In late January, two words caught the internet by storm: DeepSeek. The release of the R1 model by a small Chinese hedge fund and research organization sparked a frenzy of speculation, with many dubbing it a “Sputnik moment” for artificial intelligence. But as Martin Casado, General Partner at a16z, and Steven Sinofsky, former Microsoft executive and current a16z board partner, discussed on a recent episode of the a16z Podcast, the significance of DeepSeek extends far beyond the model itself.
“R1 comes out and it looks pretty good,” Casado said. “And then all of a sudden, they’re saying, well, if you can do it just as cheap, is this going to actually drive the next wave? And so there’s a lot of buildup to O1, which led to the R1 hype.”
But it’s not just about performance or cost. DeepSeek’s release came with a permissive MIT open-source license and reasoning traces that enable smaller model training through distillation. This, the guests argued, has the potential for much wider proliferation and adoption of AI technologies.
Sinofsky drew parallels to the early days of the internet, highlighting the tendency for industry leaders to focus on the wrong layer for monetization.
The Internet is such a great example because there’s no way this doesn’t play out like the Internet, and what we saw was for a while, building one app seemed like a crazy thing because you had to own Windows and you had to own Office. But then a new app came along that didn’t own any of those, and it was Search.
The conversation also touched on the differences in capital investment and financing between the current AI wave and the internet era. While the fiber buildout was driven by investors seeking internet exposure, today’s AI landscape is backed by cash-rich tech giants like cloud providers and Nvidia. This, the guests suggested, reduces the risk of a glut and crash like the one seen during the fiber bubble.
Looking ahead, Casado and Sinofsky emphasized the importance of scale-out and commoditization of models, with a shift towards endpoint computing and specialized models. They also called for a reevaluation of benchmarks and metrics in the AI era, moving away from a focus on parameters and coding tests and towards application-specific measures like truth in research.
Perhaps most importantly, the DeepSeek release serves as a wake-up call for AI policy in the United States. “For me, actually, the biggest aha of DeepSeek is nothing we’ve talked about right now,” Casado said. “The biggest aha of DeepSeek is how blind our policies have been around AI. They’ve been so wrong-headed.”
Rather than imposing export controls and limiting domestic labs, the guests argued for increased funding and investment in AI research. “What we should be doing is funding and investing in our research labs,” Casado emphasized. “And we should be going as fast as we can. And it really is the AI race, just like we went through the space race. And we need to win.”
Despite the hype and speculation surrounding DeepSeek, Casado and Sinofsky remain optimistic about the opportunities for established players and the overall growth of the AI industry. With a massive TAM expansion and room for many participants, they encouraged frontier labs to build apps for feedback and competition.
“It’s a good reminder that there are always pockets of people innovating,” Sinofsky noted. “Worldcom and AT&T did not predict that the Internet was going to come out of universities.”
As the AI landscape continues to evolve at a breakneck pace, the DeepSeek moment serves as a reminder of the global nature of innovation and the importance of embracing the transformative potential of this technology. With the right policies and investments, the United States can lead the charge in the AI race – but it will require a fundamental shift in mindset and a willingness to learn from the lessons of the internet era.
