The Substrate Compounding

On May 18, Anthropic acquired Stainless for more than $300M, a figure reported by The Information and not officially confirmed by Anthropic. Stainless was founded in 2022 by Alex Rattray, a former Stripe engineer, and reached an approximately $150M valuation in a December 2024 round led by Andreessen Horowitz with Sequoia participating. Its product was a code-generation pipeline that turned OpenAPI specifications into production-grade SDKs across the languages everyone ships in, and more recently into MCP servers. The customer list is what makes the deal legible. OpenAI, Google DeepMind, Cloudflare, Perplexity, Replicate, Runway, LangChain, Groq, and DigitalOcean had built their developer surfaces on top of Stainless, and the generated SDKs were being downloaded tens of millions of times per week as of late 2024. On day one of the acquisition, Anthropic announced that all hosted Stainless products would be wound down. The SDK generator goes away, new signups and new projects close, and the hosted maintenance loop ends. Customers keep whatever has already been generated, frozen as it stands. The founder and team move to Anthropic to focus on Claude Platform capabilities. This is the first Anthropic acquisition whose primary effect on the market is subtracting capability from rivals rather than adding it to Anthropic. The Bun acquisition in December 2025 added a JavaScript runtime and build tool to the Claude developer surface. The Vercept acquisition in February 2026 added a computer-use harness, a system for letting agents drive a real desktop without falling off the screen. Stainless is a different shape. The thing Anthropic bought was, in operational terms, the SDK and MCP authoring pipeline that OpenAI and Google DeepMind were using to ship their own ecosystems. Anthropic just took it off the board.

Three weeks ago Anthropic’s posture became visibly different. The May 6 SpaceX/Colossus 1 agreement granted Anthropic access to more than 220,000 NVIDIA GPUs and over 300 megawatts of capacity already physically deployed in Memphis. The May 18 Stainless deal made the developer-tooling story. Combined with Bun, Vercept, and the May 13 launch of Claude for Small Business (a vertically packaged bundle of QuickBooks, PayPal, HubSpot, Slack, Canva, DocuSign, Google Workspace, and Microsoft 365 connectors, plus fifteen prebuilt agentic workflows and fifteen skills), that is five distinct non-model levers accumulated in roughly six months. The obvious counter-reading is that every frontier lab is doing this. OpenAI has Cerebras and a stack of Microsoft commitments. Google has its TPU build-out and the Workspace bundle. xAI has Colossus 2. Substrate accumulation is not a pivot away from being a model lab; it is the current cost of staying one at the frontier. The interesting question is therefore not whether Anthropic is accumulating non-model levers (everyone is), but whether the pattern of which levers and how they are being acquired reveals something specific about Anthropic’s bet. The Stainless deal is what makes me think it does.

Run the inventory of levers with dates. Compute: approaching 10 gigawatts of contracted or committed capacity across four substrates and four counterparties. Up to 5 GW with Amazon under the Trainium-heavy Project Rainier, 3.5 GW of Google TPU through Broadcom starting 2027, a $30B Microsoft and NVIDIA Azure commitment with paired NVIDIA and Microsoft investments back into Anthropic, a $50B Fluidstack data-center build, and now Colossus. Developer tooling: Stainless, May 18, 2026. Runtime and build: Bun, December 2025, the JavaScript runtime that has been gaining ground against Node in modern toolchains. Computer-use harness: Vercept, February 2026. Vertical packaging: Claude for Small Business, May 13, 2026, the first Anthropic product whose unit of distribution is the SMB workflow rather than the API call. Five levers, each one specific, each one with a date, each one acquired or built on a timeline that suggests this has been the plan rather than a sequence of opportunistic moves. The same week the Stainless deal hit, the US Department of Commerce announced voluntary pre-deployment evaluation agreements with Google DeepMind, Microsoft, and xAI under the Center for AI Standards and Innovation framework, building on prior agreements with OpenAI and Anthropic. Even the regulatory substrate is consolidating around the labs that committed early.

