The model layer commoditised. Your advantage moved to the wiring.

Two stories from this quarter look unrelated until you put them side by side. SAP agreed to acquire the Freiburg startup Prior Labs and committed to spend more than a billion dollars over four years building it into a frontier lab specialised in structured business data. In the same window, the gap between the major foundation models narrowed again — Google shipped Gemini 3.5 with a two-million-token context window and a deep-reasoning mode, and the practical difference between the top models on most real workloads kept shrinking. One company is paying a fortune to own a model. Everyone else is discovering that the model is the part of the stack that matters least.

This is the defining commercial shift of 2026: the model layer has commoditised, and the durable advantage has moved to the wiring around it. If you are deciding where to spend the AI budget right now, the integration layer beats the model upgrade nine times out of ten — and the vendors know it, which is why every foundation provider is racing to own the delivery work that turns their API into a business outcome.

Why the model stopped being the differentiator

Three things happened at once. The frontier models converged — for the overwhelming majority of business tasks, the second- or third-best model is indistinguishable from the best in the output that reaches a customer. Inference cost fell again, another sixty to seventy-five percent over twelve months, so the model is no longer the largest line item in a build; engineering time, evaluation tooling, and observability are. And the interoperability layer matured: Model Context Protocol has quietly become the de-facto way to connect an agent to tools and data across providers, so a tool definition written once now runs against several models.

Put those together and the strategic implication is stark. If the model is interchangeable, cheap, and reachable over a standard interface, then nothing about your choice of model is a moat. What is a moat is the thing only you can build: AI wired correctly into your systems, reading from your CRM, writing to your ERP, routing the easy work to a cheap fast model and the genuinely ambiguous decisions to an expensive smart one, and passing clean audit logs back to your compliance team. That is not a model. That is integration, and it is specific to you.

The vendor-as-consultancy move, and what it costs you

SAP buying a lab is one version of the response: own the model, specialise it for your domain, sell it bundled. The foundation vendors are running a complementary play — building and buying delivery capacity so they can sit next to their API and do the integration work themselves. For buyers this is genuinely useful. The one-throat-to-choke bundle of platform plus consulting is now a real option from each major vendor, and for some organisations it is the right call.

It comes with a structural cost worth naming. A delivery team owned by a model vendor will, by the gravity of its own incentives, recommend solutions built on that vendor’s model — and the integration it builds will tend to assume that model stays. You are buying convenience and paying in portability. In a world where the model layer is the commodity and the interop layer is mature, designing your most durable asset — the wiring into your systems — to be quietly dependent on one vendor’s model is optimising against the actual trend. The whole point of the mature interop layer is that you no longer have to.

What the integration-first build looks like

• Model-agnostic by design. The tool definitions, the prompts, the evals, and the orchestration are yours and portable across providers. Swapping the model underneath is a configuration change, not a rebuild.
• Routed, not single-model. Cheap fast models handle the easy eighty percent; the expensive model is reserved for the genuinely ambiguous calls. Teams who route run production AI at thirty to sixty percent of single-model cost with no quality penalty.
• Deep into the real systems. The value is in reading and writing your actual schema — not a generic connector that ignores the seven custom fields your business runs on.
• Auditable by default. The logs the compliance team will ask for in 2027 are produced from the first commit, not retrofitted after the first incident.

This seam — between a commodity model and your specific systems — is exactly where a Cravings AI Integration engagement lives: RAG over your real corpus, copilots inside the tools your team already uses, model-routed pipelines into your CRM and finance stack, and the whole thing built so the model underneath is a swappable part rather than a dependency you cannot escape. Yours, in your accounts and repos, owned by your team.

The honest counter-argument

The case against this is the case for the bundle, and it is not nothing. Vendor lock-in is sometimes a fair price for speed and a single accountable supplier — if your AI is not strategic, if you want it running this quarter and do not intend to maintain it deeply, the platform-and-consulting package will get you there faster than an integration-first build. And SAP’s bet is not stupid either: for structured business data specifically, a specialised model may genuinely outperform a general one for years. The point is not that vendor models or vendor delivery are wrong. It is that the choice between platform-aligned and independent is now a deliberate strategic decision the vendors are forcing by acquiring the alternative — and it should be made on purpose, documented, and revisited annually, not defaulted into because the bundle was easiest to buy.

What to do in the next 30 days

• Stop shopping for the best model. Pick a capable one and put the energy into the integration. The model rankings will have changed again by the time you ship.
• Audit your AI contracts for portability. The interop layer is mature enough that switching is real. Renewing without negotiating portability is paying for lock-in that no longer needs to exist.
• Cost out routing. If you are running everything on one expensive model, a routed fleet is probably your biggest cost win this year.
• Make the platform-aligned vs independent call on purpose. Write it into the strategy document and revisit it annually. The vendors are forcing the decision; make it deliberately.

SAP can afford to buy a lab. You do not need to, and you should not try. The advantage that is actually available to you is not a better model — it is your business, wired to AI more thoughtfully than your competitors wired theirs. That advantage is portable, it compounds, and no vendor can sell it to you bundled.

Spending the AI budget on the model when the leverage moved to the wiring? Talk to Cravings about an integration-first build — model-agnostic, routed for cost, deep into your real systems, and auditable from the first commit. The model stays swappable; the advantage stays yours.