DeepSeek Proves It: Open Source is the Secret to Dominating Tech Markets (and Wall Street has it wrong)

I rarely blog or post on social media but I do write a private weekly newsletter for our staff and board and for a couple of weeks in a row have been writing about DeepSeek.  What I didn’t consider is the reaction this week in the press and markets.  Good folks like Ben Thompson, Pat Gellisnger, Tomasz Tunguz have all been commenting. With AI stocks getting pummeled and grave concerns surfacing about the impact of DeepSeek’s R1 model on the future of AI, it’s clear that fears of DeepSeek struck a deep nerve. That huge reaction merits both examination and explanation.

In short, what the markets reacted to was DeepSeek’s ability to build a model that rivaled OpenAI’s sophisticated o1 reasoning model and Anthropic’s Sonnet model for pennies on the dollar on a fraction of the compute. It also appears that DeepSeek did this using well-known techniques. There were no massive algorithmic breakthroughs, just very clever engineering. The team just went back to First Principles, asked basic questions and stacked up existing techniques in novel ways to achieve outsized results using Reinforcement Learning (RL) and various stages of fine-tuning. There’s no magic here — just a very smart reshuffling of the existing cards that produced a more refined and efficient result.

Some observers accused DeepSeek of “free riding” on work done by other large model makers like Meta (Llama) and AliBaba (Qwen). However, this perspective reflects a misunderstanding of how open-source systems function. The foundational principle of open-source innovation is the expectation that others will build upon prior work to drive progress. In the case of DeepSeek, they demonstrated this by distilling their base reasoning model, DeepSeek-R1—an evolution of their earlier open source DeepSeek-V3-Base model —and fine-tuning smaller models from the Llama 3 and Qwen 2.5 series of models using reasoning data generated by their base model. This process highlights how DeepSeek leveraged existing open innovations, not merely for replication, but to demonstrate significant improvements in small language model performance. DeepSeek then re-released those models back to the open source AI community.

Here’s the Big Takeaway. What the news and commentariat are missing is the massive opportunity that DeepSeek has opened for open source and, more broadly, the entire open movement.  Too much of the conversation is framed as U.S. vs China and the race for AI supremacy. Too much of the conversation is framed on the idea that DeepSeek training a powerful model on a fraction of the compute for a fraction of the cost means all the large companies spending big bucks on NVIDIA gear and training will be undercut.

In my (biased) view, open source innovation will win and that this will actually be good for everyone — China, the U.S., Big Tech, European Digital Sovereignty, NVIDIA, and more. Some venture investors who bet on early AI startups that have become features inside of larger AI offerings might get wiped out, but that happens in any technology phase shift. What DeepSeek proves is that we need thousands of eyes on the problem to come up with better solutions to make intelligence as close to free as possible. A small team in China took a fresh look at a problem and came up with a novel approach that reduced the cost of chain-of-thought reasoning by 50x (if DeepSeek’s postings are accurate) and then published a paper fully describing their process, allowing the community to benefit from their learnings. We need MORE of this progress, not less. This is not an arm’s race between the U.S. and China. It is a struggle over open markets between the forces of open and the forces of closed. Governments may think they can control this, but history shows that open technology, once discovered and put in the hands of the community, is like rain. You can’t pause or stop it. Artificially halting scientific development has never worked in any long-run term, and computer science and AI are no different.

In March 2013, the open source world was introduced to a lightweight, standardized way to package and run applications with all their dependencies, ensuring consistency across different environments. Unlike traditional virtual machines, Docker containers used isolated environments on a single underlying operating system to do virtualization. Docker containers were faster, more portable, and more efficient by sharing the host system's kernel while isolating processes. Docker reorganized many existing open source virtualization and container capabilities like cgroups, LXC, namespaces and more. That reorganization, or shuffling of the then-existing cards, changed the game. Virtual machines had worked well for years, but Docker containers were far better for many workloads. The openness of that technology shift helped power a new wave of cloud-native computing adoption.

Another fundamental misunderstanding is that DeepSeek will require less AI infrastructure investment. Yet, there is a boundless appetite for intelligence. We haven’t even scratched the surface and are in the very early stages of tapping into AI-powered applications.

More recent improvements in AI models have shifted the nexus of reasoning from pre-training and post-training enhancements to now test-time compute, allowing models to “reason” through their responses (chain-of-thought). This doesn’t mean we need less compute. It actually means we need more compute, when the inference layer acts more like a human brain — always thinking, reconsidering, tackling multiple tasks at once, and evolving to fill the need for new intelligence activities. This is more like electricity — a commodity. Make it cheaper so more applications are possible (as VC Tomasz Tungus explains neatly here), and more people will use it.

For open source, this opens a massive new frontier. If open source wins in AI and becomes the dominant innovation and development model, then we have an opportunity to reshape the way the world works at fundamental levels. DeepSeek is one example of making reasoning available to a much wider array of users and applications. Open source AI could be a path to deliver true interoperability and standards between applications and application stacks.

AI is the meta-layer upon which we could build a new expectation for interoperability, a new reality that Satya Nadela hinted at when he spoke in multiple forums about how AI could disrupt SaaS apps by allowing organizations to hook up different back-ends and data sources to AI engines. In other words, open source AI gives the world a chance to rewrite the rules to favor open everything, everywhere possible. In this world, power goes to the community and the maintainers.

There is, of course, a lot of nuance around open source and how it works. But through this lens, I believe the lesson of DeepSeek is about the coming AI boom and how it can benefit everyone and drive economic and technological progress far exceeding what markets perceive — if we keep it open.

Don’t believe me?  Linux is now thirty-four years old and a group of researchers at the University of Waterloo working in the open demonstrated, just last week, that “changing 30 lines of code in Linux could cut energy use at some data centers by up to 30 percent”  Guess what OS all those AI workloads run on?

Finally, I will throw out a small prediction for all the “trojan horse” uninformed naysayers.  Another firm or research lab will have a similar model using this method with amazing performance to cost ratio in the next few weeks. Feel free to guess who in the comments.