Anthropic Is Building a Sandwich Strategy
That Has Nothing to Do With Models

Anthropic spent at least $300 million on a company that generates SDKs. On the same day, it handed Claude access to 276,000 KPMG employees across 138 countries. One move points down the stack. The other points up. Together they reveal a bet most people are misreading.

The conventional take is that AI competition is about models. Bigger parameters, better benchmarks, lower latency. Anthropic is making a different wager. The company is quietly assembling control over the layers above and below the model, the connective tissue where developer time and enterprise budget actually get spent. This is not a model play. It is an infrastructure play disguised as one.

Here is the framework that explains what Anthropic is actually doing, why it matters for every builder in the ecosystem, and where the real risks hide.

The Sandwich Strategy

Call it the Sandwich Strategy. Anthropic is compressing its competitive position by owning the bread on both sides of the model layer.

Bottom slice: developer tooling. According to TechCrunch's May 18, 2026 report, Stainless software was used by OpenAI, Google, Cloudflare, Runway, Replicate, and hundreds of other organizations. Every official Anthropic SDK since the earliest days of its API was generated by Stainless. Now that toolchain is going in-house. Anthropic told TechCrunch it will wind down all hosted Stainless products. Competitors retain ownership of the SDKs they already generated, but they lose the central hosted tooling that kept those libraries current.

Top slice: enterprise deployment. The KPMG alliance announced May 19, 2026 embeds Claude inside Digital Gateway, the platform KPMG's people and clients use for tax, legal, and advisory work. Bill Thomas, KPMG's Global Chairman and CEO, framed it as a commitment to "responsible AI, prioritizing security, trust, and governance." Anthropic named KPMG a preferred partner for private equity. The two companies will co-build Claude-powered products for PE portfolio companies.

The model itself is the filling. Important, but not the moat. The Sandwich Strategy says: own what developers touch first and what enterprises buy last. Everything in between becomes harder to displace.

The Infrastructure Inversion

This is where the strategic picture gets interesting, and where most analysis falls short.

Katelyn Lesse, Anthropic's Head of Platform Engineering, said it plainly: "Agents are only as useful as what they can connect to." That single sentence explains the Stainless acquisition better than any financial model.

Consider the sequence of events. Four days later, on May 18, it acquired the toolchain that generates MCP servers and SDKs. The billing layer and the developer layer are now controlled by the same company. MCP server generation and Claude Agent SDK ergonomics will align by design rather than by community effort.

This is what I'd call an infrastructure inversion. Normally, infrastructure providers are neutral. They serve everyone equally and compete on reliability and price. Stainless operated that way for years. Now a model company owns a piece of shared infrastructure and has removed it from the commons.

The asymmetric advantage is real but narrow. Anthropic can ship new API features with polished, multi-language SDKs and MCP connectors faster than anyone else. Competitors face more friction. On May 5, 2026, Anthropic launched ten ready-to-run agent templates for financial services: pitchbook generation, KYC file screening, month-end book closing. The Stainless acquisition makes that integration pipeline proprietary.

But here is the contrarian case, and it deserves serious weight. Stainless is one layer of the stack. It does not give Anthropic control over IDEs, CI/CD pipelines, vector databases, observability tools, or cloud deployment. AWS, GCP, and Azure still own the compute substrate. Open-source alternatives exist: OpenAPI Generator, Swagger Codegen, Fern, Speakeasy. For plenty of developers, a thin wrapper around fetch or httpx with typed data objects is good enough. The notion that Anthropic now "owns the SDK layer" probably overstates the lock-in.

There is also a signal risk worth watching. By pulling a shared tool out of the ecosystem, Anthropic tells every future tool vendor and open-source community that collaboration might end in acquisition and shutdown. That could make partners more cautious. OpenAI, Google, and Cloudflare are almost certainly already evaluating or building alternative SDK generators. The competitive response may arrive faster than the strategic advantage compounds.

