This article was AI-generated as part of an experimental historical-content project. The date reflects the period being analyzed rather than the date the article was originally written.
Cloud contracts usually read as procurement news. Perplexity’s agreement with Microsoft, reported by Bloomberg on January 29, deserves a second look from anyone who cares how companies are described in AI answers.
What was reported
According to Bloomberg, Perplexity made a three-year, $750 million commitment to use Microsoft’s Azure cloud, which lets it deploy AI models through Microsoft’s Foundry service, including models from OpenAI, Anthropic and xAI. A Perplexity spokesperson said the company was excited to partner with Microsoft “for access to frontier models,” that it had not shifted spending away from Amazon Web Services, and that “AWS remains Perplexity’s preferred cloud infrastructure provider.” DatacenterDynamics notes that this is Perplexity’s first cloud agreement outside AWS, and that it comes while Perplexity and Amazon are in a legal fight over the Comet browser agent making purchases on Amazon’s store.
So by Perplexity’s own account the servers mostly stay where they were. The deal is about models.
An answer engine is a router
From the outside, Perplexity looks like one product with one voice. In practice it is a retrieval system that finds sources, plus language models that read those sources and write the answer. Paying users have long been able to choose which model does the writing. The Microsoft deal, as reported, secures access to several leading model families through a single platform.
That matters for reputation in a way that is easy to miss. The same question about a company, sent to the same answer engine, can be written up by different models with different habits: how cautious they are with allegations, how they handle sources that disagree, how readily they fill a gap with a plausible guess. Retrieval decides which evidence is on the table. The model decides what the paragraph says about it.
The same few writers, everywhere
The wider pattern is concentration at the model layer. On Microsoft’s earnings call in late January, Satya Nadella said that over 1,500 Foundry customers have used both OpenAI and Anthropic models. Perplexity will now source several of the same frontier models through that platform. Many products that look independent to the people using them draw, underneath, on a small number of model makers.
For a company, that has two consequences. A flaw in how one model family tends to describe you can show up in several products at once, because they share the writer. And improving the public sources those models retrieve helps across all of them, for the same reason.
In early 2024 I wrote that Perplexity’s funding round was a bet that cited answers become the new homepage. The citations are still the visible part. The less visible part is which model is deciding what to say about them.
What to do with this
When you test how answer engines describe your company, record which model produced each answer wherever the product shows it. Differences between models are information, not noise.
When an answer is wrong, work out whether the cited sources are wrong or the model misread them. If the source is wrong, it can often be fixed or outweighed, and the fix carries across every product that retrieves it. If the model misread a correct source, the remedy is clearer wording and corroboration elsewhere, and patience. Only the first is fully in your hands.