You'd Never Run Two ERPs. Run Multiple AI Models Anyway.

Twenty years of enterprise IT discipline can be summarized in one rule: one of each. One ERP, one CRM, one CMS. Running two of anything meant a merger nobody finished integrating or a migration somebody abandoned, and either way it was a problem to fix. Consolidation was what competent looked like.

Last week the CEO of Microsoft went on CNN and told enterprises to do the opposite with AI models.

On Fareed Zakaria GPS, Satya Nadella laid out a warning that would sound extreme coming from an analyst, let alone from the man whose company sells Copilot seats by the hundred million: "Any firm that doesn't have this control, I will claim will not remain a firm because you've essentially outsourced your thinking."

The control he means: your prompts, your metadata, your context — retained by you, portable across models. His prescription, in his own words: "By keeping the harness separate from the model and the context and memory separate from the model, you absolutely can use multiple models for what they're great at."

The vendor is telling you to keep the vendor swappable. That deserves a closer look — both at why he'd say it and why he's right.

What did Nadella actually say about AI lock-in?

Three things, and they stack.

First, the survival claim: companies that rely wholly on a single proprietary AI lab have outsourced their thinking, and firms that outsource their thinking stop being firms. Second, the data claim: "Every time you use the model, all of the metadata around it is retained by you, so that you could use all of that to train perhaps your own weights or your own open model." Your usage history is an asset — if you're the one holding it. Third, the architecture claim: put a gateway between your business and the model, keep the harness and the memory on your side of it, and model choice becomes a routing decision.

Is he talking his book? Obviously. Microsoft hosts every major model family on Azure and wins whenever models commoditize into interchangeable line items on a cloud bill. A world of portable prompts and swappable models is a world where the margin lives in infrastructure — which is to say, with Microsoft.

The motive is worth noticing for what it confirms, though: the company with the most complete view of enterprise AI consumption just told you which side of the architecture the leverage sits on. When the house explains how the odds work, listen.

Why AI lock-in works differently from ERP lock-in

The one-of-each rule was always about state, whatever the licensing spreadsheets said.

An ERP is where your general ledger lives. A CRM holds every customer relationship you have; a CMS holds a decade of content, workflow, and permissions. Migrating any of them means moving state — years of accumulated data, integrations, and process wired into every corner of the business. That's why those migrations run eighteen months and blow their budgets, and it's why running two of them was organizational malpractice. The lock-in was structural. The vendor didn't even have to try.

Now look at what a model actually holds: nothing. A model is stateless reasoning-for-hire behind an API that looks nearly identical from vendor to vendor. The prompt goes in, the answer comes out, and everything that made the exchange valuable — the context you assembled, the instructions you refined, the workflow it feeds — exists on your side of the wire. Swapping ERPs is a two-year program. Swapping models, in a sanely built system, is a config change and an eval run.

The market keeps demonstrating why you'd want that config change available. Frontier leadership has changed hands repeatedly in twenty-four months. Capability leaks through the very API that sells it and reappears at commodity prices within months. And availability is a dependency you don't control — a model your workflows depend on can disappear overnight, as everyone running Fable learned in June when an export-control directive switched it off mid-quarter. Consolidating on the current leader means consolidating on a position in a race, and the position changes faster than an enterprise procurement cycle.

So the old rule inverts cleanly. Systems of record: one of each, because state has gravity. Models: several, because there's no gravity to fight — match the model to the workload the same way you'd match compute to the job, and re-match when the leaderboard moves.

Which raises the obvious question. If models hold no state, why does AI vendor lock-in exist at all?

How companies get locked in anyway: rented surfaces

Because lock-in with AI has to be manufactured, and the manufacturing process is running at full capacity.

Every major AI vendor is racing to move up the stack from selling stateless inference to selling stateful surfaces: the chat workspace where your team's projects accumulate, the vendor-side "memory" that learns your preferences, the app platform where workflows get built inside the vendor's walls, the fine-tune that lives in one provider's infrastructure and runs nowhere else. Each feature is genuinely convenient. Each one relocates a piece of your state — prompts, context, institutional knowledge, workflow logic — from your side of the wire to theirs.

That's the rented surface. You don't own the floor your process is standing on, and every month of use pours more of your operational knowledge into a building you can't move out of. It's the same mechanism that made ERP migrations miserable, except this time the state gravity is opt-in — assembled voluntarily, one convenient feature at a time, by teams who'd never dream of running their general ledger in a vendor's chat app.

I've paid this tax personally. My health journey — years of accumulated context — is state stuck inside one frontier model, and I still haven't organized getting it out. Zenpo's business and sales knowledge got built after that lesson: LLM-agnostic, context on my side of the wire. I can swap frontier models whenever I want, or route it all to Kimi, open source's current darling, tomorrow.

The evaluation discipline here is the same one that governs any buy-versus-build decision: the question is what you stop owning when you build on the vendor's surface, however good that surface is. Renting the reasoning is fine — that market is competitive and getting cheaper. Renting the surface where your prompts, context, and workflows accumulate is how a stateless commodity quietly turns back into an ERP.

Nadella's gateway point is the practical test. If there's a layer you control between your business and the model — one that holds the context, logs the metadata, and routes the calls — you're a customer. If the vendor's surface is the system, you're a tenant.

What a multi-model AI architecture looks like

The pieces are unglamorous, which is a good sign — durable architecture usually is.

A gateway you own, so every model call routes through your infrastructure and your prompts and metadata land in your logs, the asset Nadella says you'll want when training your own weights stops being exotic. The harness — task logic, tool definitions, workflow integration — written against your own abstractions, so a model swap touches configuration instead of application code. Context and memory in your data stores, assembled per-call, handed to whichever model currently earns the task. And standing evals, so "model B got better and cheaper" is a measurement you act on rather than a rumor you debate. This is ordinary systems work — the same integration engineering as any custom line-of-business build — and it's a fraction of what any company has spent making two systems of record talk to each other.

The broader ecosystem is already organizing around this shape. On July 27, Nvidia launched the Open Secure AI Alliance — thirty-plus companies, including Microsoft, IBM, Cisco, and Hugging Face, building an open, shared stack for AI security: identity, permissions, isolation, logging. The conspicuous absences were OpenAI, Google, and Anthropic — the closed-model labs. Read the room: the infrastructure layer is standardizing around portability and shared tooling, while the model layer keeps building taller walled gardens. Your architecture gets to pick which of those layers it depends on.

Is AI vendor lock-in real?

The dependency is real; the inevitability is manufactured. Models are stateless and swappable behind near-identical APIs — lock-in appears when your prompts, context, and workflows accumulate inside a vendor's surface instead of your own infrastructure.

Should enterprises use multiple AI models?

Yes. The one-vendor consolidation rule existed because systems of record hold state and migrations are expensive. Models hold no state you don't hand them, and capability leadership changes hands too often to consolidate on the current leader.

What is an AI gateway?

A layer you control between your applications and model providers. It routes calls, holds context and memory, and retains prompts and metadata in your logs — which keeps model choice a configuration decision and preserves your usage data as an asset.

The one-of-each rule was right for its era, and the reason it was right is the reason it inverts now: it was always about who holds the state. Keep the state, rent the reasoning, and no AI vendor can lock you into anything. So run the audit question in your next architecture review: how many of your AI workflows could change models by Friday without a migration project? Whatever that number is — that's your real exposure, and every rented surface you adopt moves it the wrong way.