Meta AI Timeline: Llama, Muse, Scale AI, and Manus

An advertising company funding a frontier bet

Meta runs Facebook, Instagram, WhatsApp and Messenger, and it sells ads against them. Roughly 98 percent of its revenue is advertising, and the second quarter of 2026 alone brought in $60.8 billion, up 28 percent from a year earlier. Mark Zuckerberg remains chairman and CEO, holding voting control through a dual-class share structure. [2][3]

That business is now funding a much larger one. Meta's 2026 capital-expenditure guidance rose three times over the year: from an initial increase announced in January, to a $125 billion to $145 billion range in April, to a narrower but still higher $130 billion floor by the July earnings call. Free cash flow fell as the spending grew. [4][5][3]

Atlas interpretation: Meta is not short of cash, so the spending itself is not the story. The story is that investors have reportedly been more willing to extend Google the benefit of the doubt on AI capex discipline than Meta, which says less about the dollar figures than about how each company's bet reads from the outside. [6]

Buying a lab and losing its chief scientist

In June 2025, Meta put $14.3 billion into Scale AI for a 49 percent stake, a deal that valued the data-labeling company at about $29 billion and brought its 28-year-old founder, Alexandr Wang, to Meta as its first Chief AI Officer. Wang now leads Meta Superintelligence Labs, the unit built around this deal. [7][8]

Five months later, Yann LeCun left, after twelve years as Meta's chief AI scientist, to start a company built on the position he had argued for throughout his time there: that scaling language models further will not reach human-level intelligence. He left a lab Meta had just spent $14.3 billion building around the opposite bet. [9]

Atlas interpretation: Reported reorganizations followed at a pace that outran the new lab's output: one account counted a fourth restructuring of Meta's AI effort inside six months, alongside a round of roughly 600 layoffs inside the unit itself. Buying talent turned out to be the easy part; giving it a stable structure to work inside has been harder. [10]

Llama opened the door, Muse tried a different one

Meta's standing contribution to AI history is Llama. The original LLaMA release in February 2023 leaked onto BitTorrent within a week of its research-only distribution, and Llama 2's commercial license that July turned the leak into a strategy: publish capable weights, attach usage terms, and let the ecosystem build on your model instead of a rival's. Much of the open-weight boom that followed traces back to those two releases. [11][12]

Meta Superintelligence Labs' first model broke that pattern. Muse Spark arrived in April 2026, nine months into rebuilding the model stack under Wang, and shipped closed: API access only, no downloadable weights. For the company that made open weights a competitive identity, the closed release was the news as much as the model itself. [13]

Four months later Meta reversed again. Muse Glimmer, a 30-billion-parameter model distilled from Muse Spark, published under an Apache 2.0 license and quantized to run on a single consumer GPU. Muse Spark's own weights stayed shut. [14]

Atlas interpretation: Reading the two releases together, openness at Meta looks less like a principle than a lever the company pulls depending on where it sits competitively that quarter. That is a defensible way to run a model line. It is a different story than the one Llama's early marketing told. [13][14]

The acquisition China would not allow

Meta's other 2025 AI purchase was smaller in dollars and larger in complication. In December, it agreed to buy Manus for more than $2 billion, a Singapore-based agent startup that had gone from launch to over $100 million in annualized revenue in eight months. Meta's own condition on the deal was that Manus wind down its operations in China and end any continuing Chinese ownership. [15]

Four months later, Beijing pulled the thread Meta had left hanging. China's National Development and Reform Commission ordered the deal unwound, citing the cross-border transfer of AI technology developed in China. By July, Tencent had stepped in to lead a deal absorbing the unwound stake, and Manus returned to operating as an independent company. [16][17]

Atlas interpretation: The Manus reversal is a specific, dated instance of a general problem: an AI acquisition can clear every American regulator and still be undone by a Chinese one, if the technology or the founders have Chinese origins. It is worth watching for whether other AI-buying companies with China-linked targets, xAI's and SpaceX's included, hit the same wall. [16]