DeepSeek hype sparks a run on Nvidia H20 chips

When DeepSeek shocked the world last month with its dirt-cheap, high-performance models, Wall Street went into panic mode. The instant narrative was simple: if state-of-the-art AI can be built on a shoestring budget, Nvidia's gold-rush era must be coming to an end. Nvidia stock even tanked 20% in the immediate aftermath.

Well, that thesis just hit a massive reality check. According to a Reuters report citing six people familiar with the matter, Chinese tech giants are furiously placing new orders for Nvidia's H20 chips, directly driven by DeepSeek's skyrocketing popularity.

DeepSeek sparks a buying frenzy

Tencent, Alibaba, and ByteDance have significantly ramped up orders for the H20 since DeepSeek took off. The H20 is Nvidia's tailored processor designed specifically to comply with U.S. export controls for the Chinese market.

These tech giants need the silicon for two reasons: their own internal product workloads and their massive cloud platforms, where businesses are lining up to host DeepSeek models.

It is not just big tech buying, either. A source at a major Chinese server maker revealed that companies in healthcare and education are actively buying AI servers equipped with H20s and DeepSeek. In the past, only cash-rich banks and telecom operators could justify those server racks.

There is also a looming threat that the Trump administration might restrict H20 sales. While preemptive stockpiling could play a role, Reuters' sources insisted that the real spark behind this rush is pure, unadulterated DeepSeek demand.

Nvidia declined to share exact sales figures, simply stating its chips win on merit in a competitive field. The company reports earnings on Wednesday, and its stock has already clawed back nearly all of the DeepSeek selloff, down just 3% year to date.

The inference boom is real

The fundamental mistake critics made was confusing training costs with inference volume. DeepSeek proved you can train smarter and cheaper, but inference—running everyday queries and autonomous agents—scales with adoption.

As Nori Chiou, investment director at Singapore-based White Oak Capital Partners, pointed out, many assumed compute demand would stagnate. In reality, better and cheaper models integrate deeper into regular life, driving inference compute needs up exponentially.

Why it matters

For companies and developers, this completely reframes how we look at AI efficiency. Cheaper open models do not kill compute; they democratize it. When running a powerful model becomes affordable, millions of businesses that sat on the sidelines suddenly need chips to deploy it.

For the broader industry, it shows that Nvidia's moat is far more resilient than skeptics believed. DeepSeek did not cannibalize the hardware market; it unlocked a flood of downstream inference demand that still runs predominantly on Nvidia's architecture.

We think the immediate rush to grab H20s makes total business sense for Chinese firms looking to cash in on DeepSeek today. However, the elephant in the room is geopolitical risk. If the White House moves to ban the H20, Chinese firms could face a harsh ceiling on their inference capacity just as everyday user adoption is taking off.