摘要：TMTPOST -- Meta Platforms, Inc. plans to maintain its burning cash pace this year even though Chinese artificial intelligence (AI)
TMTPOST -- Meta Platforms, Inc. plans to maintain its burning cash pace this year even though Chinese artificial intelligence (AI) startup DeepSeek released highly cost efficient models, sowing increasing doubts on ongoing huge spending on the technology.
Credit：Shanghai Observer
Meta CEO Mark Zuckerberg in a post on Facebook last Friday announced the company plans to invest $60 billion to $65 billion in capital expenditures (Capex) in 2025 amid massive efforts to ramp up AI development. The planned spending is way above Wall Street projected $51.3 billion of Capex this year. It represented a nearly 58% increase from the annual Capex of around $38 billion expected by analysts, and more than double that of 2023.
Calling 2025 “a defining year for AI”, Zuckerberg expected Meta AI to become a leading assistant serving more than billion people, the open-source model Llama 4 to be the leading state of the art model, and launch of an AI engineer that will start contributing increasing amounts of code to the company R&D efforts.
To power the aforementioned goals, Meta is building a 2GW+ data center that would cover a significant part of Manhattan, and will bring online 1 gigawatt (GW) of compute in 2025, ending the year with more than 1.3 million graphics processing units (GPUs), Zuckerberg said.
Besides the up to $65 billion of Capex for 2025, Zuckerberg said the year will also see Meta growing its AI teams significantly, adding that the company has the capital the continue investing in the years ahead.
Despite the ambitious spending plan, Meta is reported to turn into a panic mode in the face of new challenges from DeepSeek whose AI models deliver performance comparable to leading offerings at a fraction of the cost.
Some leaders of Meta’s AI teams, including AI infrastructure director Mathew Oldham, recently worried new AI made by DeepSeek meant the Facebook owner was falling behind in the AI race, The Information reported at weekend. These leaders are concerned the next generation of Meta’s flagship AI Llama, which the company has suggested would be released this quarter, won’t perform as well as DeepSeek’s model, the U.S. tech news media outlet cited two Meta empolyees. The sources said Meta has set up several war rooms, or specialized groups of researchers, to dissect DeepSeek and use the insights to improve Llama.
It was reported that Meta’s generative AI group and infrastructure team have started four war rooms to learn how DeepSeek works. Two of these war rooms are trying to understand how High-Flyer, the founder and backer of DeepSeek, reduced the cost of training and running DeepSeek as Meta wants to apply those techniques to Llama. A third war room is trying to figure out what data Higher-Flyer might have used to train DeepSeek models, and the fourth one is weighing new techniques for restructuring Meta’s models based on attributes of the DeepSeek peers.
In a post at Team Blind, an anonymous forum-style social media application used by millions of professions to seek and share advice, echoed the reported panic mode inside Meta. A Meta employee said in the post that DeepSeek’s V3 model started the panic mode across the generative AI team, and the “unknown Chinese company with 5.5 million training budget” added insult to injury.
“Engineers are moving frantically to dissect DeepSeek and copy anything and everything we can from it. I’m not even exaggerating,” said the employee. “Management is worried about justifying the massive cost of generative AI organization. How would they face the leadership when every single ‘leader’ of generative AI organization is making more than what it cost to trained DeepSeek V3 entirely, and we haev dozens of such ‘leaders’.”
来源：钛媒体
