U.S. officials target DeepSeek for training new AI model on restricted Nvidia chips

DeepSeek plans to release a new AI model as soon as next week, and U.S. officials say it was trained on Nvidia’s most advanced Blackwell chips.

A senior official in Donald Trump’s administration allegedly said on Monday that this could break current U.S. export control laws. Those rules block the sale of Blackwell processors to China.

The official said, “we’re not shipping Blackwells to China.” He did not explain how the government got the information or how DeepSeek obtained the chips.

U.S. officials believe DeepSeek may remove technical signs that could show the chips came from the United States. The same official said the Blackwell units are likely grouped together at a DeepSeek data center in Inner Mongolia. That region is an autonomous area in northern China known for large energy projects and server farms.

If the U.S. confirms that DeepSeek accessed these chips, it could deepen political tension in Washington over how far export limits should go.

Officials clash over chip sales and security rules

The Chinese embassy in Washington rejected the accusation. In a statement, it said China opposes “drawing ideological lines, overstretching the concept of national security, expansive use of export controls and politicizing economic, trade, and technological issues.”

On Tuesday, foreign ministry spokesperson Mao Ning told reporters she was not aware of the specific case. She added that China had already made its position clear on U.S. chip restrictions.

Export controls fall under the U.S. Commerce Department. Blackwell shipments to China are currently banned.

In August, Donald Trump signaled that Nvidia could sell a reduced Blackwell version to Chinese buyers. He later changed direction and said the company’s top chips should stay inside the United States.

In December, he allowed Chinese firms to purchase Nvidia’s second most advanced chip, the H200. That decision drew sharp criticism from lawmakers who take a hard line on China. Shipments of H200 chips are still paused because of approval guardrails.

White House AI czar David Sacks and Nvidia CEO Jensen Huang argue that allowing limited sales could slow the rise of Chinese rivals like Huawei. Some lawmakers disagree. They fear advanced chips could support military systems and weaken U.S. dominance in artificial intelligence.

The official also said the new DeepSeek model likely used “distillation” from leading U.S. AI systems built by Anthropic, Google, OpenAI, and xAI.

Distillation means a stronger model checks and grades the answers of a newer one. The smaller system then learns from those results. OpenAI and Anthropic have raised similar concerns before about Chinese developers using this method.

Hangzhou-based DeepSeek drew global attention early last year when it released models that rivaled leading U.S. systems. That release raised alarms in Washington about China closing the AI gap despite trade barriers.

Investors watch for Nvidia earnings and AI spending data

All this unfolds days before Nvidia reports fourth-quarter earnings on February 25. Last quarter, revenue beat guidance by almost $3 billion and showed $10 billion in sequential growth.

Despite those numbers, the stock fell from $186.52 on earnings day to $180.64 the next session. Shares had earlier peaked at $207.04 on October 29. They later dropped and closed at $171.88 on February 5 after a partial recovery.

Morgan Stanley analyst Joseph Moore expects strong Q4 results. Joseph said he sees at least $2 billion above the current $64 billion revenue guidance.

Consensus sits near $72 billion. He said the ramp of the Vera Rubin chip should stay on schedule. He also noted demand for AI compute remains strong across DRAM, NAND, HDD, optical, CPU, and power supply chains.

Bank of America analyst Vivek Arya updated his outlook before earnings. Vivek said cloud capital spending could reach $748 billion in 2026 and $869 billion in 2027.

That represents growth of 56% and 16% year over year. He expects the data center systems market to expand 64% in 2026, while AI systems could grow close to 100% because of new accelerator deployments.

Even without Google, which uses its own TPUs, 2026 cloud capex may rise about 45% to $180 billion. Vendors are expected to pass higher HBM and DDR costs to customers to maintain margins as system sales increase.

If you're reading this, you’re already ahead. Stay there with our newsletter.