DeepSeek Pauses Fundraising After Founder's Remarks Leaked

Authored by Arthur Zhang via The Epoch Times,

Chinese artificial intelligence (AI) startup DeepSeek has paused a second fundraising round after founder Liang Wenfeng’s remarks at a conference circulated online, including a statement that the company could obtain some processors he described as “noncompliant.”

Liang made the remarks during a May 20 closed-door investor meeting on why DeepSeek was receiving fewer Huawei processors than China’s major internet companies. The transcript of the meeting was shared online and posted on GitHub on July 22.

“For us, we can buy some noncompliant cards,” he said.

He did not identify the chips, suppliers, purchasing route, or rules he was referring to.

DeepSeek has not publicly confirmed the transcript. Chinese state media National Business Daily reported that institutions involved in investing in the company confirmed that the meeting occurred and the transcript’s content was accurate.

Bloomberg reported on July 25 that DeepSeek representatives verbally told some prospective investors they would not be signing agreements in the coming days. Discussions continued, and the company could resume the fundraising process later, they said.

DeepSeek was seeking a valuation of about 500 billion yuan ($74 billion) after raising approximately $7.4 billion in its first outside financing round. The pause was driven partly by Liang’s frustration over online reports about his comments to investors, according to Bloomberg.

Chip Access

Liang described advanced processors as DeepSeek’s most valuable use of capital.

He said the company would be willing to turn all available funding into Nvidia chips and pay a premium if enough hardware could be obtained.

“The best thing would be to turn all the money into cards,” he said, using an industry term for graphics processing units (GPUs), the processors used to train and run advanced artificial intelligence models.

Liang said spending 20 billion yuan ($2.8 billion) on processors during the year would represent an exceptional performance by DeepSeek’s purchasing team. Nvidia B200 processors would be worth buying in any available quantity at a reasonable price, he added.

The remarks come as Washington continues to examine how DeepSeek obtained its computing hardware.

In April 2025, the House Select Committee on the Chinese Communist Party said DeepSeek appeared to use tens of thousands of Nvidia processors that are restricted from export to China.

The committee separately asked Nvidia CEO Jensen Huang for records concerning DeepSeek’s acquisition of export-controlled chips, including communications with customers in China and Southeast Asia.

The committee’s findings did not establish that the processors Liang called “noncompliant” were Nvidia products or identify how DeepSeek obtained them.

U.S. controls have continued to evolve. In guidance issued on May 31, the Commerce Department’s Bureau of Industry and Security said a license is required to export covered advanced computing items to companies headquartered in China or whose parent companies are based there—even when the immediate recipient is located in another country.

Compute Gap

DeepSeek drew global attention in January 2025 after releasing lower-cost models that appeared to challenge assumptions about how much computing power companies needed to compete with leading U.S. developers.

Investor concern that less expensive models could reduce demand for costly data centers and processors drove Nvidia shares down nearly 17 percent in one trading session, wiping almost $594 billion from the company’s market value.

Liang’s private account described limits that DeepSeek’s engineering efficiencies had not eliminated.

He said the company possessed computing power equivalent to approximately 20,000 Nvidia H-series processors. Liang did not specify which H-series model he used for the comparison and said much of the equipment had arrived recently or was still awaiting installation.

Even if DeepSeek spent 50 billion yuan ($7 billion), it could not train a model at the scale of the largest U.S. systems, he said.

Liang estimated that such a training run would require approximately 50,000 Nvidia GB300 processors or 200,000 Huawei Ascend 950 processors. That figure did not include the numerous experiments needed before a final model-training run, he said.

He put Chinese developers one to two years behind their U.S. competitors while operating with about a 20th of the computing resources.

Chinese companies could outperform U.S. rivals in selected tasks through engineering efficiency, Liang said, but could not achieve broad superiority while the hardware gap remained so large.

Hardware is not DeepSeek’s only constraint.

Liang said roughly half of the company’s core researchers were working on data annotation—the process of reviewing and correcting material used to refine an artificial intelligence model after its initial training.

He said U.S. developers OpenAI and Anthropic had accumulated an advantage in high-quality post-training data that additional funding could not quickly erase.

Huawei’s Limits

Huawei has become central to Beijing’s effort to reduce China’s dependence on Nvidia and other U.S. semiconductor companies.

Liang said Huawei was allocating DeepSeek about 16,000 Ascend 950 processors, while major Chinese internet companies were receiving allocations in the hundreds of thousands.

DeepSeek’s allocation would provide computing power equivalent to about 4,000 Nvidia B-series processors, by Liang’s estimate.

That would be enough for DeepSeek’s current model generation but not its next one, he said.

Liang said four Huawei processors were needed to match the computing capacity of one Nvidia processor. He also estimated that Nvidia hardware could remain economically useful for about five years, compared with no more than three years for Huawei processors.

Huawei systems would become more expensive to operate as they aged because of their electricity consumption, he said.

Production capacity posed another problem. Huawei could not provide enough processors to satisfy demand from both DeepSeek and China’s major internet companies, according to Liang.

He said DeepSeek was purchasing the Huawei processors partly to help Huawei improve its chip and software ecosystem, even though the allocation was too small to train DeepSeek’s next model.

DeepSeek was also developing software intended to make its models easier to move among Nvidia, Huawei, and other processors. Nvidia’s Compute Unified Device Architecture, its widely adopted programming platform, has made switching hardware costly because developers have built years of software and expertise around it.

Liang also described lower prices as a way for Chinese developers to compete despite their hardware disadvantage. DeepSeek could accept smaller margins than U.S. rivals, he said, while companies trying to operate its publicly released models independently would struggle to match its costs.

Before the May meeting opened to investor questions, the moderator said Liang had disclosed sensitive information and instructed participants not to circulate figures—including processor quantities—or share recordings and screenshots outside the meeting.