Meta’s big, splashy bet on Scale AI — a $14.3 billion investment and the hiring of Scale’s founder Alexandr Wang to run a new Meta Superintelligence Labs (MSL) — was supposed to be a fast track out of the company’s recent AI doldrums. Instead, barely two months in, the relationship is showing early strain: key hires have left, researchers inside Meta are reportedly reaching for competing data vendors, and the company’s ambitious timeline for the next Llama release is adding pressure.

The June deal that tied Meta and Scale together was extraordinary by any measure — both in cash terms and in signal. Meta didn’t merely write a check; it imported leadership and, implicitly, the hope that Scale’s data and operational know-how would plug a glaring hole in Meta’s AI supply chain. That headline move put Wang at the center of Meta’s new superintelligence push and set high expectations inside and outside the company.

But early reports suggest the honeymoon has run into friction. One of the executives who followed Wang to MSL — Ruben Mayer, Scale’s former SVP of GenAI Product and Operations — left Meta after only about two months, according to people speaking to TechCrunch; Mayer himself described his exit as for personal reasons and disputed some characterizations of his exact role.

More revealing than a single departure, though, are the whispers from Meta’s labs: engineers and researchers in the TBD unit — the centrepiece group inside MSL tasked with building the most ambitious models — have reportedly been relying on other data labeling vendors, including Mercor and Surge. In short, after investing billions in Scale, Meta researchers are leaning on Scale’s competitors. That’s both awkward and unusual for a relationship this deep.

A data problem — or just growing pains?

Scale made its name by scaling labeling cheaply — huge crowds, fast turnaround. But modern foundation models increasingly demand curated, expert-level annotations: clinical review by doctors, legal context from lawyers, scientific judgment from domain experts. Where Scale has tried to pivot toward higher-quality work with offerings like its Outlier platform, rivals such as Surge and Mercor were built from the start around higher-paid, specialized talent — and with that, they’ve grown quickly. Several people familiar with MSL say researchers prefer the higher-touch offerings from those rivals for the kind of datasets that will matter to cutting-edge LLMs.

Meta has pushed back on the “Scale data is low quality” characterization through a spokesperson, and Scale’s own statement after the Meta deal pointed to an expanded commercial relationship. Still, the optics are stark: a company that took a multi-billion dollar stake in a data provider is simultaneously working with that provider’s competitors. Labs often use multiple vendors for redundancy and specialization, but investors and executives normally expect a closer alignment after a deal of this magnitude.

Scale’s shakeup and the wider fallout

Scale hasn’t exactly been immune to turbulence. After OpenAI and Google reportedly stopped working with Scale AI in the wake of Meta’s investment, Scale cut roughly 200 roles in its data annotation business in July — a move its incoming CEO, Jason Droege, attributed to shifting market demand. The company has tried to reallocate resources into other growth areas, including government sales; it recently landed a roughly $99 million contract with the U.S. Army, according to reporting. Those moves underscore a business in transition from bulk crowdsourcing to a more specialized, higher-value service.

For Meta, the deeper worry isn’t only vendor selection. The company’s aggressive recruiting — luring senior researchers from OpenAI, DeepMind and elsewhere — has created both a surge of talent and a management headache. New hires accustomed to startup speed and small-team autonomy have reportedly bristled at Meta’s bureaucracy. Several longtime Meta AI staff have also left amid the reorganization, and some new entrants have departed as well. That churn complicates any plan that depends on rapid, tightly coordinated progress.

Racing the clock: Llama 4.X and what’s riding on it

Compounding the people and vendor issues is a hard deadline: Business Insider reports a team inside TBD is pushing to get a new Llama model (internally nicknamed Llama 4.X or Llama 4.5 in some accounts) production-ready by the end of the year. Rolling out a next-gen foundation model under that timetable is a logistical challenge even in the cleanest of circumstances — add the vendor friction and staffing churn and you have a recipe for stress.

Meta publicly framed the Scale investment and the hiring of Wang as a way to accelerate its AI ambitions after a lackluster Llama 4 launch in April and mounting pressure from competitors. But the company is now attempting to execute a multi-vector strategy: massive data center builds, wide-ranging acquisitions (from voice startups to partnerships with image model vendors), and an internal reorg that centralizes AI reporting lines. Those moves are bold — and they magnify the penalties for missteps.

What’s next — for Meta and Scale

There are a few possible outcomes here. The partnership could smooth out: Scale could double down on higher-quality, expert labeling and Meta could tighten integration and incentives so research teams stick with the vendor they helped elevate. Alternatively, Meta could formalize a multi-vendor strategy, using Scale where it fits and other vendors where they perform better — effectively treating Scale as one tool among many rather than a centerpiece. For Scale, the risk is asymmetric: Meta can diversify its sourcing without existential damage; Scale — having lost other big customers and faced layoffs — has fewer ways to absorb the blow.

Either way, the episode is a reminder that vaunted talent grabs and headline investments don’t erase the nitty-gritty work of model building: high-quality datasets, stable teams, and operational trust. Superintelligence is built on small things — labels, expert reviewers, and the day-to-day collaboration between modelers and data ops. If those pieces don’t line up, even huge checks and marquee hires won’t stop delays, departures, or disappointment.

The human element

One of the ironic byproducts here is how much the story comes back to people. Ruben Mayer’s brief tenure — and his insistence in interviews that his departure was personal and that he had been “part of TBD Labs from day one” — underscores how messy transitions can be when you transplant talent into a different corporate culture. Other departures, including product and research leaders at MSL, have been framed publicly as normal attrition; insiders say the reality is murkier. For a lab chasing something as audacious as “personal superintelligence,” keeping the core team intact may prove as important as any vendor relationship.

Bottom line

Meta put a lot on the table when it bet on Scale AI and recruited Alexandr Wang: money, jobs, and the hope of a shortcut to parity with the likes of OpenAI and Google. What’s playing out now isn’t necessarily a fatal rupture — but it is a test. Can a behemoth reorganize quickly enough, manage incoming talent and old teams, and coordinate the messy supply chain of annotations and expert labels? If Meta wants Llama 4.X out by year-end, it had better hope the answers come fast — because the clock is already running.