Meta Dropped $14 Billion on an Open-Source Killer, and It Actually Worked
On April 8, 2026, Meta announced Muse Spark. The company’s first proprietary large language model since restructuring AI operations last year. This breaks with the Llama strategy. Meta had spent more than $14 billion hiring Alexandr Wang and his Scale AI team in June 2025 to make this happen.
The timing is no accident.
The Bet Paid Off—At Least on Day One
Meta’s stock surged nearly 9% after the reveal. That added over $100 billion to market cap in a single day. Investors clearly believe this finally moves Meta from “company spending way too much on AI” to “company with an actual AI product.”
The numbers are staggering. Meta told Wall Street it plans to pour $115 billion to $135 billion into capital expenditures in 2026. Double 2025’s figure. That $14 billion acquisition of Wang’s expertise created Meta Superintelligence Labs as a new elite unit. The question has been whether this spending would ever produce revenue.
Muse Spark is the first answer.
Why Muse Spark Now
Llama 4 launched in April 2025. It flopped. Developers didn’t care. The open-source strategy wasn’t working for Meta’s business model. Zuckerberg shook up the entire AI operation after that.
The Muse series takes a different approach. Each generation validates and builds on the last before scaling up. Muse Spark is small and fast by design. It handles science and math questions. It handles health questions too. But the real focus isn’t developers—it’s consumers.
Arun Chandrasekaran at Gartner called this a shift away from Llama. The proprietary model changes everything about how Meta competes.
What Muse Spark Actually Does
It powers the Meta AI assistant across the app and meta.ai. Two features stand out.
Multimodal perception lets the AI see what you see. Snap a photo of airport snacks, and it ranks them by protein content. Share a medical chart, and it explains what’s happening. This isn’t text-in, text-out anymore.
The parallel subagents feature is weirder. Plan a Florida trip, and three agents work at once—one on the itinerary, one comparing Orlando versus the Keys, one finding kid activities. They all run simultaneously. The answer comes back faster.
Visual coding is the other big one. Prompt it to build a dashboard for party planning or a retro arcade game, and it generates playable code. Share the link with friends. The shopping integration pulls from content across Meta’s apps—styling inspiration, brand stories, things creators are already talking about.
The Real Business Model
Here’s where it gets interesting. Meta has 3 billion monthly users across Facebook, Instagram, and WhatsApp. Andrew Boone at Citizens called this the crown jewel.
OpenAI and Anthropic sell developer access. Meta sells ads. Advertising generated 98% of Meta’s $200 billion in revenue last year. The company wasted $60 billion on the metaverse since 2020. The ad business is the thing that actually works.
Morningstar’s Malik Ahmed Khan put it simply: make ads better. Targeting, engagement, conversions. Advertisers already spend heavily on Meta. If AI improves those returns, they’ll spend more. That’s the business case.
Doris Xin runs AI startup Disarray. She pointed out that Muse Spark benchmarks show strength in image and video processing. That matters for Reels. That’s where the money is.
Developers Get the Shaft
This is the problem with going proprietary. Joseph Ott runs Samu Legal Technologies. He put it bluntly—the only reason he used Llama was fine-tuning. You could customize it for specific use cases. Open-weight models are the foundation for countless specialized AI products.
Muse Spark doesn’t offer that. Meta plans a private API preview for select partners. Maybe paid access later. But developers have plenty of free alternatives now. Chinese models are getting better. Llama itself is still out there. Why pay Meta?
Ott’s question hangs over everything: what makes Muse Spark worth paying for?
Ulrik Stig Hansen from Encord frames this differently. It’s about sovereignty. Meta doesn’t want to depend on OpenAI or Google. One of the few companies with the compute and cash to build its own foundation models. They want to be an AI company, not an AI customer.
“We just gave you a state-of-the-art frontier model,” Boone said. “What are you going to do with it?”
That’s the question now.
What Happens Next
The upgraded Meta AI is rolling out in the US now. Coming weeks bring it to more countries and all Meta’s platforms. Instagram, Facebook, Messenger, WhatsApp. AI glasses too—that’s where the visual perception really shines.
The roadmap promises richer answers. Reels, photos, posts woven directly into responses. Credit back to creators. Meta also emphasized safety and privacy safeguards.
The bet is massive. $14 billion for Wang, $115 billion in capex this year, a complete strategic pivot from open source to proprietary. Muse Spark works. The stock jump proves investors approve.
But the real test starts now. Can Meta actually make money from this? Or will this be another metaverse—another expensive distraction from the ad business that actually prints cash?
Wang built something impressive. The model benchmarks are solid. The integration across Meta’s products is happening. The pieces are there.
The execution? That’s on Zuckerberg. Again.
Comments