DeepSeek and the DKI Portfolio

As you all know by now, DeepSeek caused an implosion of the US-based technology and energy space on Monday. On Tuesday, the market decided it has over-reacted and many of the heavily-sold stocks regained ground. I want to address the speculation on both sides of the discussion and to update you on how this affects the DKI portfolio. I’ll also acknowledge that most of what’s “known” right now represents estimates based on logical inferences rather than confirmed information. I will also note that while we’ve all read a lot about AI in the past two years, I don’t have a background in computer science.

UPDATE:

In the few hours since I posted this there have been two additional pieces of news.

– Alibaba announced they have an AI model that exceeds the performance of DeepSeek. They announced this on the Lunar New Year (Year of the Snake) when most of China wasn’t working potentially indicating they don’t want to lose even one day to a rival company. This would imply even more competition coming in the non-US AI space and that US limitations on Nvidia exports to China have been ineffective.

 – OpenAI is claiming that DeepSeek obtained their proprietary intellectual property and used it to train DeepSeek’s AI. In the article below, I express skepticism that DeepSeek actually was able to equal the performance of ChatGPT on a total of $45MM – $55MM of spending. If OpenAI’s claims of IP theft are true, it would provide an explanation for how DeepSeek was able to achieve comparable performance for one-twentieth of the cost of the US AI models.

Capability:

This is one place where we don’t need to make guesses. DeepSeek is performing around the level of ChatGPT on standard benchmark tests.

 

Efficiency:

It’s apparent that DeepSeek has come up with a solution that’s almost as good as the US-based AI models, but with far fewer resources. As I understand it, the US models use massive amounts of computing power, to cover as many topics as possible great depth. The Chinese DeepSeek model created efficiency by focusing on the topics that are most likely to be queried and doing so to a lesser level of depth. The computational effort required by DeepSeek is a fraction of what is required of models like ChatGPT. That’s probably fine for most inquiries. It is potentially a limitation for some more detail-oriented or technical ones. Stated simply, most of us might not notice a difference for daily use, but for some specialized and corporate applications, the US approach might be worth the extra cost.

 

While DeepSeek is more efficient in training the model, because it has to do more inquiry and research to answer each individual query, it uses more time, computing power, and energy to answer each question as its posed. This means there are energy savings, but probably not as much as is being widely reported.

 

None of This is Static:

These AI models are constantly being revised. It’s possible that DeepSeek will access greater computing power and pursue certain topics in greater depth. It’s likely that the US hyperscalers will learn from DeepSeek’s approach and become more efficient. In a move that I applaud, DeepSeek is open source meaning they make their code available to the public to study, copy, and use. It’s not speculative to assume that teams at Facebook, Google, Apple, and Microsoft are working around the clock to study and understand DeepSeek’s model.

 

Three Things I Believe Are Not True:

Cost:

DeepSeek claims they used only $40MM – $50MM of hardware and about $6MM of computing costs to develop a model that is almost as good as the US ones that cost about 20x more. China had at least three ways of circumventing the Biden restrictions on Nvidia delivering their best GPUs to Chinese companies.

 

First, while Chinese companies couldn’t buy Nvidia’s best GPUs in the quantities the big US technology firms did, it’s pretty hard to track the location of millions of computer chips. I’ve seen estimates that tens of thousands of the high-end versions made it to China. This is something I can’t confirm, but it seems likely that companies in China found a way to source some of the best GPUs.

 

Second, it’s been an open secret for a long time that Chinese companies have been able to rent time on high-end GPUs in other countries. Renting time on someone else’s hardware might not be the most cost-efficient way to proceed, but I believe it happened.

 

Third, the Biden restrictions related to data transmission speed. Nvidia designed a special chip for the Chinese market that complied with these restrictions, but were otherwise comparable in performance. The common narrative was that Chinese companies were forced to use GPUs that had been severely limited in performance. The truth is they were offered chips that were closer to what US companies had than had been reported.

 

I’d like to specify that I’m not criticizing the prior White House on these matters. AI was the fastest-moving part of the technology world during the past two years. I think anyone in any Administration from either party would have struggled to quickly come up with comprehensive regulations that would anticipate the next generation of GPU progress. They did the best they could under challenging circumstances, and I respect that. Nvidia found a way to comply with the restrictions as written. This was technically legal.

