Mark Zuckerberg paid $14billion for Alexandr Wang’s AI; Meta’s own engineers still reach for Claude

A year after Mark Zuckerberg spent more than $14 billion to bring Alexandr Wang and a team of Scale AI engineers to Meta, the social media giant has its first proprietary frontier AI model. What it doesn’t have yet is proof that Muse Spark can close the gap on Claude, Gemini, and ChatGPT, or convince investors that the company’s aggressive AI push will be more than just another pricey detour.

Why did Meta abandon open source for Muse Spark?

Over the years, Meta’s identity in AI has been built on Llama, a family of open-heavy models that has positioned the company simply as the industry’s most accessible AI developer. That reputation took a significant hit in April last year, when the release of Llama 4 failed to generate meaningful excitement among developers, prompting Zuckerberg to fundamentally reconsider the company’s direction.

Two months later, Zuckerberg announced a $14.3 billion investment to acquire nearly half of Scale AI and, more importantly, bring Wang on board as Meta’s Director of Artificial Intelligence, along with its most senior engineers.

Wang’s first major delivery was Muse Spark, which was released in April of this year and represented Meta’s first step from open source into private edge model territory. Wang has since offered a more careful framing of Meta’s open source commitments, saying the company will continue to publish models it deems “appropriate and safe” to publish while locking down boundary work. When asked if Llama would continue to brand for this effort, Wang sidestepped the question: “We have exciting discussions about branding within the company and nothing to share at this time.”

Why Muse Spark remained registered: biosecurity concern

The decision to keep Muse Spark proprietary was not purely commercial. Wang acknowledged at Bloomberg Tech that internal testing flagged security concerns that made the open version untenable.

“It actually triggered some high-risk areas during early training, especially related to biological risk, but it also increased a number of risks,” Wang told Bloomberg. He added: “This is something the whole industry has seen as models have improved significantly in the last year.”

As part of establishing Meta Superintelligence Labs, Wang updated what the company describes as its advanced AI scaling framework; This is an internal document that outlines how Meta evaluates and mitigates model risks. He suggested that embedding the Muse Spark into Meta’s own products allowed the company to implement safety guardrails that would not have been available when model weights were made public.

How was Muse Spark created and its place in the Meta ecosystem

Instead of targeting third-party developers, Muse Spark is designed to integrate directly into Meta’s core apps like Facebook, Instagram and WhatsApp, as well as AI-powered hardware like Ray-Ban Meta glasses, according to Thomas Randall, an analyst at Info-Tech Research Group. It also supports the standalone Meta AI app and website.

“There will be a lot of these leading model providers that will fundamentally change in many different ways, and Meta needs a consistent, reliable, proprietary model that it has itself,” Randall said. He added that Meta would have been “lost” if Zuckerberg had not opened its wallet for Wang and other high-profile AI employees, calling the move a “strategic restructuring” for the company.

Randall acknowledged Meta didn’t take the “most optimized route” but said he could now see “a vision of what they were trying to accomplish and what Wang was trying to accomplish.”

Why is Muse Spark still following Claude and Gemini?

Despite all these repositionings, Muse Spark has yet to emerge as a credible competitor. The Financial Times reported that Meta employees wanted to test the model for software development tasks and continued to prefer Anthropic’s Claude.

Although Wang was praised for its visual understanding, he acknowledged that the model fell behind its rivals in coding. According to the FT, some insiders are comparing parts of the system to DeepSeek’s latest model, while others note that Muse Spark is based on Llama 4 code and datasets, although Wang has previously described it as “built from scratch.”

Access was also narrow. The model lives primarily inside Meta’s own applications, with a proprietary API offering defined by the FT as limited. A Meta spokesperson said the company was “already testing with some partners and looks forward to launching this month.”

Developer trust issue Meta not resolved

Meta faces a more fundamental challenge beyond security and performance: rebuilding credibility with the developer community it alienated from the Llama 4 disappointment.

“I think the AI ​​community has largely ignored Meta at this point.” CNBC He quoted Rob May, CEO of a startup called Neurometric, which works in the field of token engineering.

May said it’s difficult to gauge how much progress Wang has made at Meta Super-Intelligence Labs, considering the company has only released one AI model so far, calling it a “yawn” among the AI ​​community because the technology isn’t widely accessible. He stated that he was in regular contact with Meta regarding Lama-related issues, but that he could no longer “get them to return messages.”

Andrew Moore, chief executive of enterprise startup Lovelace and former Google Cloud AI chief, argued that Meta’s focus on computational efficiency could still make a meaningful difference.

“If they make proprietary, computationally efficient models, it will be very different from what’s going on in this death match between the big guys,” Moore said, as quoted by CNBC. “They could really benefit from it.”

Moore added that Meta needs to show an advantage somewhere “in terms of cost, latency, or other technical nuances that matter to developers.”

Wall Street is not convinced and addiction to advertising continues

Despite reporting revenue growth of 33 percent in the first quarter of 2025, the fastest growth rate since 2021, Meta’s shares have fallen 18 percent over the past 12 months, making it the worst-performing company among the megacap tech group alongside Microsoft.

The underlying numbers illustrate this challenge. The Wall Street Journal reported that 97.6 percent of Meta’s 2025 revenue will come from advertising, steeper than that of Google, Microsoft or Amazon relative to the size of the company’s planned AI capital spending this year.

To generate revenue outside of advertising, Zuckerberg is currently testing $4-per-month subscription tiers on Instagram, Facebook, and WhatsApp, as well as a $7.99 Meta AI chatbot subscription in select markets.

Analysts at Truist Securities quoted: WSJIt pegs the subscription opportunity at $20 billion annually by 2030, while Deutsche Bank predicts $15.6 billion for next year alone. These are ambitious estimates for a company that failed to generate $5 billion in non-advertising revenue last year.

William Blair analyst Ralph Schackart, who recommends buying the stock, said he wanted to see “concrete evidence of the growing list of new, AI-first products created by Muse Spark, even if monetization is delayed.” That’s what investors are looking for, he said.

“Meta needs to provide more evidence of both adoption and commercialization,” Schackart said.

Pressure on Wang and leadership issue

Wang described the Muse Spark as an “appetizer” for the future and promised more powerful and “bigger models” in the future. However, the AI ​​industry operates at a non-stop pace of launches and updates, and Meta has yet to reach the pace set by OpenAI, Anthropic, and Google.

There are signs of stress on the inside. Meta cut nearly 8,000 jobs in May; Reductions across departments, including teams working in trust and security roles, have raised concerns about risks posed by AI development among people familiar with the matter.

There’s also reportedly tension in the upper echelons of the AI ​​organization as pressure puts pressure on both Wang and former GitHub CEO Nat Friedman, who joined last summer, to turn Muse Spark into meaningful revenue growth.

Andrew Bosworth, Meta’s long-serving technology chief, a 20-year company veteran and close confidant of Zuckerberg, is someone the CEO could turn to for a larger role in AI if newcomers are perceived to be underperforming, according to people with knowledge of the matter. Wang dismissed the reported infighting while speaking on the Core Memory podcast last month, saying: “One of the things that is very important to me is the safety of these models.”

Ultimately, analysts and observers agree that the burden falls on Zuckerberg himself. Yu noted that the CEO’s metadata and virtual reality ambitions have resulted in losses totaling more than $80 billion since late 2020, a track record that makes the AI ​​rollout a tougher sell for investors.

(With input from Bloomberg Tech, Financial Times, Wall Street Journal)