Anthropic just paid roughly $300 million for Stainless, the startup whose tooling generates the SDKs (software development kits) and MCP (Model Context Protocol) servers behind the official client libraries for OpenAI, Google, Cloudflare, and Anthropic itself.

Models aren’t the story anymore

For the last two years, the frontier AI conversation has been almost entirely about models — bigger context windows, better reasoning, lower cost per token. That story isn’t over, but it’s no longer where the most interesting moves are happening. The interesting moves are at the interface between models and the actual work — the SDKs, the agent frameworks, and increasingly, the Model Context Protocol.

MCP started as a clean idea: a standard way for models to connect to tools, data, and systems. It was easy to dismiss as plumbing. Plumbing turns out to be where the value lives.

What Anthropic actually bought

When Anthropic spends $300M to buy the company that generates the SDKs everyone uses, they are not buying a code generator. They are buying the standard layer that decides how every model talks to every tool. In a world where most enterprises run multiple models, swap models often, and connect them to dozens of internal systems, owning the canonical SDK and MCP layer is structurally more valuable than owning any one model.

It’s not just Anthropic

The pattern shows up everywhere right now. Google and Blackstone launching a $5B TPU (Tensor Processing Unit) AI infrastructure venture. The Linux Foundation’s Agentic AI Foundation adding 43 new members, heavy on regulated industries. Dust raising $40 million to build “multiplayer” agentic AI for enterprises, explicitly pitched as moving past isolated assistants. These are not unrelated stories. They are the same story told from three vantage points: compute, standards, and workflow.

Here’s what this means if you’re operating in a regulated space — clinical trials, financial services, anything where audit trails matter.

Models are substitutable. Workflows aren’t.

The model layer will keep changing. The model your team standardized on this year will not be your default model in 18 months. The same is increasingly true of model vendors, fine-tuned variants, and even underlying chip architectures. The layer that needs to be stable is the orchestration layer between models, tools, processes, and people. That layer needs to be model-agnostic, MCP-fluent, governance-native, and audit-ready from the first day. It cannot be retrofitted. The organizations getting this right are the ones treating the workflow layer as durable infrastructure and treating models as substitutable inputs.

This is the layer we’ve been investing in at AG Mednet, in clinical trials specifically. Process orchestration that sits across sponsors, CROs, and sites, model-agnostic by design, with the audit trail that regulated environments demand. The frontier model wars are noisy. The orchestration layer is quiet. The orchestration layer is also where the work actually compounds.

One question worth asking

A practical takeaway for anyone building or buying enterprise AI right now: when you evaluate a platform, ask not “which model does this run on” but “how cleanly can I swap the model six months from now without rebuilding the workflow.” If the answer is “you can’t,” you are buying a lock-in dressed up as a feature.

‍

John Paul (JP) Lee, COO, AG Mednet. McKinsey alum and Kellogg MBA, JP drives operational strategy for Judi across regulated clinical trial environments.

‍