This week, The Epoch Times reported that DeepSeek paused a fundraising round due to comments made by its Founder, Liang Wenfeng, at a conference. These comments focused on the superiority of “noncompliant cards” and DeepSeek’s ability to obtain them. While Nvidia wasn’t named, it’s clear he was commenting on his ability to access Nvidia GPUs which he believes are superior to the Huawei ones that the CCP insists Chinese LLM AI makers use. These comments both confirm some things that I wrote about 18 months ago and also shed some light on the business.
As a starting point, I’m curious about Liang Wenfeng’s intentions. He’d be well-aware that the CCP draws no distinction between the interests of the party and individual businesses. He’d also be aware of the strange disappearance of Jack Ma, one of China’s wealthiest businessmen, following his criticism of Chinese regulators. Liang wasn’t caught by a hidden camera or the surprise leak of a private email. He was speaking to a group of investors and there is a transcript of the meeting circulating online. He may have thought he was speaking in private although if that’s the case, it’s not clear to me how there’d be a transcript. At a minimum, one of the investors in the room decided to ensure the meeting wasn’t private. DeepSeek isn’t confirming the transcript and Chinese state media says it’s accurate. Under the circumstances and based on what I know, I think it’s more likely than not that Liang believes what he said. Let’s go through the details:
Liang said that buying Nvidia’s best GPUs would be DeepSeek’s best use of capital and that he’d want to buy as many as possible. The Chinese government has indicated it wants Chinese companies to rely on Huawei. In April last year, the US government determined DeepSeek used huge amounts of Nvidia processors that were illegal to export to China. I came to this same conclusion three months before the US government did.
While the House Select Committee on the Chinese Communist Party didn’t identify how DeepSeek obtained the Nvidia GPUs, In January of 2025, I offered two obvious possibilities. First, GPUs are small, lightweight, and expensive. It would have been easy to have people or companies in other countries buy the best Nvidia AI accelerators, and trivial to put them on a plane to China. I remember reading about one small town in Vietnam that purchased more GPUs than they had power to run them. If Vietnam couldn’t use all the cards they bought, then those cards had to have been purchased for others. The second option is even more simple. There are plenty of companies that buy GPUs and rent them by the hour. Chinese companies don’t need to import “noncompliant” GPUs to make use of the ones purchased by others.
In January, 2025, I was clear that I didn’t believe DeepSeek trained its models on the widely-reported estimate of $50MM. The Epoch Times article claims DeepSeek had tens of thousands of the best Nvidia cards. Secondary prices for H100 cards in 2024 were in the $35k – $45k range. If DeepSeek had 25,000 cards at $40k each, that adds up to a billion dollars. DeepSeek may have done more with less, but $50MM was always a number that wasn’t credible. However, many did believe it as Nvidia’s stock lost almost $600B of value the day after the DeepSeek announcement.
Liang confirmed DKI’s conclusion saying that even with $7B of spending, they couldn’t reach the quality of the US models. He thinks that it would take 50,000 Nvidia GB300 processors to compete with the US entrants. He also said that to get the equivalent amount of computing power would require 200,000 Huawei processors. That implies that the Chinese GPU maker’s best chips are about 25% as fast and effective as those made by Nvidia. He said that the Chinese chip makers are 1-2 years behind Nvidia. Given the rate of improvement we typically see in GPU development, those two comparisons are comparable. Note that this doesn’t mean Huawei will catch up in 1-2 years. Liang was saying that Huawei would be where Nvidia is now in that timeframe. Nvidia will continue to develop new and better GPUs. Jensen Huang isn’t going to sit still and wait to be caught.
There has already been a response to Liang’s comments becoming public. In addition to putting the current fundraising round on hold, Huawei has allocated DeepSeek only 16,000 of its high-end Ascent 950 processors. Using the previous comparison, that’s the equivalent of just 4,000 of the best Nvidia cards. Chinese competitors are receiving hundreds of thousands of Huawei processors while US competitors are deploying hundreds of thousands or even millions of the (currently) better Nvidia versions. I’ve always questioned DeepSeek’s ability to produce results with their claimed level of minimal resources. Unless they continue to access high-end Nvidia products, something that will be increasingly difficult given the attention now on the company, we’re going to find out what DeepSeek really can do with less.
Liang said something else that, if true, affects the depreciation schedules of the US hyperscalers. He said that Nvidia GPUs had a useful life of five years and that the less-energy-efficient Huawei ones had a three-year useful life. Previously, I’ve commented that the US hyperscalers are on an unprofitable treadmill. Even as losses from AI development accumulates and token pricing falls, the big LLM developers have to keep spending or see their previous spending become worthless because new AI models surpass the previous leaders every few months. Other sharp-pencil researchers have pointed out that these big US companies were probably using an unrealistic depreciation schedule that’s making earnings look better than they should. I agreed with their accounting concerns. After all, who’s doing frontier work on a 5-year-old Nvidia card? I’m skeptical that Liang is right on this point, but if he is, the current depreciation schedules of the big US companies is probably correct and earnings aren’t overstated.
The final point emphasized by The Epoch Times was that Liang said Chinese developers could accept lower margins than US companies. I agree and believe we’re seeing that now. In multiple recent editions of DKI’s weekly 5 Things to Know in Investing, we’ve commented on the Chinese models that have been 90% as good as the US models, but as much as 90% cheaper. Recent Chinese releases appear to have caught up to the US models in some tests and more end users are gravitating towards the better value. As much as Liang’s comments seem to have caused consternation in the CCP, the Chinese fundraising community, and the executives at Huawei, his last point illustrates a strength of the Chinese approach. US companies have better funding and better access to the best GPUs, but they should focus on the fact that they’re facing price competition from Chinese models that are rapidly closing the performance gap.
Information contained in this report is believed by Deep Knowledge Investing (“DKI”) to be accurate and/or derived from sources which it believes to be reliable; however, such information is presented without warranty of any kind, whether express or implied and DKI makes no representation as to the completeness, timeliness or accuracy of the information contained therein or with regard to the results to be obtained from its use. The provision of the information contained in the Services shall not be deemed to obligate DKI to provide updated or similar information in the future except to the extent it may be required to do so.
The information we provide is publicly available; our reports are neither an offer nor a solicitation to buy or sell securities. All expressions of opinion are precisely that and are subject to change. DKI, affiliates of DKI or its principal or others associated with DKI may have, take or sell positions in securities of companies about which we write.
Our opinions are not advice that investment in a company’s securities is suitable for any particular investor. Each investor should consult with and rely on his or its own investigation, due diligence and the recommendations of investment professionals whom the investor has engaged for that purpose.
In no event shall DKI be liable for any costs, liabilities, losses, expenses (including, but not limited to, attorneys’ fees), damages of any kind, including direct, indirect, punitive, incidental, special or consequential damages, or for any trading losses arising from or attributable to the use of this report.
