Speaker 1: All right, now it is time for Tech Check. And today, let's talk about AI, but not in the way that you all think, because there is a new model that has all of the Valley buzzing. And it does not come from OpenAI, or Meta, or Google, or any of those names. Dear Dribosa, this comes from a rather unusual source.

Speaker 2: It also comes from China, Brian. It's called DeepSeek. And here's why it matters. It took Google and OpenAI years, and billions and billions of dollars to build the latest AI large language models. But now, a Chinese research lab has built a competitive model in just two months with dumbed down GPUs for less than, get this, $6 million. Not billion, $6 million. Now, DeepSeek looks and acts just like ChatGPT. I've been trying it out this morning, and when I asked, what model are you? It answered, I'm an AI language model created by OpenAI, specifically based on the GPT-4 architecture, suggesting that it was trained on ChatGPT outputs, which, leaving aside terms of service violations, means that entirely new state-of-the-art models, they can be built on what is already out there. In other words, OpenAI's moat may be shrinking. If a model like DeepSeek can emerge with competitive performance, minimal cost, and reliance on existing outputs, it signals a rapidly shrinking barrier to entry in AI development, challenging the current dominance of industry leaders like OpenAI and Google and Anthropic and Meta, by the way. There's also key geopolitical implications in this for the AI race. DeepSeek is backed by a Chinese quant trading firm, High Flyer Capital Management, and it used NVIDIA H800s, a lower performance version of the H100 chips that are cheaper, more available, and tailored for restricted markets like China. So they were able to sort of go around the H100s that everyone seems to be looking for here in the West. Now, in fact, it costs DeepSeek just $5.5 million to train it versus hundreds of millions of dollars for Meta's latest Lama model, and billions of dollars for GPT and Gemini models. This all raises an important question, Brian, for investors as the AI trade evolves and technological progress stalls, is training frontier models still a good investment? DeepSeek should make investors look twice.

Speaker 1: Why the difference in price? Like what am I getting for 5.5 million versus a billion?

Speaker 2: That's the thing. You're basically getting a model that's as good as the frontier models that OpenAI and Lama have created. So there's third parties that check key benchmarks. These are key capabilities of the different large language models, and they found that the ones from DeepSeek, which was presumably built and trained on public OpenAI chat GPT data, is as good or better in some cases than Meta's or OpenAI. So it's really remarkable. You get just as much bang for your buck for a lot less, like $5.5 million versus billions and billions of dollars that OpenAI has shelled out. So this is really a remarkable development that has huge implications for the race. I spoke to one source who questioned some of the benchmarks, but nonetheless said this is a massive development and could make investors question, do you need the high-end GPUs if you can get just as much out of the H800s, the dumbed down version?

Speaker 1: If that's the case, it changes a lot of the game, including for energy, but this is one thing. To your point, Deidre, we'll see where it all goes. DeepSeek, we're watching you and you, Deidre Bosa. Thank you very much.