The analytical core is the shape of the Stainless deal, not the size of it. A 2x premium five months after a $150M round is not by itself unusual for a hot-sector acquisition. What is unusual is the wind-down decision. Anthropic could have kept Stainless operating, served its existing customers, and collected the SDK-generation revenue while integrating the team. They chose not to. The hosted product line is being shut. New customers cannot sign up. The OpenAI SDK, the Google DeepMind SDK, the Cloudflare SDK, and every other lab and developer-tools company that had been getting Stainless to maintain their generated client libraries now has to plan a migration or absorb the maintenance burden internally. This is a deliberate decision to make those competitors absorb a switching cost. Forrester analyst Biswajeet Mahapatra and Omdia analyst Lian Jye Su, speaking to InfoWorld about the deal, both read it as the moment differentiation moved from model performance to “developer tooling, orchestration layers, and ecosystem connectivity” (Mahapatra) and “from pure model performance to system integration and application-level performance” (Su).

What Anthropic captured at the reference-implementation layer is worth spelling out, because the protocol itself stays open. Stainless was the production-grade pipeline from OpenAPI spec → MCP server that every major lab and developer-tools company was relying on. MCP as a protocol has to stay open; that is the property that makes it a standard. What changes is that the canonical authoring tool for the protocol now sits inside Anthropic. The closest precedent is Chrome and Blink. The HTML and CSS standards remained open; Chrome’s market share and engineering velocity in implementing them produced a state where the de-facto rendering engine became the one inside the dominant browser. Opera retired Presto and switched to Blink in 2013. Microsoft retired Trident when Edge moved first to EdgeHTML in 2015 and then to Chromium and Blink in 2020. The protocol stayed open at the spec while the reference path consolidated. With Stainless wound down as a hosted product and Rattray’s team integrated into Anthropic, the equivalent consolidation inside MCP is now visible. Customers who had managed Stainless workflows have until those workflows end to migrate; what they migrate to is, in practice, whatever Anthropic ships next on top of the same pipeline, or a still-immature alternative like Fern or Speakeasy.

The operator-level translation matters because it changes how routing and tooling decisions compound in a production workload. The Claude Code CLI I run every day rides a Python SDK that Stainless lists on its own customer page, so I am one layer downstream of the wind-down even though I am not in the customer cohort that gets disrupted. That SDK is frozen now and will continue working, but the next time the Anthropic API surface changes in a way that requires a regeneration, the upgrade path is whatever Anthropic ships internally, not what the open market produced. An operator who has accumulated multiple Anthropic substrate decisions (their MCP servers were generated through the Stainless pipeline, their runtime is Bun, their computer-use harness is downstream of Vercept, their packaged workflows are Claude for Small Business connectors) is several layers deep into a stack whose maintenance schedule is set by one vendor. The model is the layer that has commoditized and that you can swap on a per-call basis through a router. The other four are decisions that compound over months and years. Once an SDK is in production, you do not regenerate it casually. Migrating off Bun once it is your runtime is similarly expensive. Once workflows are anchored to a specific connector catalog, the cost of moving them grows with usage.

The thesis only earns its keep if it can be falsified, so it is worth naming what would weaken it. If Anthropic open-sources the Stainless pipeline within ninety days in some meaningful form (the announcement makes no such commitment), the consolidation reading weakens substantially. If OpenAI ships its own SDK-generator and MCP-authoring tooling at comparable quality on a similar timeline, the substrate gap closes and the deal becomes defensive rather than offensive. If Stainless-generated SDKs for OpenAI, Google DeepMind, and Cloudflare continue working and getting maintained for the next twelve months without measurable degradation, the wind-down was a graceful sunset rather than a denial-of-service. And if it turns out the entire deal was driven by Anthropic needing to deepen its own SDK velocity (the talent-acquisition framing), with the rival-disruption a side effect rather than the goal, the strategic read is wrong even if the outcome looks the same. Of those four, the graceful-sunset scenario is the most likely on a sixty-day horizon, because letting established customer SDKs keep working is the cheapest way for Anthropic to manage the reputational tail of the deal. If that is what we see, the wind-down was a soft landing and the take-out-of-circulation reading was overstated. Any of the four would shift the interpretation. None is impossible.

Absent those reversals, the frame I have been trying to operate inside (which model is best this week, what is the cheapest API per token, where do the benchmarks land in June) was never going to produce the right decisions. That frame puts the model layer at the center of strategic differentiation. Three weeks of Anthropic capital deployment do not. Colossus added watts, Stainless changed the tooling layer, and the remaining three added runtime, harness, and packaging. None of the five was a model move. The model is the layer Anthropic ships on top of the substrate it has been quietly acquiring around the model. The router on my workstation has a config line for which model to call. It does not have a config line for which substrate I am inside. By the time it needs one, I will have already made it.