saahityaedams

08 Jan 2025

Are Tech Stocks Overvalued ?

WIP

My good friend Rishabh argues that US Tech stocks are overvalued (on account of AI hype) and specifically that the “Edelweiss US Technology Equity FoF is not a good investment option. He cites this very well written memo from Howard Marks1 that suggests that the Tech Stocks and the S&P 500 shows signs of potential bubble behaviour.

I want to document my thoughts about this and hopefully clarify my thinking (for myself primarily) about why I think some tech stocks are still so attractive to me. [DISCLAIMER: I COULD BE COMPLETELY WRONG AND I OWN A SMALL NUMBER OF AMZN STOCK]

Narratives

AI

I personally believe that AI (specifically LLM systems) will have a significant impact on society in the medium term future. In the short term, it is already augmenting labour / increasing productivity for a ton of professionals (folks harping on about AI 2-3-4-5) (counter point: 6). I personally use Anthropic Claude 3.5 Sonnet API to a significant extent for reviewing my writing, generating software, building basic tools, and just general researching.

More specifically, the overarching question seems to be whether companies (specifically the mag 7) can capture significant value from “AI”. Here I have jotted down some trends (LLM commodization) and nvidia dominance (?) ) and have some theories - selling cloud services, improve internal operations.

LLM Commodization

It is partially true that there’s been a commodization of LLMs on the market. You can get light weight models (that perform basic tasks) running on your laptop (CPU or basic GPU) (7-8).

You can use large open sourced models like Meta’s Llama as APIs from certain providers for dirt cheap 9. A Chinese hedge fund called High-Flyer made an impressive model called DeepSeek V3 (but what was actually impressive was their technical report10 where they talk about how it cost only $6M to train and was done under sanctions on powerful GPUs in China).

This trend is a significant risk to flagship AI labs like OpenAI and Anthropic (where they make ground-breaking technical contributions but end up bust like Netscape or Sun because they didn’t have a economic moat). On the other hand, the flagship AI labs have been pursuing new techniques like reasoning11 in models in recent times, so they might have edge there.

Nvidia Dominance (?)

It is true that most companies today are training/running their models on Nvidia chips, which has led to Nvidia having fat profit margins and be valued at a high premium on the stock market. Even though Nvidia is the current leader by being at the right time and at the right place, I don’t think this moat is sustainable. There’s growing competitors in AWS (Tranium chips which Anthropic uses12), Google (TPU chips), Groq chips (super fast inference time 13).

We also see this in a saturation of certain Nvidia chips on the market. The rental market for H100s has filled from shortage ($8/hr) to oversupplied ($2/hr) 14.

Selling Cloud Services

Amazon, Google, Microsoft have a gigantic sprawl of infrastructure (data centers, servers, networking) and the technical chops of operating it (primary leaders: since this is their core competence, plus a decade of experience in operating large scale public clouds).

These trends - the commoditization of LLMs and NVIDIA’s weakening grip on AI hardware - favor big tech companies. It’s eerily similar to how they built their cloud business by capitalizing on open-source software and commodity hardware a decade ago15.

Improve Internal Operations

Another interesting scenario that could play out is that big tech companies start figuring out how to realize significant economics gains from “AI”. One early indicator of this would be headcount reduction (big tech companies have a large base of contractors for fairly routine activities like annotating, data entry).

I would be pretty dissapointed if the most significant use case of LLMs for big tech end up being just cost saving by headcount reduction and not some improvements in internal systems via some other metrics like better operations and increased customer satisfaction.

Financial Engineering

Something I’ve noticing is that the tech giants have started getting serious about making more money (ie. becoming operationally efficient, tightening their belts,taking a larger tithe from customers, and engaging in more aggressive business practices). This has already started showing up in metrics like operating margin, free cash flow, etc and I expect this trend to continue with big tech companies pulling many more of such tricks from up their sleeve.

Plus, I firmly believe that some big tech companies are clearly monopolies based on certain characteristics (like scale, competition, data moats, network effects). This also means that they have an amount of additional significant leverage due to demand inelasticity.

The flip side of this, the big tech companies has been under legal scrutiny recently with antitrust and monopoly cases. But if I had to make an uneducated guess, this isn’t going to really make a significant impact on any of these company’s bottomlines especially during a Trump administration and with a conservative supreme court.

Examples

  • Amazon warehouse regionalization (supposedly led to decrease in cost to serve by $0.45 dollars) 16
  • Introduction of Prime Video Ads 17
  • Youtube making it harder to block Ads 18
  • Chrome’s V3 manifest switch (leading to ad blockers breaking) 19

AI - An Indian (?)

This is slightly controversial and sensitive, but I get the sense that the big tech companies are aggressively hiring in India (or Philippines) vs the muted hiring in the USA20 for white collar professionals. They can have their choice of young, agreeable employees who cost significantly less from the gigantic stockpile that is the indian youth. Globalization of high paid white collars employees like tech workers is increasingly becoming true. I expect big tech companies to decrease their payroll costs over the next few years this way.

More datapoints (hopefully quantitative)

Fair Valuation

Amazon, Google, Microsoft, Apple

TODO:

  • ton of data and metrics
  • discuss likely hypotheticals.

Specifically the “Edelweiss US Technology Equity FoF”

TODO

  • look at expense ratios
  • one of its kind in the market to get exposure to US Tech

USD to INR

TODO

  • fundamentaly believe that inr is going to get weaker.

  1. https://www.oaktreecapital.com/insights/memo/on-bubble-watch ↩︎

  2. https://antirez.com/news/140 ↩︎

  3. https://newsletter.pragmaticengineer.com/p/ai-tools-for-software-engineers-simon-willison ↩︎

  4. https://nicholas.carlini.com/writing/2024/how-i-use-ai.html ↩︎

  5. https://www.geoffreylitt.com/2023/03/25/llm-end-user-programming.html ↩︎

  6. https://xcancel.com/sh_reya/status/1873431565650502060#m ↩︎

  7. kick-ass vision model: https://moondream.ai/ (eg: describe an image, get license plate no from an image, intruder detection, etc) ↩︎

  8. speech recognition model: https://github.com/openai/whisper/blob/main/model-card.md (eg. subtitle a video, make a transcript of a meeting recording) ↩︎

  9. https://compass.arc53.com/ ↩︎

  10. https://arxiv.org/pdf/2412.19437 ↩︎

  11. https://openai.com/index/learning-to-reason-with-llms/ ↩︎

  12. https://www.anthropic.com/news/anthropic-amazon-trainium ↩︎

  13. https://news.ycombinator.com/item?id=41162875 ↩︎

  14. https://www.latent.space/p/gpu-bubble ↩︎

  15. https://gwern.net/complement ↩︎

  16. https://www.aboutamazon.in/news/company-news/amazon-ceo-andy-jassy-2023-letter-to-shareholders ↩︎

  17. https://www.aboutamazon.in/news/entertainment/prime-video-ads-update-2025-india ↩︎

  18. https://news.ycombinator.com/item?id=40666304 ↩︎

  19. https://news.ycombinator.com/item?id=38369820 ↩︎

  20. I am trying to find some statistics (combing through 10Ks) around this or atleast some proxies to confirm this but am unable to find good data. Though anecdotally this is the sense I got when I was previously worked at a big tech company. ↩︎