Mark Zuckerberg spent $14 billion to bring Alexandr Wang in-house last summer. On Wednesday, he finally showed Wall Street what he bought. Meta debuted Muse Spark, the first proprietary AI model out of its newly minted Meta Superintelligence Labs, and in doing so it walked away from the open-source philosophy that had defined the Llama era. The bet is simple and staggering. Zuckerberg is spending like a hyperscaler, hiring like OpenAI, and pivoting like a man who has decided that the next version of Meta is an AI company with a social network attached, not the other way around.
Muse Spark, originally code-named Avocado inside Meta’s labs, is a multimodal model that accepts voice, text, and image inputs. It produces text-only output for now. The company says the model closes a meaningful portion of the performance gap with frontier models from OpenAI, Anthropic, and Google, particularly in multimodal perception, reasoning, health, and agentic tasks. It is already powering Meta AI inside the standalone Meta AI app and the desktop website, and it will roll into Facebook, Instagram, WhatsApp, Messenger, and the Ray-Ban Meta smart glasses in the coming weeks.
THE CLOSED-SOURCE PIVOT NOBODY SAW COMING
For the better part of three years, Meta’s AI story was Llama and open weights. Zuckerberg went on every podcast he could find to argue that open source was both a moral good and a strategic moat. Developers loved it. Regulators tolerated it. Wall Street was mostly confused by it. That story ended on Wednesday. Muse Spark is a closed model. Its code and weights are not being released to the public. Meta says it “hopes” to open-source future versions, which is corporate speak for, we will see how this goes.
This is a direct response to what Zuckerberg has said privately for the past year: that the open-source model was costing Meta the ability to attract and retain the kind of research talent that now chooses OpenAI and Anthropic by default. Pay aside, top AI researchers want to work on frontier systems with real commercial stakes. Giving the weights away for free made that pitch harder. Closing the model fixes that problem. It also aligns Meta’s incentives with the rest of the frontier lab ecosystem for the first time.
The pivot will be read in different ways depending on where you sit. Open-source advocates will call it a betrayal. Enterprise customers will probably call it a relief. Regulators in Brussels, who have been nervously watching open-weight models proliferate, may quietly breathe easier.
THE WANG EFFECT
Alexandr Wang was brought in last summer to run Meta Superintelligence Labs after his Scale AI deal. The price tag was $14 billion for the team and the infrastructure, a number that at the time looked like a panic buy. Meta’s AI efforts had been drifting, Llama 4 had underperformed Meta’s internal benchmarks, and the company was losing ground to OpenAI’s ChatGPT and Google’s Gemini on every measure that mattered to consumers.
Nine months later, the panic buy is starting to look like foresight. Muse Spark is the first shipped product of the Wang era, and Meta says the model’s improved training techniques and rebuilt infrastructure have allowed the company to build smaller models that match the performance of midsize Llama 4 variants “for an order of magnitude less compute.” If that claim holds up under third-party evaluation, it is arguably more important than the model itself. Efficiency is the real battleground in frontier AI right now. Everyone has compute. Nobody has enough compute.
$115 BILLION TO $135 BILLION, AND COUNTING
The number that will keep Meta shareholders up at night is not Muse Spark’s benchmark scores. It is the 2026 capital expenditure guidance. Meta has told investors to expect $115 billion to $135 billion in AI-related capex this year, nearly twice its 2025 total and roughly the market capitalization of Boeing. That is the kind of spend that makes sense only if AI monetization inside Meta’s properties accelerates faster than anyone currently expects, or if the company eventually sells API access at commercial scale.
On that second point, Meta confirmed that a “private API preview” is open to select partners now, with broader paid API access coming later. In plain English: Meta is about to become a competitor to OpenAI, Anthropic, and Google in the enterprise AI developer market. That is a market currently worth tens of billions and growing fast. OpenAI has reportedly crossed $25 billion in annualized revenue and is sketching out an IPO path for late 2026. Anthropic is approaching $19 billion. Meta wants a seat at that table, and Muse Spark is the ticket.
CAN META’S DISTRIBUTION EDGE CLOSE THE GAP?
Meta’s secret weapon has always been distribution. Three billion people open one of its apps every day. Plugging Muse Spark into WhatsApp alone gives Meta an instant consumer AI footprint that dwarfs ChatGPT’s reach. The question is whether Meta can turn that engagement into revenue without wrecking user trust, and whether the model is actually good enough to hold users when they have free alternatives one tab away.
The early reviews from independent developers are cautiously positive. Muse Spark is said to be strong at reasoning and health-related queries, weaker at creative writing than Claude, and roughly on par with GPT-class models on coding tasks. The Ray-Ban Meta integration is the wildcard. If Meta can deliver a voice-first AI assistant inside a pair of sunglasses that actually works, it will be the first mainstream form factor to put generative AI into a truly ambient product.
THE REGULATORY SHADOW
One more thing worth watching. The closed-source pivot gives Meta new leverage with regulators who have been fighting over AI safety frameworks for two years. By closing the model, Meta can credibly argue it has more control over misuse, which is exactly the argument OpenAI and Anthropic have been making in Washington and Brussels. Expect Meta’s policy team to lean on that framing hard in the next round of AI rulemaking hearings.
The risk is that regulators do not buy it. The European AI Act treats companies with frontier models as high-risk regardless of their open-versus-closed posture. In Washington, the antitrust scrutiny that already surrounds Meta’s core advertising business will not get easier because the company is now also a dominant AI platform. If Muse Spark succeeds, the regulatory file on Meta will get thicker, not thinner.
The bottom line: Zuckerberg has made the biggest strategic pivot of his post-metaverse career. Whether Muse Spark lives up to the spend is a question for the next four quarters. Whether the market rewards him for trying is a question for tomorrow’s open.
Meta’s own announcement is live on the Meta AI blog. For independent coverage and benchmark analysis, see the Bloomberg writeup.