 

Finally, one DKI subscriber contacted me with his opinion that DeepSeek may have only spent $45MM – $55MM, but they probably had assistance from the Chinese government. I think this is highly likely as do many of the other finance and technology people I’ve spoken with this week. Many observers have noted that it’s quite a coincidence that a Chinese company claims to have equaled US technology dominance with a fraction of the cost just days after President Trump announces The Stargate Project to ensure future US dominance of AI.

 

Security:

Many people have speculated this week that any queries or data requests on DeepSeek meant your information was going straight to the CCP (Chinese Communist Party). I believe this is technically true but misleading. If you download the DeepSeek application and use that, then this speculation is likely to be correct. The CCP will have access to this version of your “browser history”. The civil libertarians in the crowd will point out that many US-based technology companies are cooperating with the US government in similar uncomfortable ways. That is also unfortunately true.

 

The reason I say this claim is misleading is because DeepSeek is open source. For companies and people with the resources and technical knowledge, it is possible to download DeepSeek’s model to your own server and use it privately with your own data. So, there is a way to use the model without tracking by the CCP.

 

This Means the End of Nvidia and US AI:

What DeepSeek did isn’t unusual in the history of technology. The incredible thing is the rate of change. Computer technology always advances at an exponential rate. The big news with DeepSeek was they cut 95% of the cost and the price to high-use customers overnight. The fear in the market is that the big US tech firms overspent on hardware and now everyone will become more efficient and no longer order new GPUs from Nvidia.

 

There could be some imbalance in this market and some future orders could be scaled back leading to risk in the current $NVDA stock price. That’s never the end of the story. There used to be an expression that “What Grove giveth, Gates takes back”. It meant that as Intel developed ever-faster processors, Microsoft designed more complicated operating systems that used up the increased processing power. It’s a never-ending cycle. DeepSeek may have just made a huge leap, but this isn’t and won’t be the end of the GPU business. The next generation of GPUs will be better and faster, and with or without DeepSeek, multiple AI firms are designing their own specialized chips right now.

 

When I was in college, I had an internship at Comerica Bank in downtown Detroit. I had to pass a class to use the company’s mainframe computer. Not everyone had their own PCs and the ones they did have couldn’t do everything required. As computers became more efficient, powerful, and cheaper, their use didn’t decrease. Use of computing power increased exponentially. We went from mainframes in clean rooms that only large corporations could afford to PCs for office use to PCs for home use. I use a home computer, a laptop, two tablets, and a cell phone. Anyone with kids uses a multiple of that many devices. Everyone’s devices regularly connect to cloud storage and applications which use computing power on many other computers. Anyone who looked at the massive improvement in computing power and efficiency decades ago and concluded that we’d have LOWER demand for computers and the energy required to power them would have been terribly wrong.

 

The DKI Portfolio:

As I wrote in a piece earlier this week, the place where we have exposure is in energy. Independent power producers and energy positions got hit hard on Monday and made a partial recovery on Tuesday. The fear is that DeepSeek’s efficiency means we won’t need more power for AI-related data centers. First, please note the previous section. More efficiency in technology leads to more use.

 

Second, as friend and fellow stock analyst, Enrique Abeyta, wrote to me yesterday, “We have a power crisis if we DIDN’T have the AI demand”. The US needs more energy infrastructure. Multiple large States including California and Texas have had recent shortages. In my home State of Connecticut, everyone is complaining about rapidly rising energy bills, and that’s with a relatively stable price of oil in the past couple of years. This is a widespread problem, not a local one. We don’t have the infrastructure for EVs if people buy what’s projected let alone AI datacenters.

 

Not only does the US need more energy production; but also, the rest of the world is screaming for it. There are a couple of billion people in China and India that want a higher material quality of life and that’s almost 100% correlated with energy usage. Africa is still burning wood, and through mismanagement, Germany is somehow burning wood and coal. The world needs more energy production with or without AI datacenters and if we want to do it without massive carbon emissions, that’s going to mean more nuclear power.

 

Nuclear plants are planned, built, and operated over a 50-year time horizon (or so). DeepSeek startled the world just a few days ago, but I don’t think it changed the way we live, or our energy needs over the next few decades.

 

I have a research call late tomorrow night (Vietnam time) with an AI infrastructure expert. I’ll update DKI premium subscribers if anything I learn causes me to update my thoughts on the subject and the portfolio.

 

For those of you with questions and opinions, you are always welcome to reach me at IR@DeepKnowledgeInvesting.com

 

 

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. 

Leave a Comment

Recent Blogs

Referral program

Invite & Earn

X
Signup to start sharing your link
Signup
background banner image
loading gif

Available Coupon

X