Developer Tools, MLOps and Data Infrastructure

Between the model labs and the applications sits a layer of companies that sell the tools to build, feed, test, run and watch AI systems: data platforms (Databricks, Snowflake), data movement and governance (Fivetran, dbt, Informatica), retrieval and vector databases (Pinecone, Weaviate, Qdrant, pgvector), data labelling and expert data for training (Scale AI, Surge AI, Mercor), evaluation and observability (Braintrust, LangSmith, Arize, Datadog), experiment tracking and training tools (Weights & Biases), inference platforms and gateways (Together AI, Fireworks, Baseten, Modal, OpenRouter) and, the largest commercial category by 2026, AI coding tools (GitHub Copilot, Cursor, Claude Code, Codex, Devin, Replit). It is the "picks and shovels" layer, and like every picks-and-shovels layer it is a mix of a few very large winners and many companies whose product turned out to be a feature of someone else's platform. Its economics differ sharply by sub-layer: data platforms earn consumption revenue with ~70–80% gross margins; labelling marketplaces report gross billings of which most is paid to contractors; inference platforms resell GPUs at thin margins; coding tools resell model tokens and see their gross margin set by the labs that supply, and compete with, them. A developer-tools business is therefore an answer to four questions: is this a product or a feature, who owns the workload, how much of revenue is pass-through cost, and what happens when the model lab or the cloud ships the same thing? This section maps the layer, the business model of each part, the 2023–2026 consolidation wave (MosaicML, Tabular, Neon, Weights & Biases, Informatica, Windsurf, Cursor), the coding-tool boom, the gross-versus-net revenue trap in data labelling, landmark cases, pitfalls and how experts read the layer.