Why DeepSeek had to be open-source (and why it won't defeat OpenAI)

By now, you’ve heard of DeepSeek. It’s the Chinese AI lab that trained R1, an open-source reasoning model as good as OpenAI’s o1, but trained on inferior hardware for a fraction of the price.

‍

They achieved this with novel training methods that are more efficient than the ones OpenAI, Anthropic and other well-funded competitors use to train their proprietary models. But why would they open-source it?

‍

On the surface, it goes against every business textbook you’ve read: If you innovate to build a market-leading product at a fraction of the cost, you should exploit that advantage. Coca-Cola doesn’t open-source its recipe, right?

‍

Not in the world of LLMs. I believe DeepSeek almost had to open-source its models—and that open-source models will become more and more dominant as time goes on.

‍
Why DeepSeek had to go open-source

‍
DeepSeek is in a unique position. It’s a Chinese company, which probably makes businesses feel uneasy about building with them, especially when you start to deal with customer data—and even more so when you want to be HIPAA compliant or SOC2-certified.

‍

A Chinese AI API would likely receive skepticism in the West. But an open-source model instantly builds trust. You have full control if you self-host or use an AI vendor like Together AI.

‍

To gain a foothold in Western markets, DeepSeek had to open-source its models. But that’s not just an economic decision. But it’s not just a political decision. I recently heard the quote that “Open-source is not just a technological behavior, it’s also a cultural one”.

‍

And open-source companies (at least in the beginning) have to do more with less. It’s precisely because DeepSeek has to deal with export control on cutting-edge chips like Nvidia H100s and GB10s that they had to find more efficient ways of training models.

‍

OpenAI, Meta, Google etc. have billions of dollars, massive compute resources and world-class distribution. They don’t need to find a more efficient way to train models when their expensive solution is the only one. In fact, making it easier and cheaper to build LLMs would erode their advantages!

‍

Now that has changed.

‍
Models are getting commoditized

‍
It feels like a new GPT-4-level LLM gets released every week. In the AI apps I use, I can’t tell if I’m using LLaMa, GPT, Claude or Mistral models. They’re pretty equal in performance both in my personal experience and in benchmarks.

‍

OpenAI is still the leader. They were the first to release a reasoning model and the first to release GPT-4. But models are getting commoditized—and it’s worth asking whether it’s worth paying the premium the OpenAI API charges compared to open-source models.

‍

DeepSeek might be the starkest example of this. Compare $60 per million output tokens for OpenAI o1 to $7 per million output tokens on Together AI for DeepSeek R1.

‍

If your end user doesn’t know the difference, why would you pay that much more? This is especially important in infrastructure.

‍
In infrastructure, open-source wins (eventually)

‍
There’s often a tradeoff with using open-source and proprietary software: Open-source is cheaper and more customizable, but ties up more resources (and requires technical knowledge) because you have to maintain it yourself. Proprietary costs more, but offers a smoother (if more rigid) experience.

For many product categories, this tradeoff is not worth making for most companies. You don’t want to lose your knowledge base because your self-hosted Notion alternative made an error.

‍

But infrastructure is always custom. It always requires work from you. Even a proprietary Oracle database requires a ton of work to set up and maintain. This is why open-source databases are more and more popular.

‍

The advantage of proprietary software (No maintenance, no technical knowledge required, etc.) is much lower for infrastructure. It’s actually the opposite: The more technical a product, the better it is for the user (engineers) to work with open-source because they can audit the codebase.

‍

This is also why we’re building Lago as an open-source company. We know billing gets complex whether you build your own or use a vendor, so the engineers prefer to work with Lago.

‍

The same is true for LLMs. To build any useful product, you’ll be doing a lot of custom prompting and engineering anyway, so you may as well use DeepSeek’s R1 over OpenAI’s o1.

‍

This is why there are a lot of successful open-source infrastructure companies and almost no successful open source consumer companies.

‍

Does that mean proprietary AI is done? No.

‍
OpenAI is far from over

‍
There’s a lot of talk about how OpenAI will be obsolete because of DeepSeek R1 or other open-source models. But that’s not true. First, OpenAI has always been first to market, both with LLMs like GPT-4 and reasoning models like o1.

‍

Without OpenAI’s models, DeepSeek R1 and many other models wouldn’t exist (because of LLM distillation). This does beg the question of whether it’s still worth it to build new frontier models if you provide the breakthrough and someone else ships something similar for much cheaper.

‍

But R1 might also wake up the well-funded incumbents and force them to find more efficient methods—and who knows what they can build when they have both efficiency and all the resources in the world.

‍