Meta spent a decade being the open-source darling of the AI era, and then Llama 4 landed flat. Benchmarks were massaged, the reasoning model slipped, the community grumbled, and Mark Zuckerberg did what founder-CEOs do when the flagship product disappoints: he reorganized the company around the thing he actually wants, which is superintelligence. Meta Superintelligence Labs (MSL) is the result. It is the most aggressive talent grab the industry has ever seen, and it is also an admission that the previous strategy stopped working.
What actually happened with Llama 4
Llama 4 arrived in spring 2025 with big claims and a lopsided reception. A specialized chat variant scored well on LMArena, then it emerged that the arena-tested build differed from the released one. The flagship reasoning model slipped its timeline badly. Enterprises that had standardized internal tooling on Llama 3.x suddenly faced a model family that looked great on slides and middling in production. None of this was catastrophic, but for a company whose open-weights strategy depends on developer trust and ecosystem momentum, it was a genuinely bad quarter.
The deeper issue wasn't one model launch. It was that the open-weights treadmill stopped being differentiated. DeepSeek undercut Meta on price and matched or beat Llama on reasoning. Qwen kept shipping quickly. Mistral held its niche. Meta was spending billions to stay roughly even while burning nine figures a month on compute and getting roasted on every release day.
Enter Superintelligence Labs and Alexandr Wang
Meta's answer was to hire a 28-year-old who had dropped out of MIT, built Scale AI into the dominant data-labeling and eval shop in the industry, and turned it into a recruiter for frontier labs. Alexandr Wang joined Meta as chief AI officer in mid-2025, and the deal structurally included a massive investment in Scale with Wang leading Meta's superintelligence effort. The optics were unmistakable: Meta bought the best recruiting pipeline in AI and put its owner in charge of the lab.
Superintelligence Labs consolidated Meta's various AI efforts — the FAIR research org, the genAI product teams, the infrastructure groups — under one roof with one mandate: reach superintelligence, full stop. Zuckerberg made clear this was the company's top priority and that capex would follow. Guidance ballooned into the hundreds of billions of dollars over coming years, which, when you put it next to Reality Labs' losses, tells you where the board's patience lives.
The hiring spree
MSL's first move was not shipping a model; it was acquiring the people who ship models. Reports pegged Meta's offers to top researchers at tens of millions of dollars per year, with some packages reported at nine figures over multi-year horizons. Notable moves included pulls from OpenAI, Anthropic, Google DeepMind and Apple, with whole teams lifting out in some cases. Meta also took non-controlling stakes and made acqui-hire-style deals with several small research shops, including talent associated with Thinking Machines and other labs. Scale's own data and eval machinery now serves Meta directly, which is a quiet but enormous structural advantage: the best eval pipeline in the industry, running on Meta's side of the table.
There's a fair criticism that this is buying research the expensive way, and that research bought by the yard doesn't compound like research grown at home. Fair enough. But frontier labs are also talent markets, and for once Meta was the buyer instead of the seller. In a field where five people can move the loss curve, paying nine figures for them is not irrational — it's just expensive.
What changed technically
Under MSL, Meta has moved toward a more deliberate, less release-date-driven cadence. The open-weights commitment is still publicly affirmed, but the releases now come with heavier caveats and more staged rollouts. Meta's reasoning models post-Llama 4 have shown more of the test-time-compute approach — longer chains, more RL — rather than pure pretraining scale, mirroring where the whole field went after DeepSeek R1. Meta also leaned into multimodal and agentic tooling for its social platforms, because that's where its data and distribution advantage is real. Nobody else on earth has billions of daily user interactions to train against.
The strategy question nobody answers cleanly
Why does Meta want superintelligence, exactly? The honest answers have shifted over time. The 2023 answer was "to win friends and influence developers via open source." The 2025 answer is "because everyone else is racing and we can't be left out of the defining technology of the century." The commercial answer is "ads, engagement and social agents," and that's fine. But investors keep asking a different question: where's the revenue line for a few hundred billion in capex? Meta's ad business is enormous, and AI-driven targeting and assistants can plausibly pay for a lot of this. Still, the gap between "we spent $400B on compute" and "here's the P&L" is the widest in big tech.
There's also the open-source tension. Open weights are fantastic for commoditizing competitors' edges and terrible for capturing direct revenue on the model itself. Meta is trying to walk that line — keep the open ecosystem warm enough to stay relevant to developers, keep the truly frontier stuff a little closer to the chest. Historically Meta has leaned liberal on openness; the Superintelligence era may lean more cautious. Watch the license terms on whatever the next flagship is called. They will tell you everything.
