The Sequence Opinion #896: Spark, Compute, and the Two Metas
Substack Jesus Rodriguez ● Covered by 2 sources
Meta just launched Muse Spark 1.1 — its first paid AI model with closed weights, after years of open-source talk. Zuckerberg broke a 3-year social silence to announce it, and that says everything.
Based on reporting by Substack, Jesus Rodriguez — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Mark Zuckerberg posted on X last Thursday for the first time in three years. Not a meme, not a photo of his cattle ranch — an announcement. Meta Superintelligence Labs had shipped Muse Spark 1.1, its second model, and the first Meta model ever to carry a price tag. It landed with a public API, an OpenAI-compatible endpoint, pricing of $1.25 per million input tokens and $4.25 per million output tokens, and closed weights. That last detail matters more than the pricing. This is the same company that spent three years telling everyone open weights were the responsible, strategic choice for AI. Now the CEO comes out of self-imposed social media exile to announce the opposite.
Spark 1.1 wasn't a solo act. Two days before it, Meta released Muse Image, the lab's first image generator. A week before that, reports surfaced about Meta Compute, an internal effort to sell spare AI infrastructure to outside companies, the same playbook AWS ran two decades ago. Layer in the custom MTIA chips inching toward production and a pattern emerges: in about eighteen months Meta has moved from an open-weights research group bolted onto an ads business into something assembling the full stack — chips, datacenters, cloud, models, API, apps, devices. Google is the only other company that has ever held every one of those pieces simultaneously.
So the obvious question is whether Meta can actually compete with the frontier labs. I'd argue that's really two separate questions dressed as one, and they don't share an answer. At the layer where models meet actual users — the apps, the agents, the billions of people already inside Instagram, WhatsApp and Facebook — Meta may genuinely have the edge nobody else can match. At the layer where the models themselves get built, the raw research and training frontier, the evidence for Meta leading is thin, and the structural incentives inside a company this size cut against fast iteration.
That split explains why this launch feels contradictory instead of triumphant. Meta is trying to be both an infrastructure landlord and a frontier lab at once, using pricing and distribution to buy time while research catches up. Whether that works depends less on Spark 1.1's benchmark scores and more on whether Zuckerberg's newly vertical company can out-execute rivals who only have to worry about one layer of the stack, not six.
My take — AI-written commentary, not fact-checked reporting
I've watched Meta wrap itself in the open-source flag for three years while quietly building every piece of a closed, vertically integrated empire, and Spark 1.1 is the moment the costume finally slips. This isn't hypocrisy exactly — it's a company that realized owning distribution matters more than owning ideology, which is a very Silicon Valley lesson learned a decade late. Watch Meta Compute closely; renting out spare GPU capacity is the oldest trick in the cloud playbook, and it tells you Zuckerberg thinks the real money is in being everyone's landlord, not just another tenant on the frontier.
Read more about this at: Substack