NVDA Vera Rubin 每 TCO 吞吐量较 GP300 高出 67 倍，并在代理工作负载中占据主导地位，是终极之选。

as well as some things affecting MetaPlatforms and Nvidia.

I invest in companies such as Nvidia, AMD, Credo and Astera Labs, as well as modern cloud computing companies such as Coreweave and others.

Now, the last thing I want to talk about is Nvidia. Nvidia's market capitalization is currently $5.3 trillion , which is about $220. Jensen is a real giant, isn't it? Jensen is a beast in the AI world .

What I found interesting was a report published on September 14th, about four days ago, titled " Genie Inference: 67x Better Performance Per Dollar." They stated that Jensen is underperforming again, achieving double the annual profit per gigawatt. The more you buy, the more you earn.

overall, what they're saying is that Jensen is significantly improving performance again. And, importantly, all of this is happening on pre-release software , which is already seeing six times better token productivity.

For every unit of power. According to their estimates, Robin can generate more than double the profit per gigawatt of the Blackwell platform . Everyone will buy Robin, right? It's impressive that Jensen is promising something, but workloads show a completely different scenario where you can make more money.

So, anyone who gets Robin will be able to command better prices because they can acquire more tokens, and that's the source of the additional revenue for each Vera Robin.

They also talk about their valuation of Vera Robin in terms of a proxy workload. And that's the kind of workload I personally enjoy the most , isn't it? I think the market as a whole is now moving towards proxy workloads.

So, even if the chip performs excellently in inference , that's only a small part of the challenge, isn't it? The challenge lies in this multi- stage work, the long context, the intensive reuse of prefixes , and the flow of sub-agents.

That 's proxy workload, and that's the market trend. So, even if you have a great chip—I mean, an OpenAI Jalapeño chip —if you don't run proxy workload inference, that's what shows the whole picture.

I believe a significant portion of the revenue will come from it. Therefore, be careful when choosing the performance metrics you base your decisions on. I think the best metric currently for evaluating any segment is proxy workflow, as it's the true revenue stream for cloud computing providers.

This doesn't mean there is n't a market for pure inference workloads, where multi- stage processes or many tool calls aren't required, but it won't be the primary market. I believe the primary market will be through proxy workloads, and the Vera Rubin significantly outperforms other performance metrics, delivering far better results than any other available solution.

The Vera Rubin is a remarkable improvement over the GP300 processors in terms of performance versus cost. The Vera Rubin offers 67 times higher total throughput per total cost of ownership compared to the GP300 Dynamics processor.

They also have a wide variety of proxy inference workloads. You can see significant throughput improvements in the Vera Rubin from Nvidia. I really wanted to showcase this because Nvidia is the ultimate choice. It truly is.