US gov't allows Chinese telecom giant ZTE to purchase Nvidia H200 AI chips — firm joins Alibaba, Tencent, and ByteDance in access to Hopper tech

**Skyler Monroe** (0:10)
Hey, everyone. Welcome back to the AI Hardware Show. I'm your host, Skyler Monroe, and this is the show where we dig into the silicon, the systems and the supply chains powering the AI revolution. Quick shout out to our sponsors, LimitLess AI, helping businesses actually integrate AI into their workflows in a meaningful way, and a Go Consulting, that's our Go, your go-to for silicon development consulting.
All right, we've got a packed episode today. Nvidia chips heading to China, TSMC crushing revenue numbers, and a major crackdown on AI chip smuggling. Let's get into it.
Okay, first story, and this one is genuinely wild when you think about the context. The US government has approved a license allowing ZTE, the Chinese telecom giant, to purchase Nvidia H200 AI chips.
Yes, that ZTE, the same company that's been on and off various US trade restriction lists for years. And now, they're joining a pretty notable club, Alibaba, Tencent, ByteDance, all of whom have also been cleared to access Hopper generation technology. So let's break down what's actually happening here. The H200 is Nvidia's current flagship data center GPU built on the Hopper architecture.
And when I say flagship, I mean it, this chip is a beast. It pairs the same GH100 compute die as the H100, but swaps out the memory for HBM3E, which dramatically increases memory bandwidth. We're talking around 4.8 terabytes per second of memory bandwidth. For AI inference and large-language model training, memory bandwidth is often the limiting factor, so that upgrade is not trivial. Now, the Hopper generation was originally restricted for export to China in late 2022 That's when the US government started getting serious about preventing advanced AI compute from flowing to Chinese entities. Nvidia had to pivot and create export-compliant chips, the H800 and eventually the H20, which were significantly cut-down versions. The H20 in particular was interesting, because it kept some of the memory bandwidth but throttled the compute interconnect. So it was still useful for inference workloads but not ideal for large-scale training clusters. So the fact that full H200s are now being licensed to Chinese companies, including ZTE, specifically, is a meaningful policy shift. And it comes in the context of a broader thaw between the US and China on trade. Following the Trump XI meeting, we'll talk about more in a later story. But here's the thing. And this is the nuance that matters. Even with the license approved, there are real questions about how much of this actually translates into chips in Chinese data centers. Chinese domestic procurement initiatives are pushing companies to favor homegrown silicon wherever possible. There's real political pressure inside China to not be seen buying American chips. And frankly, the domestic chip ecosystem, Huawei's Ascend line, Cambricon, more threads has been advancing. Not at the same level as Nvidia, but advancing. So the market impact of this ZTE license might be more symbolic than transformative. Still, it signals something. The conversation around export controls is shifting, and that matters for every company in this supply chain. All right, let's pivot to story two and three together, because they're really two sides of the same coin. Both involve TSMC, which is basically the backbone of the entire AI chip industry. First, TSMC's Q2 earnings came in with revenue up 36% year over year.
That is a stunning number for a company of TSMC's size. We're talking about the largest contract chipmaker on the planet posting growth rates that most startups would envy. So why is TSMC growing so fast? Three letters, AI. The demand for advanced silicon for AI training and inference workloads has been relentless. Nvidia's H100s, H200s, the Blackwell B200s all fabbed at TSMC, Google's TPS, TSMC, Apple's M-series and A-series chips, TSMC, Amazon's Tranium and Inferentia chips, TSMC, AMD's MI300 series, TSMC. When you're the world's most advanced semiconductor foundry and AI is the hottest technology sector in human history, you're going to have a very good quarter. And let's give credit where it's due, TSMC's process technology leadership is not an accident. Their N3 node which is their 3 nanometer class process and now N2 on the horizon. These represent years of incredibly difficult engineering. We're talking about transistors so small that we're measuring them in atoms. The gate all around transistor architecture in N2 is a fundamental change to how the transistor is physically constructed. This isn't just incremental improvement, it's a rethinking of the basic building block of every chip. The reason TSMC dominates AI chipmaking specifically is that AI accelerators need the highest transistor density possible to pack more compute into a given die area. More transistors means more CUDA cores or tensor cores or systolic array elements, whatever the accelerator uses and that directly translates to more teraflops of AI compute. So TSMC's process leadership is directly correlated to AI performance leadership. They're not just a manufacturer, they're a strategic asset. The headline question around TSMC earnings was whether they could rescue the AI chip trade, which is a bit of dramatic framing, but the underlying concern is real. There's been volatility in AI chip stocks, questions about export control impacts, and general macro uncertainty. TSMC's strong numbers essentially serve as a real-world data point that AI infrastructure spending is still very much on. Hyperscalers are still spending, cloud providers are still spending, cloud providers are still building. The demand signal from TSMC's fab utilization is about as real as it gets.