An open-source model from a Chinese AI lab tanked the markets today - shedding nearly a trillion dollars from the market caps of leading AI companies and shattering a narrative that has defined the AI landscape for too long. We think that's a good thing.

Last week the Chinese AI lab DeepSeek introduced R1, a so-called reasoning model which uses test-time compute to answer complex questions using chains-of-thought that approximate multi-step reasoning. The model - which was open sourced and made available through a consumer app - was accompanied by a 22-page technical report that showed R1 matching the performance of OpenAI's leading model (o1) despite being trained on a small fraction (2-3%) of the compute. More importantly, early users have found that R1 is up to 30x cheaper to run on a per-token basis; which could add up to a ten-fold decrease in inference costs which have exploded with the new generation of models.

They were able to do this - with a small team - through a number of clever tricks and meaningful research breakthroughs, most notably end-to-end reinforcement learning and low-level hardware optimization which allowed them to get the most out of their limited stockpile of high-end chips, which the US has restricted since 2022.

While the model is impressive, the sell off was clearly a response to a deeper realization: that if compute is no longer needed to build state-of-the-art models, then capital is not the competitive advantage it once seemed.

The idea of AI as an arms race for compute and capital has become pervasive since the launch of GPT-3 and has singularly shaped the AI landscape: from research, to venture capital funding to the historic rise of semiconductor companies that took the biggest hit today. What began with an observation about empirical “scaling laws” for large language models has become embedded into the origin stories of OpenAI and Anthropic and has been used to justify everything from corporate restructuring, to billion dollar seed rounds and the half a trillion dollar infrastructure investment announced last week.

While scaling laws did drive much of the early progress in LLMs, there were signs as early as 2023 that they were beginning to plateau. Not surprisingly the loudest voices pushing the arms race (and pushing back against the scaling wall) narrative were those that stood the most to benefit from it - the biggest cloud providers, chip makers and best funded AI labs with access to the deepest pools of capital, compute and talent.

But the more damaging effect of this narrative is that it has narrowed the focus of the field and embedded a set of assumptions that have made it hard to pursue ideas outside of the orthodoxy and cut out critical institutions.

We’ve seen this most clearly in AI research, where the center of gravity has shifted from academic to industrial labs, which have lured away top researchers with higher salaries and greater access to compute, but are all-in on a small number of methods and architectures. It’s also affecting startups, where highly pedigreed founders are leaving AI labs to start companies (whether or not they have an idea) and raising massive rounds out of the gate before writing their first line of code: often pursuing the most capital intensive ideas instead of looking for market opportunities. All of this distorted capital markets - from early stage venture capital to large cap public investors - who have concentrated capital in the “leading” hardware and foundation model companies while staving off resources from anything that isn’t clearly a winner in an arms race defined world.

Over the past three years, this narrative has become so dominant and self-reinforcing that it could only have been taken down by a company operating outside of the system in an environment with the kind of constraints that have always encouraged disruptive innovation.

The collapse of the dominant narrative is good news for AI researchers building foundation models because it shows that small teams with novel ideas can advance the frontier. It’s good for the AI research ecosystem because it allows room for a variety of new directions to be explored and hopefully a more diverse set of models, built on different assumptions, to mature.

We’re already seeing this play out in the rise of non-transformer based architectures - like state-based models (Mamba-3B), liquid neural networks (LFM-40B) and evolutionary models (Transformer²) which have seen an influx of funding since concerns around a ‘scaling wall’ first surfaced last summer. It has also driven interest in domain specific foundation models in areas like molecular biology (Boltz-1, Evo), material science (Orb) and physical simulation (Genesis) where scaling laws have never been as pronounced. It’s no surprise that many of these models have their roots in academic labs.

This is also good news for early-stage founders, because it will free up capital from the biggest players and shift attention away from the most capital intensive ideas. We’re already seeing increasing investor interest in AI applications, which are closer to the end-user and more capital efficient, giving teams an opportunity to iterate and learn.

Competition and variety at the model layer also creates new and better primitives for founders to build on top of – enabling new applications while keeping costs in check. Diversification of components and decentralization of core technologies have always been key ingredients of vibrant startup ecosystems.

Lastly – this is good news for investors, especially early-stage venture capitalists, whose core value proposition and business model was upended by the arms race. In a world where capital is the competitive advantage and everything is moving fast, it’s understandable why founders would skip seed all together and raise large rounds from the deepest pocketed multistage and growth funds. This has been a hallmark of previous technology waves, like the dotcom bubble, which were shaped by similar bigger-is-better stories and like then founders who bought into. We’re already seeing early signs of companies that raised too much, too early being crushed by their preference stacks or cracking under the pressure their piles of cash and early valuations imply. We've seen this play out with companies like  H Company, Stability, Adept, Character.AI and Inflection and many more stories that are still playing out and have yet to be told.

Far from being an existential crisis for AI or the start of another arms race for America, we think the R1 moment is a positive development for the entire AI ecosystem. It’s an opportunity to lean into democratization of a powerful technology and let a hundred (or a hundred million) flowers bloom. For us it’s a chance to get back to basics and focus on what has always moved technology forward – a wide field of brilliant researchers and scrappy entrepreneurs pursuing disruptive and innovative ideas.