DeepSeek won't kill OpenAI, but it does show us what's settled – and what isn't

The end of 2024 saw consensus build around the slowing pace of AI improvement, but don't confuse that with a steady state on where value lies in the AI stack

The financial markets are currently having a very broad and fairly deep reaction to the technically impressive debut of DeepSeek’s V3 and R1, and Janus-Pro-7B models – a lot of which is showing up as a significant sell-off of US tech stocks, all the way up and down the stack. Some are saying this undermines not only all of OpenAI’s value premise, but also the recent funds of the venture firms that put money into model companies altogether.

This is definitely an overreaction: OpenAI’s value won’t go to zero, and NVIDIA’s wedge as the anchor of the AI revolution isn’t dissolved to dust.

In my last post for this newsletter, I talked about how the need for AI compute is just going to keep scaling (possibly faster) because of the growing need for inference-time compute resources. I briefly addressed DeepSeek there, but it’s worth expanding why I don’t think R1 or its amazing efficiency is an extinction-level event for anyone else out there building at the foundational level.

What DeepSeek’s impressive launches do show is that 1) AI and the distribution of its value across the chain are far from a settled question, and that 2) a focus on fundamentals continues to be the key way for investors and builders to suss out and make bets on the allocation of said value.

In the era of highly attenuated attention spans, it’s easy to forget that ChatGPT launched in 2022. A little over two years just isn’t long enough for a nascent and particularly explosive technology like generative AI to achieve any kind of reliable stasis.

Because markets and funders need to try to seek out some kind of stability against which to place bets, and because there was initially an immense amount of upside in just continuing to iterate on what had already been achieved in a linear direction, it started to look like gen AI was settling into a set of comfortable ‘knowns.’ All along, however, some of the smartest people in the room were cautioning that this likely was not, and could not be the end state with respect to even the fundamentals of AI technology. Meta’s Yann LeCun, for instance, has long held that some of the key assumptions around LLMs and transformer-based technologies just won’t hold over time for how AI is developed and deployed.

DeepSeek is impressive because it’s a master class in escaping constraints through non-linear thinking and orthogonal approaches to problem solving. But the technical improvements and advantages that it incorporates are not locked to this one company or this one effort.

Companies like OpenAI, Anthropic and others (as well as their open source equivalents like Meta’s Llama) can and will incorporate these learnings into their own approaches, and the result will be massive resource savings along existing parts of the technical stack. But these savings won’t just drive costs for these services down to the point that well-capitalized companies collapse under their own weight and nimble ones offer the same services at cut-rate prices.

We already know that one way in which the AI landscape is shifting is that the importance of having a well-resourced training facility is being replaced with the importance of inference-time compute. DeepSeek-R1 and what it reveals to others in the space is just another way to shift resource allocation to that new and exciting part of the stack, and to pursuing other, complimentary technologies that are equally demanding at the data center but that can also fundamentally change the quality and nature of output at the level of the model supplier.

This reveals that for venture investors, the right approach remains to look for businesses that are focused what offers fundamental value to their customers, and what addresses their specific pain points. I still don’t think the foundational layer becomes ‘commodity’ in the way that bears would have you believe, but it reinforces that the best bets for most will be around companies and founders that understand the core problems they aim to solve, and that are nimble, smart and flexible enough to adopt the best of the base layer even as it shifts and transmutes underneath them.