DeepSeek's Next Move: Trained on Forbidden Nvidia Chips, Coming March 2026

They did what now? DeepSeek—the Chinese AI lab that shook Silicon Valley by matching OpenAI's best model for 3% of the cost—just revealed its next trick: training on Nvidia's most advanced AI chip, the Blackwell, despite US export bans. Here's why this matters and what's coming in March.

The News: Blackwell Despite the Ban

According to a senior Trump administration official via Reuters, DeepSeek's next-generation model—expected to release as soon as March 2026—was trained on Nvidia's Blackwell chip. This is the same chip the US government explicitly banned from being shipped to China.

The US believes DeepSeek will likely remove technical indicators that reveal use of American chips. The Blackwells are reportedly clustered at a DeepSeek data center in Inner Mongolia.

This isn't just a regulatory绕 (workaround)—it's a middle finger to US export controls. And it raises a uncomfortable question: if China's most advanced AI lab can get hold of banned chips anyway, what's the point of the restrictions?

What's Coming: The March 2026 Model

DeepSeek hasn't officially named the model yet, but here's what we know:

• Trained on Nvidia Blackwell — The most powerful AI chip Nvidia makes
• Expected release: March 2026 — Could drop any week now
• Trained using distillation — Using outputs from Anthropic, Google, OpenAI, and xAI models to improve reasoning
• Withheld from US chipmakers — DeepSeek didn't give Nvidia or AMD access to benchmark or optimize the model

In a twist, DeepSeek gave Chinese chipmakers—including Huawei—a weeks-long head start to optimize their software for the new model. That's a major win for China's domestic chip industry.

The Price That's Still Blowing Minds

Let's recap what DeepSeek already delivered. Their R1 model matched OpenAI o1 on benchmarks—for 3% of the cost:

That's 27x cheaper than OpenAI o1 for output tokens. A 100,000-token conversation costs $6.50 with OpenAI—or $0.27 with DeepSeek.

And DeepSeek trained the entire V3 model for an estimated $6 million. OpenAI reportedly spent over $100 million on o1. The new Blackwell-trained model could narrow that gap—or widen it further.

The Geopolitical Mess

This is where things get complicated:

And then there's the privacy issue. DeepSeek's own privacy policy states data is stored on servers in China. Security researchers found code connecting to China Mobile, a state-owned telecom.

For enterprise users: this matters. If you're handling sensitive data, DeepSeek's API isn't your friend. But the open weights? Run locally, and you get the performance without the privacy risk.

What This Means for You

As a user:

• DeepSeek's API is absurdly cheap—great for non-sensitive prototyping
• The open-weight R1 model runs locally if you have the hardware
• Wait for the March model if you want something even better

As an industry:

• US chip export controls are ineffective—if DeepSeek can get Blackwells, so can others
• AI pricing will keep plummeting ($2.19 → ?)
• China's domestic chip industry just got a massive incentive to accelerate

The Bottom Line

DeepSeek isn't just disrupting AI technology—they're disrupting geopolitics. By training on forbidden chips and withholding the model from US companies, they're making a statement: the AI race won't be constrained by trade walls.

For developers and businesses, the takeaway is simple: the pricing revolution is real. Whether the new March model drops next week or next month, one thing is clear—the old AI pricing models are dead. And that? That's good for anyone who wants to actually use AI without going broke.