US official says Nvidia has begun shipping powerful H200 AI chips to China - Reuters

Skyler Monroe (0:10)

Hey, everyone, welcome back to the AI Hardware Show. I'm your host Skyler Monroe, computer engineer turned silicon obsessive, and this is the show where we dig into the chips data centers and hardware that's actually making AI happen. Before we dive in, a huge thank you to our sponsors, Ada, who helps businesses integrate AI into their workflows in smart, practical ways. Ago Consulting pronounced Ago, your go-to for silicon development consulting across everything from AI to SoC design. And Zen Semiconductor, an umbrella company with a serious portfolio of silicon industry ventures, including Sierra RISC-V CPUs, their Loom AI fabric, and Ada for deploying AI at scale. These folks are literally building the processors and AI infrastructure behind modern data centers. Alright, we've got a loaded episode today. We're talking H200 chips heading to China, the unsung hero stock that Nvidia and AMD both depend on, Nvidia's new Blackwell server GPU, an eye-popping market forecast for data center AI chips, and Alibaba swinging big with a chip that could challenge Nvidia's dominance. Let's get into it. Okay, first story, and this one is going to raise some eyebrows. A US official has confirmed that Nvidia has begun shipping its H200 AI chips to China.

Now, if you've been following along with the ongoing saga of US chip export controls, you know this is a pretty significant development. So let's unpack what's actually going on here.

So first, a quick refresher. The H200 is Nvidia's current generation high bandwidth memory flagship. It's the successor to the H100, which was already a monster for training large AI models. The H200 essentially takes the same hopper architecture, but slaps on HBM3E memory, giving it dramatically higher memory bandwidth. We're talking around 4.8 terabytes per second. That's not a typo.

For AI workloads, memory bandwidth is off in the bottleneck, so this is a big deal. Now here's where the policy angle gets interesting.

The US has been tightening export controls on advanced AI chips going to China since 2022

The H100 was restricted. Then Nvidia created a watered down version, the H800, specifically to comply with those rules. Then that got restricted too. So you'd think the H200, which is even more powerful, would absolutely be on the restricted list, right? Well, apparently not. And this is where it gets nuanced. The export control rules are based on specific performance thresholds, things like interconnect bandwidth and compute density.

It's possible the H200's particular configuration, depending on how it's being sold, might technically thread the needle on those thresholds.

Or we could be looking at a policy gap that regulators are now scrambling to address. Either way, the fact that a US official is publicly confirming these shipments suggests this is very much on the government's radar. And you have to imagine there are conversations happening right now about whether the rules need to be tightened again. This has been a bit of a cat and mouse game. Nvidia designs chips to comply with export rules. The rules get stricter. Nvidia adapts and so on. From a business perspective, China is an absolutely massive market for Nvidia. Before the export controls kicked in, China accounted for something like 20 to 25% of Nvidia's data center revenue. Losing that market has been painful, which is why Nvidia has been so motivated to find compliant configurations to sell there.

And from a geopolitical standpoint, this matters because AI compute is increasingly seen as a strategic resource, like rare earth minerals or oil. Whoever has the most advanced chips has a significant advantage in training frontier AI models. So every H200 that lands in China is, in some people's view, a chip that could be used to close the gap on USAI capabilities. Now, I'm not here to make a policy argument. Either way, that's above my pay grade. But what I will say is that this story is a reminder that the semiconductor industry doesn't exist in a vacuum. These chips are geopolitical assets, and the decisions about who gets them are just as consequential as the engineering decisions about how to build them. This one's going to keep developing, so stay tuned. All right, moving on to story two, and this is one of those wait-who moments that I love. The headline from Motley Fool is that Nvidia and AMD can't make AI chips without a particular growth stock, and they're talking about a company that doesn't get nearly the attention it deserves given how central it is to the entire semiconductor ecosystem. The company they're highlighting is ASML. And if you haven't heard that name, let me fix that right now because ASML might be the single most important company in the chip industry that most people have never heard of. ASML is a Dutch company headquartered in Eindhoven, and they make the machines that make the chips. Here's the analogy I love for this. Imagine you want to build the most intricate, detailed painting in the world. We're talking brush strokes measured in atoms. You can't do that with a regular paintbrush. You need a specific, incredibly specialized tool. ASML makes that tool for the chip industry. Their machines use a process called Extreme Ultraviolet Lithography, EUV, to etch circuit patterns onto silicon wafers with mind-boggling precision. And here's the thing, ASML is essentially the only company in the world that can make these EUV machines. It took decades of development, billions in R&D, and deep collaboration with hundreds of suppliers. The machines themselves cost around $200 million each, and the newest generation, high-end EUV, cost around $350 million per machine. They're also about the size of a double-decker bus, just for context. So when Nvidia wants to manufacture its H100s or H200s or Blackwell chips at TSMC, TSMC needs ASML's machines to actually do the manufacturing. Same for AMD, same for Apple, Intel, Qualcomm, everyone. Without ASML, the cutting-edge chip fabs simply cannot operate. It's one of the most powerful monopolies in the history of technology, built on genuine engineering excellence. From an investment standpoint, The Motley Fool's argument is essentially that, as demand for AI chips explodes, demand for ASML's machines explodes too, because fabs need to buy more machines to produce more chips. And since ASML is the only game in town for leading-edge lithography, they're in an extraordinarily strong position. There's also an export control angle here too, which is fascinating. The US has pressured the Dutch government to restrict ASML from selling its most advanced EUV machines to China, and the Netherlands has complied, meaning Chinese chipmakers like SMIC are locked out of the most advanced lithography. So ASML is right at the intersection of AI, geopolitics and semiconductor manufacturing. Pretty remarkable position for a company that most people outside the industry have never heard of. The broader lesson here is that the AI hardware story is not just about Nvidia and AMD and the flashy GPU numbers, there's an entire supply chain that makes those chips possible. Understanding the full stack from lithography machines to packaging to memory gives you a much more complete picture of where the real strategic value lies. Okay, story three, and this one comes straight from the Nvidia developer blog. It's about scaling AI ready data centers using the new Rtx Pro 4500 Blackwell Server Edition, paired with Nvidia's VGPU 20 software platform. This is a little more technical and a little more enterprise focused, but stick with me because there's some really interesting stuff here. So first, let's place this chip in the Nvidia lineup. When people talk about Nvidia's Blackwell architecture, they usually jump straight to the B100 or B200.