Speaker 1: First, we were reeling from the success of China's AI upstart DeepSeek. Now TikTok's parent company, ByteDance, has just entered the AI race in a big way, escalating competition between the U.S. and China. Deirdre Bosa has more in today's Tech Check. Deirdre?

Speaker 2: Kelly, I feel a little bit like a broken record this week, but the momentum from Chinese AI players has been unrelenting. ByteDance, as you mentioned, now claims that its model, Duobao 1.5 Pro, outperforms OpenAI's latest reasoning model products. And just days before you mentioned this, an open-source model out of Chinese AI lab DeepSeek was released that rivals OpenAI's O1 on several third-party performance benchmarks. But those two have something else besides performance in common that makes it starkly different than our American ones, and that is cost. They were many times cheaper to build and are many times cheaper to access. So developers are really interested in these models. Big American players too at Davos are

Speaker 3: taking note. We should take the development out of China very, very seriously. What we've found is that DeepSeek, which is the leading Chinese AI lab, their model is actually the top performing or roughly on par with the best American models. If the United States can't lead in this technology,

Speaker 2: we're going to be in a very bad place geopolitically. Each of them, extremely influential CEOs working in AI, they mentioned DeepSeek by name. And so we've been digging into this lab that's very mysterious, and also the breakthroughs that has everyone talking. We'll be publishing our long-form tech check take tomorrow morning, diving into what it all means. So don't miss that, Kelly.

Speaker 1: So going back to what TikTok is up to, what's most significant about it, do you think?

Speaker 2: The cost. I mean, these models coming out of China are just built at a fraction of the price. When you think about open AI that's spending $5 billion a year, burning through billions of dollars a year, these models, the DeepSeek one, for example, they say they built it for less than $6 million. ByteDance as well shows that it was built and you can access it at much, much lower prices. So this really turns on sort of this truth that we have thought about generative AI for the last few years, that you need hundreds of millions of dollars to develop bigger and better models. What the Chinese labs and companies are doing is they're going straight to the frontier. They're building with sort of infrastructure and outputs that are already out there, built in many cases by American companies and startups, and they're improving on it. They're innovating on it and producing models that are just as good in some cases at a fraction of the cost.

Speaker 1: So maybe we don't need to buy the nuclear stocks after all. Maybe we don't need a million GPUs down in Abilene. It's a good question. Interesting. And it does raise a lot of questions even about

Speaker 2: Project Stargate, right? What kind of infrastructure are we building here? Is it for pre-training, like we've had for the last two years, or is it for reasoning and inference, which is a different

Speaker 1: cost proposition? Exactly. And you almost wonder if that's why Microsoft is evolving its partnership with OpenAI as well, as these changes happen. Deirdre, thanks. Deirdre Bosa, always appreciate it.