By Adam Levine

On Thursday The Wall Street Journal reported that Meta Platform's largest and most ambitious AI language model, Llama 4 Behemoth, would be delayed for the second time. Meta stock fell on the news, finishing Thursday down 2.4%.

In delaying its next generation of language models, Meta has good company in AI start-ups OpenAI and Anthropic.

OpenAI's ChatGPT 3.5 came out in November 2022 and shocked the world with its capabilities; it was the spark that set the AI world ablaze. Just a few months later, it released the much larger GPT-4, a big leap in capabilities.

Observers projected this rate of progress into the future, leading to bold predictions of superintelligent computers within just a few years.

But it's been two years since GPT-4's release, and GPT-5 is nowhere to be found. Anthropic announced a new Claude model, Opus, last June. It, too, isn't ready.

Meta's Behemoth is set to be by far the largest AI model Meta has produced, and it looks like it is running into the same walls that OpenAI and Anthropic hit. The report from the Wall Street Journal said that Meta staffers were questioning "whether improvements over prior versions are significant enough to justify public release."

A Meta spokesman declined to comment to the Journal. The company didn't respond to a request for comment from Barron's.

With new large-language models bumping into technical roadblocks, the AI model makers have had to find other ways to make their chatbots perform better. New "reasoning" models, like OpenAI o3 and DeepSeek r1, have long internal dialogues with themselves and work through problems step-by-step, checking itself as it goes along. They work well, but they're very slow.

For example, Barron's asked OpenAI o3, "Explain how much Meta is spending on AI data centers, and how much their pre-training runs for Llama 4 Behemoth cost." o3 spent over four minutes talking to itself before spitting about a thousand-word answer. Meanwhile, all that "thinking" costs money; for developers, o3 is five times more expensive than OpenAI's best non-reasoning model.

New, larger models offer the hope of an o3-type response, without four minutes of thinking.

The disappointment from Meta investors on Thursday is compounded by the massive amounts of money the company is spending to make a new model happen. After spending $14 billion in capital expenditures in the first quarter, Meta projected it would spend another $50 billion to $58 billion on AI data centers in the remaining nine months of 2025. Training runs to create these very large models are expensive; Meta's research and development costs for the first quarter were up by 22% from the year before.

Investors are still wondering when the payoff for all this investment comes. Advertising is the likely path for Meta to monetize AI, given that 98% of its total 2024 revenue came from advertising.

But it's not on the near horizon. "Our initial focus for Meta AI is really about building a great consumer experience, and that's frankly where all of our energies are kind of directed to right now," Chief Financial Officer Susan Li said on Meta's fourth quarter earnings call. "There will, I think, be pretty clear monetization opportunities here over time."

More recently, CEO Mark Zuckerberg sounded optimistic about Meta's return on investment when it comes to AI spending. After laying out a range of monetization possibilities, he said, "We don't need to succeed in all of these areas to have a good ROI. But if we do, then I think that we will be wildly happy with the investments that we are making."

Shares of Meta were down 1.4% in mid-day trading on Friday.

Write to Adam Levine at adam.levine@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

May 16, 2025 12:55 ET (16:55 GMT)

Copyright (c) 2025 Dow Jones & Company, Inc.