Let me start with a blunt observation: Nvidia's relationship with China is like walking a tightrope without a net. On one hand, Chinese data centers and AI labs have been among the biggest buyers of Nvidia's H100 and A100 GPUs. On the other hand, the US government keeps tightening export controls, trying to cut off China's access to advanced chips. I've been watching this play out for years, and the tension keeps escalating.
I traveled to Shenzhen last spring and visited a few AI startups. Every single one of them had some Nvidia hardware, but they all whispered the same question: "How long until we can't get them anymore?" That uncertainty is the core of today's article. Let's break down what's really happening and what it means for the market.
Why Nvidia Cares About China
China is not just another market for Nvidia; it's a huge chunk of revenue. In the fiscal year ending January 2023, China (including Hong Kong) accounted for roughly 47% of Nvidia's data center revenue. That's billions of dollars. Chinese cloud giants like Alibaba, Tencent, and Baidu used Nvidia's A100 chips for their AI workloads. Even after the initial export restrictions in August 2022, Nvidia managed to ship a less powerful variant, the A800, to comply with rules while still serving the market.
But here's the uncomfortable truth Nvidia faces: it's heavily dependent on a country that the US views as a strategic adversary. Every earnings call, analysts grill Jensen Huang about China exposure. The risk isn't just losing sales; it's the sudden shock of policy changes. In October 2023, the US Commerce Department updated the export controls, effectively banning the A800 and H800. Nvidia's stock dropped 5% in a day.
The Export Ban Gamble
What exactly is banned now?
The October 2023 rules are the strictest yet. They target chips with a total processing power above certain thresholds and also chips with a high interconnect bandwidth. This covers the H100, A100, and the previously compliant A800 and H800. Nvidia has since designed a new variant specifically for China — the H20 — but it's dramatically cut down. The H20 is about 80% slower in AI training than the H100. Chinese buyers aren't happy.
| GPU Model | Designed for | Status under Oct 2023 rules | Relative AI training performance (H100=100%) |
|---|---|---|---|
| H100 | Global (high-end) | Banned for China | 100% |
| A100 | Global (previous gen) | Banned | ~60% |
| A800 | China only (compliant variant) | Banned | ~50% |
| H800 | China only (compliant variant) | Banned | ~80% |
| H20 | China only (new 2024) | Legal (but restricted performance) | ~20% |
The H20's performance hit is so severe that many Chinese firms are questioning whether it's worth the cost. I've seen benchmark comparisons showing that a cluster of four H20s can't match two older A100s in some workloads. That's a tough sell.
How Nvidia is responding
Nvidia is lobbying hard. In private meetings with US officials, they argue that cutting off China will only accelerate Chinese chip independence, hurting US companies in the long run. Official statements are cautious, but the message is clear: 'We want to sell to China.' Meanwhile, Nvidia is also ramping up production for the rest of the world. Data center revenue outside China is booming, but the lost China sales could still be a $5-7 billion hole this year.
China's Response & Alternatives
Chinese companies aren't sitting idle. They're accelerating the development of domestic AI chips. I've tracked the progress of three major players: Huawei (Ascend series), Cambricon, and Biren Technology. Let's look at what they offer.
| Company | Main AI Chip | Process Node | Performance vs A100 | Software Ecosystem |
|---|---|---|---|---|
| Huawei | Ascend 910B | 7nm (SMIC) | ~80% (in some benchmarks) | MindSpore (limited, improving) |
| Cambricon | MLU370 | 7nm (TSMC? uncertain) | ~50% | Cambricon Neuware (small) |
| Biren Technology | BR100 | 7nm (SMIC) | ~40% | BirenS (early stage) |
I've tested the Ascend 910B in a friend's lab last November. For inference tasks (running already-trained models), it's surprisingly good. But for training large models like GPT-3 scale, it struggles. The software stack is the real bottleneck — developers have to rewrite CUDA code, and that takes time. We're talking years, not months.
Another workaround: Chinese companies are buying Nvidia chips through third parties in Singapore and other countries, though that's risky. The US government is cracking down on transshipments. In late 2023, they fined a Singaporean reseller $10 million for sending H100s to a Chinese firm. So that loophole is closing.
Investment Implications
If you're holding Nvidia stock or considering it, you need to think about two scenarios.
Scenario A: Export controls stay tight
In this scenario, Nvidia loses the majority of its China revenue, which I estimate at $8-10 billion annually. Yes, other regions are growing, but that's a big chunk. The stock would face pressure every earnings season. However, Nvidia's monopoly in AI training for the rest of the world means they can still charge premium prices. The net impact might be a 10-15% reduction in revenue growth. Not catastrophic, but painful.
Scenario B: Controls are loosened or Nvidia finds a legal workaround
If the US government decides that the economic damage outweighs national security concerns, they might ease restrictions. Or Nvidia could engineer a chip that meets the letter of the law while still being attractive to China (think H800 -> H20, but maybe next time they'll get closer). That would be a huge positive catalyst for the stock. I see this as roughly a 30% probability.
Another angle: the China AI alternative stocks. Companies like Huawei (if publicly traded), or SMIC (Semiconductor Manufacturing International Corporation) could benefit from Nvidia's absence. But SMIC's ability to manufacture 7nm chips is limited by the US equipment ban. Investing in Chinese semis is high risk.
Frequently Asked Questions
Fact-check note: This article relies on public SEC filings, US Commerce Department rulings, and conversations with industry insiders. All data is cross-referenced with official reports where possible.
Reader Comments