Thank you, DeepSeek

3 Reasons why I'm Thankful for DeepSeek

Everyone started freaking out with the release of DeepSeek’s r1 model, which is comparable to western models, despite spending fraction of the cost. I’m not here to argue if their claim is true or not — I’ll leave that up to you to decide. However, I’m going to write this short essay under the assumption that it is somewhat true.

I don’t believe for a second they spent only $6 million on the model and got it to this scale; however, I do 100% believe the Chinese made a frontier breakthrough in efficiency. It made us rethink what we know, and puts a fire under the bellies of OpenAI, Anthropic, Meta, and the United States government to become more efficient.

That said, in this essay, I want to discuss my thoughts on this whole fiasco. More importantly, I want to explain why I’m grateful for DeepSeek and am excited about the future of AI and startups.

1) Thank you, DeepSeek, for putting a fire in the bellies of Western incumbents (and the US government).

Firstly, we’re all overreacting that this is the doomsday for infra companies. We saw NVIDIA shed nearly $600B in market cap (for context, that’s almost the size of Coca Cola market cap… but times two).

DeepSeek’s release tells one thing to companies training large language models: Be better or else you’ll succumb. Putting this fire in their bellies will only force these companies to become more competitive and efficient, which ultimately is only more beneficial to the consumer (I’ll relate this back to the app layer later on). It’s critical to understand that everything we’re seeing with AI is still extremely early. We’re still using chatbots — we’re still in phase one of AI as I like to call it. We’ve barely touched on agents, robots, fully autonomous organizations even, etc.

ChatGPT has 200 million active users, Google have 5 billion users. What we’re seeing now is nothing compared to what we’ll see in the near future. With that being said, Western incumbents have seen that there are better and more efficient ways to build their LLMs. This will only translate to more efficiency in the future, which should get all of us excited.

Clearly, this is under the assumption that these incumbents need to execute and figure it out. But I think they will, and I think Nvidia’s GPUs will be necessary, especially when we we reach the subsequent phases of AI.

Let me touch on the United States government as well. The reason why China is leaping us because 1) government support and 2) work ethic (I know it’s controversial, but gosh we’re lazy). I’ll only speak to one because I think two is too controversial. “Biden issued an executive order which sought to constrain compute under an arbitrary threshold, bar open source as an alleged threat to national security, and effectively allow regulatory capture by the biggest players. The administration and its enablers wanted to limit math, and in turn, limit code—but ended up just limiting America’s lead” (Alex Rampell, The FP). With that said, this hopefully makes us rethink entirely how the country should approach AI. This is no different than the space race — in fact, it’s more important.

That is my first reason to thank DeepSeek. Thank you for putting a fire under the bellies of western incumbents to become more efficient.

2) Thank you, DeepSeek, for making SaaS and the app layer exciting again.

As for my second reason: thank you DeepSeek for making the app layer and SaaS exciting again! In the app layer, it’s only easier and cheaper now to build companies. We’re already seeing a bunch of companies add DeepSeek’s models into their offerings (Perplexity being one of them and the company I work for is too).

Why is this exciting? Well, I’m a builder, so in startup land this makes it cheaper to run companies using AI and LLMs. Having spoken to dozens of founders, we’re all excited for this. I spend the other half of my time thinking about investments and investing in growth stage companies. Using and running cheaper models improves gross margins which flows down to profitability. As a result, we’ll see improved valuations in SaaS companies. Without getting into the finance (interesting for me, but might be boring to you), SaaS companies are usually valued based on the Rule of 40. Nowadays, it’s more the Growth-Adjusted Rule of 40, but I don’t feel like going there (lol). Companies are valued based on a mix of their revenue growth and profitability, so the more profitable they are, the higher the implied valuations.

I wrote a couple months ago about the revival for software and SaaS, and I think we’re at an inflection point where we’re truly seeing the revival. The successful ServiceTitan IPO reopening the IPO markets (Sailpoint being the first tech company to file for an S-1 this year), or DeepSeek showing us possibilities for margin expansion. Either way, this is a super exciting time for SaaS and the app layer!

3) Thank you, DeepSeek, for bringing back the fundamentals.

Everyone made fun of GPT-wrappers, but here we are now where we’ve flipped the script. We can still make the argument that GPT-wrappers are fragile and LLMs are where the true edge lies — sure. However, my belief is that we’re back to the good old days of distributions, network effects, and a good product for a company to succeed. ChatGPT didn’t win because it had the best model. It won because of it’s sleek UI/UX, dead simplicity, and because of that, it will stay as the winner.

“Frontier models are becoming commodities. User habits aren't. While everyone obsesses over the next architecture breakthrough, the real game is being played in the interface layer. The moat isn't in the model - it's in being the tool people reach for without thinking” (@gregisenberg on X).

When founders and builders ask me — how can I become differentiated? My answer will just be: go back to the basics, build something that your users will use, build a brand, create network effects, have a great UI/UX experience, be considerable close to your customers. It’s back to the basics baby :)

I remember listening to a talk by the Education GTM lead from OpenAI in which I asked her if my “GPT-wrapper” startup will succeed. In a nicer way, she basically said that we’ll get crushed. I don’t think so now.

DeepSeek — thank you for your hard work.

Sincerely,

Allen’s thoughts at 2AM