Meta delayed access to flagship neural network. Is the payback on AI spending in jeopardy?

Meta Platforms has several times postponed the release date of the developer interface (API) for its new artificial intelligence model Muse Spark and by early June had not decided on the exact release date, The Wall Street Journal (WSJ) found out. The two-month delay threatens to reinforce Wall Street's doubts about Mark Zuckerberg's company's ability to quickly recoup its multi-billion-dollar spending on AI development, the newspaper points out.

API is the only channel to access closed AI models, which includes Meta's Muse Spark: developers cannot download such a model and run it on their own, but they can use API to connect it to their applications to use the capabilities of neural networks in their products. Usually, companies launch APIs together with a new AI model or shortly after the release, during the period of maximum interest in the novelty, states WSJ. Meta introduced Muse Spark in early April.

Meta's competitors, OpenAI and Anthropic, are already making money by selling access to their APIs: using this tool, customers embed their AI into their own products without having to create AI models from scratch. Meta wanted to release the API at the same time as the launch of Muse Spark in April, but that didn't happen. The release was first pushed back to Ma due to bugs discovered during testing and the need to build additional infrastructure; the launch was then pushed back to June, a source told the newspaper.

The current delays in launching the API are reminiscent of last year's failure: back then, Meta postponed and eventually canceled the launch of the Behemoth AI model because it failed to noticeably improve its capabilities. After that, the company dramatically stepped up hiring, rebuilt its AI teams, and created a new structure, Meta Superintelligence Labs, headed by ex-Scale AI head Alexander Wang, which launched Muse Spark. This neural network, on which Meta's corporate AI bot is now working, was the company's first closed development (before that, it made the code of all its models open source).

According to the company's internal tests, Muse Spark successfully competes with OpenAI and Anthropic products and beats Grok from startup xAI. But until the API is released, these results have only been confirmed by several third-party firms testing the neural network on Meta's servers, WSJ emphasizes.

Investors have long been pressing Mark Zuckerberg to prove that spending on AI will bring the company real money. This year, Meta's capital expenditures may reach $145 billion, and the corporation will spend a significant part of these funds on AI projects, Business Insider wrote. In April, Wall Street reacted with a sell-off on the news that the company's investment spending this year will exceed the initial forecast.

So far, paid services bring Meta barely noticeable income. In the first quarter, "non-advertising" revenue (which includes subscriptions and sales of gadgets like smart glasses and VR helmets) amounted to only $1.29 billion. The corporation earned over $55 billion from advertising during the same period, Bloomberg noted. In an attempt to increase the return on AI outside of the advertising business, Meta last week began selling subscriptions to the Meta AI chatbot.

Meta quotes on the premarket in New York on June 4 are down about 1%, correcting after a jump of more than 4% at the end of yesterday's trading. The reason for the excitement was the announcement of Meta's AI agent for automating everyday business tasks - a signal of the company's readiness to compete with B2B developments OpenAI, Anthropic and Google.

According to S&P, none of the stock analysts advise selling Meta securities - they now have 58 positive ratings (Buy and Overweight) with six neutral (Hold).

This article was AI-translated and verified by a human editor