How Meta compares to the other labs right now
It's tempting to read the MSL pivot as purely reactive, but the honest comparison is structural. OpenAI has Microsoft's capital and distribution. Anthropic has safety-brand gravity and Bedrock/GCP plumbing. Google has TPU and DeepMind research depth. Meta's unique assets are different: the largest social data firehose on earth, the best open-source distribution channel in the industry, and a founder who will spend his own money without asking the street for permission first. That combination is genuinely dangerous to the incumbents, and everyone in the room knows it. The open-source community, meanwhile, is nervous. A Meta that closes up shop on frontier releases doesn't just lose the developer leaderboard; it torches a decade of goodwill that made Llama the default open model stack.
There's also the Llama license question hanging over everything. Meta has always drawn the line somewhere — commercial use allowed, with thresholds for very large companies — and MSL has done nothing to clarify where the line sits for the next flagship. If you're building a product on Llama-class weights, that ambiguity is a real planning risk, not a theoretical one.
The culture question
FAIR, Meta's research org, was once the crown jewel of industrial AI research, home of PyTorch and a string of genuinely open contributions. MSL effectively supersedes it as the center of gravity, and the cultural friction is real: research-first timelines versus product-first ambition, openness versus secrecy, publication norms versus competitive hoarding. Alexandr Wang's org is run more like a lab with a launch problem than a research institute with a product aspiration, and that's coherent — but it will cost Meta some of the science goodwill it spent a decade accumulating. Good researchers don't love being told the release date comes first. Some will stay for the mission and the money. Some won't.
The quiet product roadmap
Under the reorganization, Meta's AI product surface is converging on one idea: a personal assistant woven through WhatsApp, Instagram and Facebook, plus business-facing agents built on the same backbone. That's distribution no lab can replicate. If MSL ships a genuinely strong agentic model and Meta wires it into three billion accounts, the usage numbers will dwarf every competitor overnight, whether or not the benchmark tables say so. The risk is the opposite one: Meta's AI features have historically felt bolted on, and a mediocre superintelligence in a chat box is still a mediocre product. Execution, as always, is where Meta's ambition goes to meet its track record.
What to watch next
- The next big model release. Its post-training quality and real-world eval behavior matter more than the benchmark table.
- Wang's eval grip. If Scale's machinery quietly becomes the de facto eval standard, Meta gains outsized influence over what "frontier" even means.
- Capex disclosures. Watch the quarterly guidance like a hawk. If it keeps climbing, so does the pressure to ship something that wows.
- Executive retention. Nine-figure packages buy presence, not loyalty. How many MSL stars are still there in 2027 is a real question.
- The open-weights fault line. Whether the flagship stays open-ish or goes fully proprietary will define Meta's developer relationship for years.
What the Scale AI pipeline means in practice
The Scale AI acquisition isn't just about Alexandr Wang & his recruiting Rolodex; it's about the machinery underneath. Scale built the dominant data-labeling, evaluation & red-teaming pipeline for frontier labs, & that pipeline now feeds Meta directly. Instead of Meta renting eval horsepower from the same boutique shops everyone else uses, Meta effectively owns the referee's whistle. When Scale's post-training datasets, preference rankings & failure taxonomies shape what "good" means inside MSL, the model gets tuned against the measurement system, not around it & a subtle but compounding advantage.
This also changes Meta's relationship with the broader eval ecosystem. If Scale's methodology becomes the de facto standard for agentic & reasoning benchmarks, every downstream buyer & enterprise CIO, regulator, open-source evaluator & inherits assumptions baked into one company's pipeline. That's not sinister by default; Scale was already the industry referee. But concentrating referee, competitor & customer in one building deserves scrutiny, & the open-source community is right to watch the license terms & the eval disclosures more carefully than they did for Llama 3.
The practical read for a builder: when MSL ships its next flagship, don't take the benchmark table at face value. Ask which eval suite produced it, whether Scale ran it, & whether anyone independent reproduced the numbers. The company that owns the measuring stick owns a slice of the game's rules, & for once Meta seems content to hold that slice rather than rent it.
My take
Meta is doing the thing disciplined incumbents do when the ground shifts: overcorrect, spend obscenely, and try to buy the learning curve. Sometimes it works — Google did a version of this with DeepMind and AlphaGo. Sometimes it doesn't — plenty of nine-figure talent orbit-outs produced very little. The Llama 4 stumble was real but it wasn't fatal; the open ecosystem still loves Llama, and Meta still has the best data engine in social. What's genuinely uncertain is whether MSL can turn compute and talent into a coherent research program before the cash runs through the patience of Zuckerberg's board. If anyone can force superintelligence with money, it's the person with the most money pointed at the most GPUs. Whether money is the binding constraint in 2026 is, as always, the trillion-dollar question.