Don't Marry One Model

The AI vendor pitch used to be simple. Pick a lab, build on it, go. That advice just got risky, and the person telling you so is the guy who owns big stakes in the two biggest labs. Let me catch you up on what changed and what it means for your roadmap.

The warning from inside the tent

Satya Nadella went on CNN and said the quiet part loud. Companies that hand everything to one AI lab, their data, their prompts, their coding tools, will not survive as firms. His words: if you outsource your thinking, you stop being a company. He wants you to keep your usage data so you could train your own model later.

Read the self-interest, because it is real. Microsoft invests in both Anthropic and OpenAI, and it now sells the exact middleware he is telling you to buy. His own writer at GeekWire admits we all blame everything on AI and that the strategy is more tangled than "AI or bust."

Still, the warning holds up. He is pointing at two things to keep separate from any model: your context and memory, and your coding tools, what he calls harnesses. Claude Code and Codex are the examples he named.

Keep the harness, swap the engine

Runway is building for a world where you never bet on one model. Its new Media Router picks the best image, video, or audio model for each request based on whether you want quality, speed, or cost. Their pitch says it plainly: rather than asking developers to bet on a single model staying ahead, the router assumes the best model keeps changing.

That matters because Runway's own video models slipped off the top rankings. Google, ByteDance, and Alibaba now fill the top 20. So Runway would rather be the orchestration layer than lose. Chief product officer Anthony Maggio also flagged a live concern: some businesses do not want Chinese models, so they can set a preference for American providers.

The lesson for you is not about video. It is the shape. A routing layer between your product and the models means any one model can go away and you keep running. That is Nadella's point, sold as a feature.

The agents are moving onto the desktop

While leaders debate strategy, the labs are racing to sit inside your team's daily work. Perplexity brought its Personal Computer tool to Windows, a local agent that reads your files and works across Office 365. It starts at $200 a month per seat. OpenAI put Voice into its desktop app so you can talk to it and direct multiple agents in one command.

Amazon is doing the same in the home, teaching Alexa Plus to route requests to the right device and adopting the Model Context Protocol, an open standard so more apps plug in. Not every push lands. OpenAI's $230 Micro keypad drew a Reddit review calling it "a prank and not a real product."

The pattern under the noise: agents want deep access to your machines and files. That is exactly where Nadella says the risk grows.

Buying the bridge, not just the model

The enterprise winners are pairing models with people who know your industry. Anthropic expanded its deal with Cognizant, which has trained more than 30,000 associates on Claude and is embedding it across its platforms. The results they cite are concrete: a contract system that cut review time by up to 40 percent, a risk tool saving underwriters roughly eight hours a week.

Cognizant CEO Ravi Kumar S named the real gap. AI capability is rising faster than companies can absorb it. His firm sells itself as the bridge, the industry context and trust frameworks that turn a model into a production outcome.

That is a useful frame for your own budget. The model is the cheap part. The integration, the guardrails, and the people who know your rules are what actually ship value.

The deep cut

Nadella's warning gives you a checklist, not a vibe. Ask what your team hands to a model and whether you keep it. Keep your context and memory separate from any model. Keep your coding agent separate from the lab that made it, so you can switch. Retain the metadata from every call, because that is the raw material for your own model later. This is not a someday project. Every agent you let onto a laptop or into a codebase raises the switching cost. The cheapest time to build a routing layer and a data-retention rule is before you have ten teams locked into one harness. Bring a one-page answer to your next review: for each AI tool in flight, name what leaves the building and what stays.

Three questions for your team

• For each AI vendor we use, what data and metadata leaves our control, and could we switch models next quarter without a rebuild?
• Are our coding agents tied to one lab's harness, and what would it cost to swap the model underneath them?
• Before we let a desktop agent touch local files and email, who signed off on the access, and where does that company's data end up?