Anthropic’s data-vendor spend races toward a ~$900M exit rate

What's the story? Wing Venture Capital partner Tanay JaipuriaDealroom has a profile for this one. Try Dealroom → (@tanayj) charts Anthropic’s estimated annualized spend on data quality and performance-evaluation vendors from the company’s S-1 line item. The ladder runs from roughly ~$20M in FY23 to ~$185M in FY24 , then ~$245M (H1’25) , ~$460M (H2’25) , ~$750M (H1’26) , and a Q2’26 exit rate of ~$900M — a near 45× jump in three years as Anthropic scales evaluation and RLHF-style data work ahead of its IPO.

Context The S-1 aggregates “vendor costs related to data quality and performance evaluation services” and does not break out named suppliers. It also does not give the spend level itself, only the increases: up $251M year on year in the first half of 2026 and up $165M in full-year 2025, according to Slator, so the run rates in the chart are estimates backed out from those deltas. The category can include human data annotation, expert-generated training data, model response evaluation and safety testing. The companies tagged here — Scale AI, Surge AI, Mercor, Turing, Labelbox and micro1 — are examples of the broader AI data-quality / eval / expert-labor vendor set that sits in this market, not a claim that Tanay (or the S-1) named them as Anthropic’s vendors. For the wider S-1 filing picture, see Anthropic’s S-1 note.

Source Tanay Jaipuria on X (26 Oct 2026), chart of estimated Anthropic data-vendor spend based on the S-1. Run-rates are Tanay’s estimates from the disclosed line item. Slator (27 Oct 2026), on the disclosed year-on-year increases.