DeepSeek's Inference Chip Is China's Quietest Threat to Nvidia

Bottom line

DeepSeek's chip project matters because it attacks the point in AI where scale matters most: inference.

Reuters reported that China's DeepSeek is developing its own AI chip for inference. That is strategically important because inference is the workhorse phase of AI: it is where a trained model serves users over and over again. Whoever controls inference economics controls a big piece of the margin pool.

This is not just about bypassing U.S. chips. It is about building a domestic inference economy.

For Nvidia, the issue is not that one Chinese start-up suddenly replaces the entire ecosystem. The issue is that a direct competitor is trying to localize the most recurring part of the demand curve.

The numbers

The market already knows that China is a meaningful revenue lane for Nvidia.

Why the DeepSeek news moved Nvidia

The market is reacting to the combination of a China-specific threat and Nvidia's still-meaningful China exposure.

Unidad: mixed scale

Nvidia share move

Lowest since April

-1.9

China revenue share

Fiscal 2025 share of revenue

13.1

H20 license revenue

Approximate licensed revenue in fiscal Q4 2025, USD bn

0.1

The threat is about more than unit sales.
Data point	Interpretation
China = 13.1% of revenue	China still matters for Nvidia, even if it is no longer the entire story.
H20 license revenue = $60M	A reminder that export restrictions already constrain the China lane.
DeepSeek focus = inference	The company is trying to own the high-volume deployment phase.
Nvidia stock = -1.9%	Investors are pricing long-run pricing pressure, not just near-term orders.

Why inference matters

Inference is the part of AI where scale and cost discipline decide who wins.

Training is visible and lumpy; inference is continuous and sticky.
If DeepSeek optimizes inference locally, it reduces the need to buy every deployment chip from Nvidia.
That does not kill global demand, but it can shift a high-margin lane into a lower-margin domestic lane.

The broader pattern is what matters. OpenAI is building its own hardware partnerships. Meta is doing the same. Google and Amazon are already pushing custom silicon externally. DeepSeek just shows that China is now fully on the same road.

Supply chain

The winners are not only the chip designers. The losers are the vendors with the weakest pricing power.

Who is exposed if China builds more of its own AI stack.
Company / layer	Why it matters
Nvidia	China pricing and mix become more vulnerable.
Huawei	Domestic competition becomes more credible.
TSMC	Still part of the fabrication bottleneck, but export controls matter.
ASML	Leading-edge lithography remains indispensable to the frontier.

The best read on DeepSeek is not “Nvidia is finished.” It is “China wants the full stack, and inference is where that stack scales.”
AI competition view

Conclusion

The threat is real, but it is a long-run pricing-power threat, not an immediate collapse story.

I would frame the story this way: DeepSeek does not need to beat Nvidia on every benchmark to matter. It just needs to prove that a large market can absorb less Nvidia silicon over time. That is enough to make investors re-think the durability of the China revenue lane.

China inference localization can pressure margins even if volumes stay strong.
The overall AI market still expands, so the story is mix and pricing, not extinction.
Watch export controls, local chip partnerships, and inference efficiency gains.