DeepSeek and the Need for Power Generation

Last night (Vietnam time), I had a conversation with an expert who has experience designing AI datacenters. In addition to some highly technical information, he made four points that I think are relevant to DKI subscribers and the portfolio. I note that there are many details still to be learned and this expert was speculating based on his experience; however, his experience in this particular area is extensive and on-topic. Let’s dive in:

 

  • He thinks DeepSeek didn’t do what they did with the resources they claim:

Earlier this week, I speculated that DeepSeek had additional resources including the possibility of illegal access to tens of thousands of Nvidia’s best GPUs, renting access to the best GPUs in other countries, and potential financial support from the Chinese government exceeding the $55MM (or so) DeepSeek claims they spent. This expert agreed and thinks it’s likely they are using the expensive high-end Nvidia chips that are subject to an export ban. He believes that models will get more efficient over time, but that there are also no shortcuts.

 

  • DeepSeek’s methodology isn’t good enough and will produce costly errors:

DeepSeek is more “efficient” because it doesn’t research every topic in great depth. That does provide for lower initial infrastructure cost, but it also greatly increases the probability of errors. Many AI users have reported AI models producing artifacts, or other “facts” and “events” presented as accurate. These confidently presented facts are completely made up and bear no resemblance to reality. Advice that I’ve gotten from programmers is that using AI models can allow them to program faster, but they need to check the machine-made results carefully because even the best models make errors. DeepSeek’s inference approach uses much less calculation intensity initially, but at the expense of accuracy of results. In addition, DeepSeek’s methodology means training a model is faster and cheaper, but responding to each individual query takes more time and is more energy-intensive. It’s not clear right now how large the long-term savings of DeepSeek’s model would be. It’s also not clear what the demand is for a cheaper model if it’s less accurate. Probably fine if you’re just making funny meme photos. Less fine if you’re doing research or programming.

 

  • As GPUs and semiconductors become more efficient, 100% of that progress will go towards more performance and 0% will go towards energy savings:

When people design for mobile devices like laptops and cell phones, they think a lot about battery life. That’s not the case for GPUs in AI datacenters which may be powered by a nuclear reactor. As technology advances and the chips become more efficient, all of that efficiency will be poured into better and faster performance. None of it will be used to save power. These chips are optimized for speed and he believes that practice will continue in the foreseeable future.

 

  • Increase in watts/rack is exponential:

Just 5 years ago, AI datacenter power usage was around 3 – 9 kW/rack. As recently as 2-3 years ago, that had risen to 30 – 40 kW/rack. Currently, he’s seeing power usage in the 75 – 100+ kW/rack range (some around 130 kW). The need for power and cooling is incredible and growing. He believes US firms have the best cooling solutions in the world and those solutions are exported widely. He also thinks it’s unlikely anyone will make a big leap in cooling technology because the laws of thermodynamics apply everywhere.

 

The conclusion:  Don’t sell your power generation or uranium positions yet!

 

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