‘A multi-model world’: Microsoft’s CEO says the future of AI is orchestration, not one single model

“There are going to be multiple models and the trick is really how you take advantage of them and, in fact, build your own model by distilling them,” he said at the World Economic Forum. “Think of these models that you orchestrate to build your own model, and more importantly, do what could be described as harness engineering.”

By “harnessing engineering”, he means companies stitching together a mix of closed and open-sourced models, layering them with their own data and shaping the whole stack to deliver a specific business outcome. 

“You take that outcome and then you can say ‘can i use all the models’, orchestrate them and feed in my context and then as a result of it the reasoning traces are really leading to some capability and models that I control as my IP,” Nadella continued. “As long as firms can answer that question, they’re going to be getting ahead.”

For advertising, the point lands clean. The conversation around AI keeps collapsing into one familiar storyline: the rise of large language models — the ChatGPT-style tools that analyze information, draft content, write code and organize data. Nadella’s message is a reminder that AI is far bigger than that narrow slice of technology, and that the real advantage will belong to companies that treat it as an ecosystem, not a single product.

“The reality is it’s a multi model world,” he added. 

It is, of course, easier to preach multi-model mastery when you run one of the companies building the models. For an industry still wrestling with legacy systems, thin margins and basic measurement problems, the path from Nadella’s vision to day-to-day reality looks steeper.

Even so, the direction of travel is becoming hard to ignore, and ad execs are slowly realizing they have little choice but to follow.

Nowhere is that clearer in advertising than inside ad tech, where there’s a growing recognition that LLMs are often not the right tool for the job. Real-time bidding at bidstream scale runs on predictive filtering and control-style optimizations — narrow, rule-bound systems built for speed in a brutally competitive market. LLMs can be powerful layers, translating marketer intent and managing workflows. But they’re not a practical replacement for a bidding engine, not philosophically, and certainly not when latency, reliability and until unit economics are measured in microseconds and fractions of pennies.

In many ways, that tension is a microcosm of a broader reality across advertising.

At CES earlier this month and again at the Digiday Programmatic Marketing Summit the month before, agencies and marketers delivered a similar verdict: sometimes the smartest answer is not an LLM at all. It might be a simple regression model, a decision tree or a few lines of rule-based logic. Even within the LLM universe there is nuance. While most Western AI forms chase ever-larger proprietary models locked inside corporate clouds, China has surged ahead on open-source models that developers can adapt and retrain.

“Choice depends on a whole variety of factors including input type, task type, data availability, latency, cost and so on,” said Ian Maxwell, CEO at ad tech business Converge Digital, which uses multiple models to connect advertisers to publishers.

A typical approach his business takes is two-layer validation training and evaluating multiple models or variants in parallel, then comparing them side by side to select the best option on performance, latency, robustness and other practical measures.

In other cases, the logic is simpler and more sequential. Converge’s Ultimate Brand Safety model, for instance, began with the simplest model class that suited the data and production constraints. Performance was “rigorously” validated, said Maxwell, with plans to move to more complex or expensive approaches only if the results fell short. The initial model cleared the bar, so there was no need to escalate further, Of course, evolution and enhancement will come but it’s not critical at this point, he continued.

The stakes for making these choices, argued Nadella, are bigger than efficiency. CEOs who fail to grasp this risk being disintermediated by rivals. That danger only grows if they do not control their own data. Or as Nadella put it “sovereignty” — the ability, he continued, to embed a company’s “tacit knowledge” into models it owns rather than leaking value to some outside platform provider. Where a data center sits matters far less, he continued, than who control;s the models and the intelligence inside them.

“When it comes to AI, this is the topic I think will be most talked about this year: the sovereignty of a firm,” said Nadella.

Get that right, and the rest can be engineered. Miss it, and the future belongs to someone else. 

“When it comes to AI, this is the topic I think will be most talked about this year: the sovereignty of a firm,” said Nadella.