In June 2025, Meta bought 49% of Scale AI for about $14.3 billion. Founder Alexandr Wang left to lead Meta's superintelligence efforts, and within weeks the market had its answer about how customers felt: Google — Scale's largest customer — wound down work, OpenAI and xAI stepped back, and competitors reported a surge of inbound interest.
If you are evaluating data labeling vendors in 2026, the Scale story matters less for its drama and more for one question it forced into the open:
who does your annotation vendor work for?
Short answer
Scale AI remains a strong vendor for large managed projects — government work, frontier-model data programs, and high-security engagements.
Consider alternatives when:
- your project is small enough that managed-service economics do not fit
- you want your labeling operations in-house, under your own control
- vendor neutrality is a procurement criterion (it increasingly is)
What actually changed at Scale
The timeline, briefly:
- June 2025: Meta's 49% non-voting stake; leadership change; Meta secures a five-year commercial commitment
- Mid-2025: major AI-lab customers reduce or exit over neutrality concerns; ~200 employees and ~500 contractors laid off; internal restructuring
- 2025-2026: Scale reports strong bookings and pivots emphasis toward government contracts (defense programs), applications (agents, GenAI platform), and evaluation leaderboards (SEAL), while the commercial data engine continues serving remaining customers
Scale did not collapse — by most accounts it had a strong financial year. But the deal clarified a risk that was always implicit: a labeling vendor can become strategically aligned with one AI company, and its other customers may become that company's competitors.
The market's response: two migrations
The post-deal demand did not vanish; it moved.
- To premium expert-data firms — independent providers (Surge AI, Mercor, Handshake AI and others) absorbed RLHF and evaluation demand, partly because they are vendor-neutral
- Back in-house — teams with ongoing annotation needs started keeping data operations internal: their own people, their own guidelines, their own tooling
Migration 2 is the interesting one for computer vision teams, because the tooling barrier collapsed. What used to justify a managed service — enterprise workflow software, QA infrastructure, versioning — is now available as affordable SaaS.
Managed service vs in-house: the honest comparison
| Dimension | Managed service (Scale-style) | In-house team + tooling |
|---|---|---|
| Best for | One-off giant projects, specialized domains, surge capacity | Ongoing labeling, evolving guidelines, sensitive data |
| Cost shape | Contract-based, typically $100K+ per project | Software subscription + annotator salaries |
| Guideline evolution | Slow iteration through vendor channels | Same-day fixes, reviewer calibration on real data |
| Data control | Contractual (NDPs, deletion clauses) | Physical — data never leaves your workspace |
| Neutrality risk | Depends on vendor ownership | None — you are the vendor |
| Domain knowledge | Vendor-trained workforce, turnover risk | Your team compounds knowledge every sprint |
Managed services still win for one-off, massive, or highly specialized efforts (a million-frame LiDAR program, expert medical judgment at scale). For continuous image annotation that feeds an in-house model pipeline, the math increasingly favors insourcing.
The in-house stack in 2026
A credible in-house operation needs four things, none of which require enterprise procurement anymore:
- Annotation workspace with review workflow — assignments, reviewer roles, rejection notes, audit logs
- AI-assisted labeling — SAM-class segmentation to cut per-image time; ideally without shipping your images to a third party
- Dataset versioning — so every training run is traceable to an exact annotation state
- A quality system — agreement metrics, honeypots, weekly QA sampling
At five to ten people, the total software bill for this stack should be in the tens of dollars per month — not a six-figure contract.
Where LabelOp fits
LabelOp is built for exactly the in-house migration:
- your team, your guidelines — assignments with priorities and due dates, OWNER/ADMIN/REVIEWER/ANNOTATOR role gates, review notes attached to the work
- AI assist without data leakage — SAM 2 smart annotation runs in your browser, so assisted labeling never uploads your images for inference
- traceability by default — dataset version snapshots with compare and rollback, plus audit logs that record who changed what
- flat pricing — $49/month for a 10-person workspace; the free tier keeps 3 people and private data, so pilots cost nothing
- GPU compute stays yours — training runs on your own Vast.ai credentials; no vendor relationship between your model and your labeling tool
For teams stepping down from a managed contract, the sequence that works: keep the vendor for the next big batch if one is already in flight, stand up the in-house workspace in parallel, migrate guidelines into your own system, and shift ongoing labeling internal batch by batch.
The QA discipline that makes in-house work — agreement metrics, calibration, release gates — is covered in the annotation QA workflow playbook.
When Scale is still the right call
Be fair to the incumbent:
- Government and defense programs — Scale's public-sector business is its growth engine for a reason
- Frontier-model data programs — RLHF, red-teaming, and evaluation at a scale no in-house team matches
- Surge capacity — a 500K-image one-off with a hard deadline is a vendor job, not a hiring plan
The 2026 skill is matching the engagement type to the vendor type — not declaring one side dead.
Final takeaway
The Meta-Scale deal turned vendor neutrality from a paranoid question into a procurement checkbox.
For ongoing image annotation, the growing answer is in-house: your team, your guidelines, your data, on tooling that costs less than one month of a managed pilot.
For one-off giants, managed vendors — including Scale — still earn their contracts.
FAQ
Why did Scale AI's customers leave after the Meta deal?
Neutrality concerns: with Meta holding 49% and Scale's founder moving to Meta, other AI labs worried their data programs would indirectly benefit a competitor. Google, OpenAI, and xAI all reportedly reduced or wound down engagements in mid-2025.
Is Scale AI still operating normally?
Yes — the company reported strong bookings through 2025-2026, with growth concentrated in government contracts, applications, and evaluation products alongside its continuing data engine.
What is the main alternative to managed data labeling?
An in-house labeling team using modern annotation software: flat-priced workspaces with review workflow, dataset versioning, and AI-assisted labeling. Software costs have collapsed from enterprise contracts to tens of dollars per month.
How do I move from a managed vendor to in-house labeling?
Stand up your own workspace in parallel, migrate your guidelines and a small pilot batch first, run QA calibration with your own reviewers, then shift ongoing work batch by batch. Keep the vendor for genuinely one-off large efforts.
How does LabelOp support in-house labeling teams?
LabelOp provides the workspace layer: assignments and role-gated review, in-browser SAM 2 assist, dataset version snapshots, audit logs, and flat $49/month pricing for a 10-person workspace — with your training compute on your own GPU account.