(B) I have listened to stock analyst and portfolio manager Dan Niles for a very long time on CNBC. Mr. Niles made an interesting comparison recently between Cisco’s P/E in 2000 and Nvida’s P/E in 2024, and made me think to push further that comparison between the two companies.

During the 2000 Internet bubble, Cisco’s share price fell 90% from a high of $80 in March 2000, with a P/E of 196 and a market cap of $546 billion, to a low of $8 in October 2022. Today Cisco’s share price is just lower than $50.

In 2000, Cisco was the most valued company, ahead of Microsoft and Intel. Google was still a private company. The Internet was highly fragmented, and Cisco was selling routers and switches to enterprises and Internet Service Providers (ISPs) while using its high market value to buy any threatening competitors or to penetrate any new networking market such as the optical transport market for ISPs.

In March 2024, Nvidia’s market cap soared to $2.25 trillion with a share price of $974 and a P/E of 79. The Internet is highly concentred and owned by Google, Apple, Microsoft, Amazon, and Meta. If we exclude Tesla, Nvidia is the only equipment provider among the magnificent seven.

So let try to discuss why did Cisco lose so much market value, and what could be the threats to Nvidia’s present valuation?

This is obviously a controversial blog post, which most content is likely incorrect. But it is a good and fun intellectual exercise for the business mind :). So please do not take the following hypotheses too seriously!

Why did Cisco lose so much market value?

• A PhD student at Stanford reinvented routing:
  • DARPA worked on it with active networks. I tried to work on it as well with various engineers on leveraging mobile agents that could run and do many things over large Internet networks. I thought that Van Jacobson would succeed but it was Martin Casado, when he was a PhD student at Stanford, succeeded to reinvent routing by decoupling the control plane from the data plane, an idea that became software-defined networking (SDN).
• Google’s data center engineers read the research papers from Mr. Casado:
  • I attended a couple of meet-ups at the Googleplex about Google’s data center architectures – read in particular – providing Large Scale Cloud Services by Virtualizing Network Services over Software Defined Networks. With SDN, Google did not need Cisco, just some cheap Linux servers for the data plane and a centralized control plane.
• Cisco is not selling to large enterprises:
  • The cloud killed the Intranets. No need for an entreprise to build campus networks and remote networks with Cisco’s switches and routers when all enterprise applications are in the cloud.
• Cisco is not selling to Internet Service Providers (ISPs):
  • Google, Apple, Microsoft, and Facebook killed the ISPs. When all the Internet traffic goes to a few sites, large ISPs are not needed any more, and large ISPs networks are disappearing. So Cisco’s sales to ISPs went down.

It is interesting to note, that if my analysis about Cisco is correct, none of those above reasons were Cisco’s management mistakes.

What Could be the Threats to Nvidia’s Present Valuation?

(Preliminary note: I assume that most of the strong and recent Nvidia’s revenues growth is due to LLMs, and so any threats to the present deployments of LLMs would be a threat to Nvidia’s future revenues growth, assuming Nvidia’s management will not compensate LLMs GPUs decreasing revenues by another business opportunity).

Business risk ~ HIgh:

• Mr Casado’s younger sister or brother is, as we speak, re-inventing the GPU at Stanford:
  • There is probably right now at Stanford, Berkeley, MIT, Ecole Polytechnique, or University of NoWhere, a very smart PhD student that is probably working on the next generation of processors.
• Top Nvidia’s researchers and engineers are leaving Nvidia:
  • All the Google researchers that worked on the Transformer left Google. As any Silicon Valley hot company, Nvidia could lose some of its talents to the venture community.
• Training and inference of LLMs and Large Models that are not Language-specific are re-invented:
  • I have heard that many times from different sources but there is a lot of work to be done to improve LLMs training and inference, and in particular how to apply the right amount of computation when it is needed.

Business risk ~ Medium:

• LLMs are not making money for enterprises:
  • LinkedIn has a generative AI app that helps you to find out which job you are a good match. I tried the apps and it told me that I was not a good match for a position because I did not have a BSCS when I do have a MSCS. Many present generative AI apps are incredibly buggy but may be it does not matter too much as long as you can make a press release about it. But how can someone expect to make money from such an apps in the long run?
• LLMs are not the future of Artificial General Intelligence (AGI):
  • Predicting the next token is probably not the best approach for reasoning and planning. Just listen to Yann LeCun‘s lecture for a strong opinion on that.
• Enterprises only need small domain specific L models not LL models:
  • Bloomberg developed BloombergGPT (blog article, paper) based on the open source HuggingFace’s Bloom for financial applications. BloombergGPT was only a 50 billion parameter LM. Now with Mistral, enterprises can build a domain specific language model with less than 7 or even 2 billion parameters.
• TensorFlow and PyTorch have copilot libraries:
  • Like Tensorflow includes recommender systems libraries and so does PyTorch, Tensorflow and/or PyTorch could integrate specific libraries to build copilot applications. I am not sure if that hypothesis will reduce the need for GPUs – it could work both ways. Enterprises will do less exploration ( => less GPUs) but developed more apps ( => more GPUs if GPUs are used in inference)

Business risk ~ Low:

• AMD, Intel, and a few start-ups have GPUs or specific processor offerings for LLMs:
  • For more on that – read: Approaches and Silicon for Parallel Training and Inference of LLMs
• Only Google for Google search and Microsoft for Bing train LLMs:
  • This is probably not a new hypothesis but an implication of a previous thought that LMs have more breadth than LLMs. Meaning that LLMs are only required as a companion to search and/or to help with a narrow landscape of applications such as language translation for rare languages, knowledge search, and coding search.
• Google is reselling its TPUs to the world:
  • I wrote about that hypothesis already.

Note: The picture above is “Danseuses Bleues” from Edgar Degas.

Copyright © 2005-2024 by Serge-Paul Carrasco. All rights reserved.
Contact Us: asvinsider at gmail dot com.