multi-modal models

@jukan05 来深圳 到你体验地道美食

深圳哪里有好吃的餐厅可以推荐吗？本地人/中国人常去的？ Any Chinese locals here — could you recommend some good restaurants in Shenzhen?

@wyek1n @WOWshiyingqi 13000一条

@WOWshiyingqi ddr5 64G*4 什么价？

在香港注册了一个公司，做了三星和海力士的代理，专卖内存。 在港交易，欢迎询价

@BgVdbJsZzTJUMVQ @WOWshiyingqi 深圳D5 崩盘了

@WOWshiyingqi ddr5 5600有货吗，现价多少

@LeonKuo2023 @WOWshiyingqi 南亚D3 颗粒 你需要哪个容量？

@WOWshiyingqi 有DDR3 颗粒吗？ 有渠道需求2KK

@LLMJunky @TheAhmadOsman @0xSero @ivanfioravanti @ZenMagnets @NVIDIAAI I need a large quantity of PRO600, could you help me purchase them?

Finally proud to announce that I've joined the GPU Minor Leagues. 2 x RTX 6000 Pro. I have six months to pay off the second GPU lol. You are all TERRIBLE influences.

@realPureNomad 这个联带责任有点大

著名服务器公司 Supermicro 联创 Yih-Shyan 今日被捕。事由是通过东南亚壳公司向中国走私 25 亿美元的 Nvidia GPU。他本人持 4.64 亿美元的 Supermicro 股票。走私手段包括，用数千个假服务器迷惑美国执法者，监控录像显示使用吹风机换条形码。Supermicro 股票盘后暴跌 12%。此人将面临 30 年刑期。

@AsiaFinance 太牛了 我们只能卖国外

两年前，车库咖啡创始人苏菂，公开炫耀如何逃避美国的AI技术制裁，走私了200片英伟达的H200显卡；两年后，苏菂的供应商被抓了，美国司法部今天公布案情：联邦调查局(FBI)调查显示，廖、张、孙三人利用东南亚公司，合谋向中国买家出售数十亿美元英伟达GPU，最长面临20年刑期。

@CuiMao 你们这个怎么卖 求合作

B300 开始涨价

@MSX_CN 高价收购服务器内存条 海外大量采购

存储芯片的领头羊 $MU.M 今夜盘后交业绩啦！ 在 HBM（高带宽内存）供不应求的背景下，全市场都在看美光的盈利上限到底在哪。 📈 数据预期： 1️⃣销售量：192.99 亿 2️⃣净利润：98.78 亿 大家现在最关心的是：美光手里被 $NVDA.M 抢疯了的 HBM 产能，是不是已经排单到了后年？如果毛利率能借着涨价潮冲破 65%，那美光就真的成了芯片界印钞机。 如果美光指引超预期，整个芯片板块都要跟着起飞。等明早看是「真香」还是「利好出尽」？

If you have RTX 5090 chips or H100/H200 servers for sale, please contact me.

@DylanOnChips If you have RTX 5090 chips or H100/H200 servers for sale, please contact me.

I got Groq'd

@jukan05 Acquire data centers and decommission H100 and H200 servers

According to a calculation by a Korean journalist estimating how much revenue Samsung Foundry could generate from Groq3 LPU production: NVIDIA has reportedly asked Samsung Foundry to produce around 500,000 LPU 3 chips to start with. That is more than double the originally planned production volume. First, each Groq server introduced by CEO Jensen Huang contains 8 LPUs. When 32 of those servers are combined, they form one rack. That means each rack contains 256 LPUs. One complete Vera Rubin platform contains 5 racks. So each platform would contain 1,280 LPUs. On a simple calculation, if all produced chips were turned into racks, that would imply about 1,950 LPU racks, enough for roughly 390 Vera Rubin platform sets. Looking at the LPU wafer shown in the commemorative photo taken a few days ago at the GTC exhibition featuring CEO Jensen Huang and Han Jin-man, president of Samsung Electronics’ Foundry Business, it appears that around 65 properly shaped LPU dies can be printed on a single wafer. To produce 500,000 LPUs, 7,692 wafers would be required if each wafer yields 65 chips and the yield rate is 100%. However, considering that Samsung’s current 4nm yield is estimated to be around 50–70%, it is possible to infer that more than 15,000 wafers annually would be needed, which is broadly consistent with recent media reports. According to what the journalist heard, the wafer price for Samsung’s 4nm process is at least around $11,000 per wafer. Assuming a possible price increase and using roughly $13,000 per wafer, that would imply about $195 million in revenue, or roughly KRW 300 billion, from LPU production. Given that Samsung Foundry’s annual revenue is known to be around KRW 20 trillion, this would still amount to a very small contribution to total sales.

@LinQingV 收购数据中心并退役 H100 和 H200 服务器

Nvidia要给中国市场供高性能推理芯片了。 路透独家：Nvidia正准备推出可在中国销售的Groq推理芯片，预计5月发布。 先说Groq是什么来头。这家公司做的不是GPU，是LPU，专门为推理设计的架构。核心思路是把HBM干掉，全用SRAM，执行路径完全确定性，不存在GPU那种调度开销。结果就是推理延迟极低，吞吐拉满，单位能耗也远低于GPU方案。Nvidia花170亿拿它的技术授权，不是收购，要的就是这套推理专用架构。 注意，这次不是阉割版。消息人士明确说"非降规、非中国特供"，跟之前H800/A800完全不是一个路径。训练靠Vera Rubin，不能卖中国。推理走Groq，可以卖。一套组合拳，两条合规通道。同时H200也重启生产了，说明出口许可和中国订单都已经到手。 如果NVDA用推理芯片重新把中国市场的门撬开，影响就很直接。Nvidia中国区收入预期要重估，华尔街此前基本把这块从模型里划掉了。反过来，百度、寒武纪们自研推理芯片的窗口期，可能比市场以为的短得多。Nvidia带着成熟生态杀回来，纯靠性价比很难扛。 如果中国这边不阻挠，主攻推理的国内GPU厂商会很难受。

@MichaelDell @jimcramer Acquire data centers and decommission H100 and H200 servers

Always fun with Jensen Huang and @jimcramer at NVIDIA GTC!

If there is an idle NVIDIA H100 H200 server that needs to be disposed of, you can contact me.

@nvidia If there is an idle NVIDIA H100 H200 server that needs to be disposed of, you can contact me.

In 2006, NVIDIA made a bold bet on a parallel computing platform that would eventually redefine the limits of science and industry. 20 years later, CUDA serves as the foundation for 6 million developers worldwide—and we are only just beginning. Join the next wave of architects and build what’s next on two decades of innovation. 🔗 nvda.ws/4bUUFUT

@Tslachan If there is an idle NVIDIA H100 H200 server that needs to be disposed of, you can contact me.

Samsung Chairman Lee Jae-yong met with AMD CEO Lisa Su for dinner in Seoul last evening. Lisa Su visited Korea for the first time since taking office at AMD, and the two are known as places to strengthen cooperation in AI memory. Samsung has been going wild lately.

@Dell @nvidia If you have B300 for sale, you can contact me.

We heard you like seafood 🦞. Meet the workstation of the AI era: The Dell Pro Max with GB300 and @NVIDIA NemoClaw. You can run AI agents locally, securely & nonstop.