META 第一阶段 AI 超级集群由数十万块 GPU 组成，满规模目标约为 100 万块，展现出强劲的 AI 执行力；其工程人才与 GPU 可靠性方面的工作，可能使其能以溢价出售算力，且财报应会开始体现 Muse 与这套基础设施正在赚钱。
逐段跳播
Meta 和 Nvidia 刚刚联手打造了目前最疯狂的 AI 超级集群之一。
Meta and Nvidia just teamed up to build one of the craziest AI supercluster out right now.
今天早些时候，Meta Platforms 确实公布了他们的超级 AI 集群，并且已经告诉我们第一阶段已经完成，这一阶段目前包含数十万块 GPU。
earlier today, Meta Platforms did announce their super AI cluster and they already told us that it finished the first phase, which includes hundreds of thousands of GPUs today.
而且众所周知，Meta 表现非常出色。Meta 在 AI 市场上表现非常出色。我不知道他们是不是第一名，但他们在 Muse 上绝对做得非常棒。
And as we know, Meta is killing it. Meta is killing it with the AI market. I don't know if they're number one, but they're definitely doing amazing with Muse.
未来的目标，虽然他们没有给出时间表，预计将达到大约 100 万块分布式 GPU 全部同步运行。那是满规模状态。
What the future target while they didn't give a timeline is expected to reach roughly 1 million distributed GPUs all functioning in lock step. That is at full scale.
Meta 能够分享一些图片。这是 GB200 对比 GB300。所以，这两者都已经部署了。但我觉得相当有趣的是，GP200 是跨两个机架系统。GP300 则是单机架解决方案的形式。
Metam was able to share some images. This is the GB200 versus the GB300. So, both of these are already deployed. But what I found pretty interesting was the GP200 was across two rack system. The GP300 was on a one rack solution style.
现在，Meta 确实谈到了他们的后端聚合架构图，这就是他们展示的图片。它显示了 GPU 结构组通过聚合节点连接，基本上只是解释了 GPU 数量。
Now, Meta did talk about their backend aggregation schematic and and this is the image they show. It shows the GPU fabric groups connecting through aggregation nodes and it pretty much just kind of explains the GPU count.
这些就是 Meta 正在使用、Nvidia 正在使用的网络解决方案类型，我很高兴看到现在有一些已经投入实际运行。
and these are the types of networking solutions that Meta is working with that Nvidia is working with and I'm glad to see some in action right now
另一件我觉得相当酷的事情是，Meta 拥有真正的工程人才。
The other thing that I thought was pretty cool is meta is a true engineering talent.
无论如何，Meta 正在努力提高 GPU 的可靠性。如果 Meta 将来想出售算力，他们确实能以溢价价格卖出。
Regardless, Meta is trying to improve the reliability of GPUs. if Meta ever wants to sell compute, they can truly get it at a premium rate.
在这个 72 GPU 域中，他们有 64 块活跃 GPU，还有八块作为备用。Meta 报告称，Muse 1/3 预训练的 goodput 高于 85%，强化学习运行高于 90%。
They have 64 active GPUs in this 72 GPU domain and they have eight kind of in in spares. Meta reports goodp put above 85% for Muse 1/3 pre-training and above 90% for its reinforcement learning runs.
Meta 确实将功耗封顶、电池和液冷描述为解决方案的一部分。
Meta did describe power capping, batteries, and liquid cooling as part of the solution.
我在这里最大的收获是，你看到 Meta 正试图为 AI 市场建设一个百万级 GPU 以上的基础设施数据中心，这让我们对 AI 解决方案的需求状况有了很多、很多的清晰认识。
My biggest win here is though is you're seeing Meta attempting to build a million plus GPU infrastructure data center for the AI market and that gives a lot and a lot of clarity about what the demand looks like for AI solutions.
我对 Meta 的财报超级兴奋，因为我们最终会开始听到 Muse 为他们赚钱，听到这个 AI 基础设施为他们赚取大量、大量的钱。
I am super excited for Meta's earnings because we're going to start to hear eventually of Muse making money for them, of this AI infrastructure making tons and tons of money for them.
观察点
这个频道对 $META 的历次判断
Jose Najarro Stocks 在这只股票上共 12 条判断，这里只列相邻的几条。