By Adam Clark
China's DeepSeek has stirred up the market with its claim of artificial-intelligence models that can match Western rivals at much lower costs. Microsoft and Meta Platforms are respectful of their new rival but have a plan on how to beat it -- and it involves copying its innovations.
Microsoft and Meta disclosed their quarterly earnings Wednesday evening, so analysts took the opportunity to ask management whether DeepSeek had changed the economics of the AI race. The CEOs of both companies avoided the more controversial issues, such as whether the start-up disclosed its full costs or piggybacked on existing models, but were clear its advantage wouldn't last long.
"I think there's a number of novel things that [DeepSeek] did that I think we're still digesting. And there are a number of things that they have advances that we will hope to implement in our systems," Meta CEO Mark Zuckerberg told analysts on an earnings call.
Microsoft CEO Satya Nadella said DeepSeek had some "real innovations" that would become broadly used.
DeepSeek used a number of techniques to both improve the efficiency of training its AI model and reduce the cost of inference, or producing answers. While the widely reported cost of $5.6 million for training its base model has been questioned, what is undeniable is that it has priced access to its latest model at a sharply lower level than rivals such as Microsoft-backed OpenAI. In response, Microsoft has made DeepSeek's R1 AI model available on its Azure cloud computing platform.
DeepSeek released its AI on an open-source basis, meaning other companies can adapt the model for their own use, making it more attractive to Western developers. Adding it to the Azure cloud facilitates usage.
The open-source format also means Microsoft, Meta, and other U.S. companies can incorporate DeepSeek's advances into their own next-generation models. Zuckerberg was upbeat on Wednesday, saying that Llama 4, the coming version of Meta's AI brand, could "become the most advanced and widely used" model.
Meta arguably has an easier task coping with DeepSeek's challenge. It doesn't sell cloud-computing services, so it won't be hurt if companies can produce AI services using less computing power. Lower costs for AI should ultimately boost the use of its advertising tools.
"Meta has already been instrumental in driving down LLM [large language model] prices over the past couple of years, with Llama being one of the cheapest high-performance models on the market currently," wrote Morningstar analyst Malik Khan in a research note.
Khan raised his fair-value estimate on Meta stock to $770 from $590. Meta shares were up 2.3% at $692.00 in morning trading on Thursday.
Microsoft faces a trickier task. On the one hand, cheaper AI could stimulate overall greater use of its cloud-computing services. Microsoft CEO Satya Nadella raised the prospect of greater demand when he posted about Jevons Paradox in the midst of last weekend's excitement over DeepSeek.
However, Microsoft is also backing OpenAI, the developer of ChatGPT and a major customer for its Azure cloud-computing business, as it seeks to build leading AI models. OpenAI's business model is more deeply threatened by commoditization of AI services. The company is investigating whether DeepSeek made unauthorized use of its data output, according to The Wall Street Journal.
"Microsoft is respecting DeepSeek's innovations and offering access to the product, somehow while simultaneously exploring whether DeepSeek might have violated OpenAI's terms and conditions, which is unusual to see in combination," wrote J.P. Morgan analyst Mark Murphy in a research note.
Murphy maintained an Overweight rating and $465 target price on Microsoft stock.
Microsoft shares were down 6.1% at $415.08 in morning trading.
Write to Adam Clark at adam.clark@barrons.com
This content was created by Barron's, which is operated by Dow Jones & Co. Barron's is published independently from Dow Jones Newswires and The Wall Street Journal.
(END) Dow Jones Newswires
January 30, 2025 12:28 ET (17:28 GMT)
Copyright (c) 2025 Dow Jones & Company, Inc.
Comments