Our Take on DeepSeek
You’d be hard pressed not to have heard something about DeepSeek. But, if you’re just coming off a self-imposed news diet, an extended stay at a monastic retreat, or something similar: DeepSeek is a Chinese company that developed an eponymous family of open source large language models and an accompanying chatbot. Beyond their provenance, what distinguishes these models is that they perform similarly to the most current models from OpenAI, Claude, Meta, and Google—while operating at a fraction of the cost (20 to 50 times lower). And, they were reportedly trained for $5-$6 million using nVidia a100 GPUs1, which were state-of-the-art over four years ago. By comparison, it is widely believed that the current generation of frontier models cost over $1 billion each to train. Even if we take DeepSeek’s claims with a considerable pinch of salt, and say for the sake of argument that they understated the true cost by 20x, that would still be far cheaper than its competitors2.
There are many facets of this development that have caused sturm und drang in the tech industry: the strong performance of the models, the fact that they are being given away, the pedigree of their creators—who do not come from backgrounds in deep learning, suspicions of IP infringement in their training, and, most significantly, that they were developed in spite of export restrictions3 on state-of-the-art hardware and so used quite outdated GPUs. Yet, they were trained at a cost potentially three orders of magnitude lower than competitors and remain remarkably inexpensive to operate.
From Rally’s investment perspective, DeepSeek’s emergence reinforces our thesis in this rapidly-moving space. We focus on investing higher up the stack, where innovation in software platforms and AI-driven applications is solving valuable problems in ways that disrupt long-held practices and older patterns of automation.
We have enormous respect for the innovation coming from chip designers, data center builders, and everyone else working on the infrastructure of AI. However, it is our belief that demand for ever-faster hardware and ever-larger data centers has been vulnerable to innovations in deep learning model design. From our seat, the race to create more powerful models has prioritized model performance over cost, enabled by the ready availability of capital.
Similarly, while we are deeply impressed by the work of standalone frontier model companies like OpenAI and Claude, their high cost of operations, the short lifespan of each generation of model, and their enormous capital intensity make them a poor fit for our investment approach.
Therefore, we see DeepSeek as amplifying our thesis: the best opportunities for early-stage investing will be in companies using these advances in deep learning. If anything, DeepSeek itself gives startups state-of-the-art AI they can completely control and operate at low cost, without being beholden to a licensor that is itself losing billions a year.
Because DeepSeek is open source, its innovations are available to other model makers. We look forward to a richer and more compelling open source model landscape, with even greater benefits to founders and builders who have the insight to use them in compelling ways to solve long-standing problems.
1. Reuters has a decent explainer.
2. A good overview of the three major performance improvements in DeepSeek comes from Ege Erdil at the Epoch AI research lab.
3. Accusations have been made that DeepSeek had access to current-generation GPUs which were smuggled to China in violation of export restrictions, but to date no compelling evidence has been shown to back up that claim.
