The buyers here are frontier labs training video-audio-text models that need every modality to agree with itself; the sellers are annotation shops racing to prove they can deliver that agreement rather than just deliver files with tags on them. iotforall.com’s framing on August 6, 2026 is really a pricing argument dressed up as a methodology piece: cheap, high-volume labeling is getting commoditized while the judgment-heavy work — RLHF preference ranking, gold-standard eval sets, cross-modal reconciliation — is where margin concentrates. That bifurcation matters for the named players in this market: MarketsandMarkets lists Scale AI, Labelbox, TELUS International, Cogito Tech, Dataloop, SuperAnnotate, Defined.ai, V7 and Clickworker as the incumbents competing for a segment it expects to hit $3.6 billion by 2027, and the ones that can show inter-annotator agreement scores and layered QA are the ones positioned to capture the premium tier rather than get squeezed on per-label rates.
Stanford’s AI Index datapoint cited by iotforall.com — training datasets doubling roughly every eight months against compute doubling every five — is the scarcity signal underneath all of this. If compute keeps outpacing well-labeled data, aligned multimodal corpora become the binding constraint, and that should push per-unit prices up for anything that requires genuine cross-modal adjudication, even as bulk single-modality tagging keeps getting cheaper or automated away. Forbes’ May 2026 reporting on Versos AI adds a second seller category to the map: licensing platforms that source raw video with C2PA provenance manifests attached, effectively pricing footage before it ever reaches an annotator’s queue, and explicitly building against model-collapse risk from synthetic contamination. That’s a new toll booth in the supply chain, and it suggests provenance verification is becoming its own billable line item alongside labeling itself.
The commodity end of labeling is getting cheaper by the day; the judgment end is getting more expensive by the doubling.
Two counter-pressures deserve skepticism. GE HealthCare’s X-ray foundation model, built on 1.2 million in-house annotated images, is a reminder that well-capitalized buyers can and do vertically integrate rather than outsource — a competitive threat to labeling vendors in domains where proprietary data moats already exist. And DGIST’s EEG-fNIRS foundation model, trained via self-supervision on 1,250 hours of unlabeled brain signal from 918 subjects, is a live example of a lab shrinking its labeling bill almost entirely by design. If self-supervised pretraining keeps eating into label-hungry corpora the way DGIST claims, the premium-judgment segment iotforall.com describes may end up being the only part of the labeling market that grows at all. Watch whether Scale AI, Labelbox and peers start pricing RLHF and eval work as a distinct, higher-margin product line separate from bulk annotation contracts — that repricing, more than any volume figure, will show where this market is actually heading.
Stanford's AI Index reports that training datasets double roughly every eight months while compute doubles every five, and image and video work now sits at the center of that growth. Corpora expanding at that pace reflect a hard truth model teams keep rediscovering: architecture and compute set the potential, but annotation sets the ceiling.
