Zuckerberg Weighs In on DeepSeek and AI Spending

In one of the most high-profile reactions since it rocked the artificial intelligence (AI) world earlier this week, Meta CEO Mark Zuckerberg on Wednesday (Jan. 29) lauded the innovations of DeepSeek’s lower-cost AI model during the company’s Q4 earnings call. But he said it is too soon to tell whether that means it’s time to reduce AI infrastructure spending.

“It’s probably too early to really have a strong opinion on what this means for the trajectory around infrastructure and capex,” he said during a call with analysts to discuss Q4.

That’s because DeepSeek’s innovations centered on pre-training, or initial training, of foundation models, when the U.S. AI industry is shifting toward AI processing, or inferencing, as reasoning AI models rise in popularity.

Zuckerberg added that they are still digesting DeepSeek’s innovations and could implement some of them in its systems.

DeepSeek’s release of a foundation model that cost $5.6 million to train — and performs at par with models from OpenAI, Google, Meta and Anthropic — shook Silicon Valley and Wall Street. On Monday (Jan. 27), Nvidia lost $589 billion in market cap and other chipmakers lost ground as well due to concerns of lower chip demand. DeepSeek was able to train with 2,000 slower Nvidia GPUs compared to tens of thousands at U.S. AI companies.

AI Infrastructure

Zuckerberg also pointed out that his social media platforms serve billions of people who will be increasingly using Meta AI. The AI assistant is being incorporated throughout Facebook, WhatsApp, Instagram, Messenger and Threads, as well as Meta’s AR/VR headsets and smart glasses. Their activities will be needing AI computing power.

“I would bet that the ability to build out that kind of infrastructure is going to be a major advantage” in the future, Zuckerberg added.

But Zuckerberg was sure about one thing: Open source is the way to go. DeepSeek’s innovations can be scrutinized and revised because its models are open source. Meta has pursued an open-source strategy with its Llama family of large language models (LLMs). It is the open-source community’s most powerful foundation model that can be freely used and revised, for most.

When more people build using an open-source AI model, it more quickly becomes the global standard, Zuckerberg contended. China has been contributing a number of open-source AI models for free public use.

“There’s going to be an open-source standard globally,” he said. “For our own national advantage, it’s important that it’s an American standard.”

Llama 4 and AI Engineering Agent

Zuckerberg also provided an update on the next iteration of its Llama family of AI models: Llama 4. He said Llama 4-mini is nearly done with pre-training, and its larger reasoning models “are looking good, too.”

He added that Llama 4 will be natively multimodal, meaning it will be built from the ground up able to interact not just with text but also potentially with images, video and audio. In contrast, some LLMs have these capabilities bolted on after the fact.

Llama 4 will also have agentic capabilities, Zuckerberg said. That means it not only can provide information to the user, but also complete the tasks it is given autonomously.

Zuckerberg added that it is “possible” Meta would unveil an “AI engineering agent” in 2025 that will have the coding and problem-solving capabilities of a mid-level AI engineer. “This is going to be a profound milestone” and over time, a potentially large market, he added.

Asked when Meta expects to monetize AI agents, Zuckerberg said he is still focused on scaling Meta AI throughout its platforms, therefore business opportunities with AI agents remain “outside of 2025.” He added that Meta usually waits until a product scales to a billion users before thinking of ways to make money from the volume of users.

An analyst also asked about Meta’s Jan. 7 announcement that it would end its third-party fact-checking program and instead let social media users pin “community notes” to posts they feel need clarification or context. Zuckerberg said it wasn’t true that “we somehow don’t care about adding context to our platform.”

He said X’s method of using community notes was just more effective than what Meta was doing. Moreover, Meta hasn’t seen any notable impact from the shift, according to another executive on the call.

Earnings Beat Consensus

In the fourth quarter of 2024, Meta reported revenue of $48.4 billion, beating consensus estimates of $47 billion, and up 21% from $40.1 billion in the like quarter a year earlier.

Net income came to $20.8 billion, or $8.02 per share, which beat the consensus estimate of $6.76 per share. In the fourth quarter of 2023, Meta earned $14 billion, or $5.33 per share. Earnings per share rose by 50% year over year.

Fourth-quarter earnings included a $1.55 billion positive impact due to smaller accrued losses in legal proceedings.

Reality Labs, which is Meta’s metaverse division, expanded its loss in the quarter to $4.9 billion from $4.6 billion a year ago.

Operating margin expanded to 48% in the quarter from 41% year over year, and up from 20% in the fourth quarter of 2022. Free cash flow fell to $13.2 billion in the quarter from $15.5 billion in the third quarter of 2024.

In its earnings report, Meta said total expenses should be $114 billion to $119 billion — mainly due to infrastructure costs.

Last week, Zuckerberg said Meta expects to spend $60 billion to $65 billion in capital expenditures. That’s up from the $39.2 billion it spent in capex last year. Zuckerberg revealed his spending plans two days after OpenAI, SoftBank and Oracle unveiled a $500 billion project to build AI data centers.

Looking ahead, Meta expects first-quarter 2025 revenue to range from $39.5 billion to $41.8 billion, reflecting an 8% to 15% increase without accounting for foreign exchange fluctuations.

Shares of Meta rose 4.8% in after-hours trading. The stock has been on a tear. Last year, Meta shares rose by 68% to close the year at $591.24. Thus far this year, the stock is up 15.5%.

The post Zuckerberg Weighs In on DeepSeek and AI Spending appeared first on PYMNTS.com.