A year after Meta's Scale AI deal, the data-labeling industry looks nothing like it did before

Meta's unusual, multi-billion-dollar investment and talent arrangement with Scale AI - which brought Scale founder Alexandr Wang into Meta to help lead its superintelligence-focused research effort while leaving Scale as a nominally independent company - triggered a wave of client defections that has permanently reshaped the data-labeling and human-feedback industry Scale had dominated for the better part of a decade. Google, OpenAI and several other frontier labs that had relied on Scale for training-data annotation and reinforcement-learning-from-human-feedback pipelines moved meaningful volumes of that work to competitors within months, unwilling to route sensitive training-data pipelines through a vendor now structurally entangled with a direct competitor. The beneficiaries of that reallocation - including Surge AI, Turing and a handful of newer entrants - have grown rapidly, and the episode has accelerated a broader shift in how frontier labs think about data-pipeline vendor risk, with several now deliberately splitting annotation and RLHF work across multiple providers specifically to avoid the concentration risk that the Scale-Meta arrangement exposed. The labeling industry's economics have also shifted upmarket, from large volumes of relatively low-skill annotation toward smaller volumes of highly specialised expert feedback - PhD-level scientific review, professional coding critique, legal and medical domain expertise - as the marginal value of more generic labeled data has declined relative to the value of frontier-quality expert judgment. India has been a direct beneficiary and a direct casualty of this reshuffling simultaneously. Indian annotation and BPO-adjacent firms that built businesses on high-volume, lower-skill labeling work have faced margin pressure as that segment commoditises, while a smaller cohort of Indian firms specialising in domain-expert annotation - drawing on the country's large pool of English-fluent engineers, doctors and lawyers willing to do part-time AI-training work - have found a growing and better-paying niche. Meta's own research output since the Scale deal has been closely watched as the measure of whether the unusual talent-and-equity arrangement actually delivered the capability uplift the company was paying for, with mixed early signals - some infrastructure and research-process improvements attributed to the new team, but no clear frontier-model breakthrough yet publicly attributable to the reorganised effort. What to watch: whether Scale AI's remaining client base stabilises or continues eroding, whether Meta's Superintelligence Labs group ships a model that closes the perceived gap with OpenAI, Anthropic and Google, and whether any Indian annotation-services firm makes a credible push upmarket into expert-feedback categories at scale.

Original source: The Wall Street Journal