Proximal risk from buyer insourcing

This risk says Proximal is selling into a market where a few buyers can decide whether vendors exist at all. Proximal serves frontier labs and large enterprises with coding training data and RL infrastructure, and its product includes turning code, traces, and workflow artifacts into evals and post training tasks. If a top customer builds that workflow in house, it is not just replacing a dataset vendor, it is removing an entire recurring spend category.1 2 3 4

• The insourcing path is real because adjacent tools already let labs run more of the pipeline themselves. Labelbox supports internal teams, outside vendors, or Labelbox managed services on the same platform, and RLHF workflows can be launched programmatically. That makes software procurement easier for buyers that already have research ops staff.5 6
• Competitors are split between service heavy and software heavy models. Surge packages expert built frontier datasets and custom engagements. Scale sells RL environments that are real software setups with graders and verification. Proximal pushes furthest toward automated generation and verification, which helps if buyers want outcomes, but hurts if they only want tooling to run internally.1 7 8 9
• This concentration matters more because the customer set is tiny. Proximal focuses on frontier labs training coding agents and enterprises deploying them internally. The company reached an estimated $200M annualized revenue by September 2026, so even one large account likely represents meaningful volume in a market with few natural buyers.1 2 4

Over time this market is likely to split in two. The biggest labs will build more of their own data engines and buy outside help only for overflow or frontier edge cases. Independent vendors will keep winning where they can show faster model improvement than an internal team, especially on long horizon coding tasks that are expensive to design, verify, and refresh.1 7 8 9 10