The KPMG side carries its own tension. KPMG has alliance relationships with Microsoft, Google Cloud, AWS, and ServiceNow. These are channel partnerships, not exclusive lock-in. A State of Tech Talent 2026 survey of 400 global IT leaders found that 97% of organizations are committed to implementing AI. But 48% cite security concerns as the top obstacle, up from 17% in 2024. And 57% report a significant capacity gap in security and risk management for AI. KPMG sells risk frameworks and compliance mapping. Anthropic's Constitutional AI positioning gives KPMG a "safer" LLM to recommend in regulated industries.

My read: the KPMG alliance is a distribution hack, not a technology moat. It bypasses traditional enterprise sales cycles by embedding Claude where 276,000 professionals already work. If even 20% become active users in year one, that is 55,000 people building muscle memory with Claude inside their daily workflows. Switching costs compound from habit, not from technical lock-in.

The real question is whether Anthropic can maintain both slices of the sandwich at once. Combining core research, cloud platform partnerships with AWS and GCP, internal developer tools from Stainless, and enterprise go-to-market via KPMG creates genuine management complexity. If OpenAI or Google remain ruthlessly focused on model quality and minimal-friction APIs, Anthropic's broader bet might actually slow it down. Whether vertical integration accelerates or fragments a company at this stage of the AI race is genuinely unclear.

2031

Three signals inside the same shift

Zoom out five years. Three structural forces are colliding, and the Stainless and KPMG moves only make sense when you see all three.

First, the model layer is commoditizing. In 2024, GPT-4 felt untouchable. By May 2026, Anthropic is valued at $61 billion in a round led by Thrive Capital and Andreessen Horowitz, according to Bloomberg's May 13 report. The company rejected a $56 billion offer, calling it "neither credible nor attractive." It is reportedly in early talks to raise at least $30 billion more. That capital is not going into model research alone. It is going into manufacturing, infrastructure, and the tooling layers that create switching costs.

Second, enterprise AI spending is entering its infrastructure phase. IDC estimated global GenAI software and services spending will exceed $90 to $100 billion annually by 2026. McKinsey and IDC both note that 10 to 20% of enterprise AI spend goes to tooling, integration, and developer experience rather than model access. That is a $10 to $20 billion annual segment. The companies that own the integration layer will capture margin that model providers cannot.

Third, the agent economy needs plumbing. Agents that can read documents, call APIs, query databases, and take actions require connectors. Lots of them. The company that controls how those connectors get built, maintained, and billed has a compounding flywheel. Every new MCP server makes the platform stickier. Every enterprise deployment creates demand for more connectors. This is the Costco hot dog logic applied to AI infrastructure: the loss leader is the model, the margin is in everything around it.

By 2031, I think the AI industry will settle into three tiers. Model providers who only sell inference will compete on price and lose margin every quarter. Platform companies that own tooling and distribution will capture the integration tax. And a small number of full-stack players will own the sandwich: developer entry point, model layer, and enterprise exit point. Anthropic is betting it can be in that third tier.

The risk is impermanence. Today's strategic advantage is tomorrow's open-source project. The 70% rule for decision velocity says: if you have 70% of the information you wish you had, move. Anthropic is moving. Whether the remaining 30% breaks for or against them will determine if this is remembered as a brilliant infrastructure play or an expensive distraction.

What to Build This Weekend

You do not need to wait for Anthropic's strategy to play out. The Sandwich Strategy reveals where value is accumulating, and you can position yourself there today.

Step one: audit your SDK dependencies. If you use Stainless-generated SDKs for any API, check whether you have the source code locally. Anthropic confirmed existing customers retain full rights to modify and extend their generated libraries. Fork them now. Do not wait for hosted tooling to disappear.

Step two: explore MCP server creation. Anthropic's agent templates for financial services show the pattern. Pick one internal tool or data source your team uses daily. Build a simple MCP server that lets Claude query it. The Anthropic docs walk through this step by step. You do not need a computer science degree. You need one afternoon and a willingness to break things.

Step three: map your own sandwich. Ask yourself: what is the developer entry point for your product or service? What is the enterprise exit point? Where is the connective tissue between them? If you do not control at least one slice, you are the filling, and fillings get replaced.

The evidence suggests the next 18 months will reward builders who think in layers, not features. Get your reps in. Build one tiny integration. See what connects.