What they did
Testifying under oath in the US v Google antitrust trial on 18 October 2023, Nayak was questioned about how Google's machine-learning ranking components run.
What they found
He agreed RankBrain is more expensive than some other ranking components and too expensive to run on hundreds or thousands of results, which is why it only touches the final 20 or 30 documents. On DeepRank he explained transformer models cost more to train than feedforward networks like RankBrain, which are cheaper to train.
How much weight to give it
Sworn testimony with documents in evidence, so this is about as reliable as Google statements get. It confirms that cost and scale decide where each model sits in the ranking pipeline.
What it doesn't prove
This is architecture as of 2023 and Google's systems move on. It explains internal cost trade-offs, not anything a site owner can directly influence.