2700 doesn't strike me as remotely crazy. I'm guessing that some of those projects all serve the same goal, e.g., for their speech system, they have: acoustic modeling; language modeling; named-entity recognition; intent classification; domain classification; grapheme-to-phoneme conversion; language detection; wake-word detection. This ignores other stuff that happens around speech (for example, I know they were using a CRF to label different types of numbers for training their spoken-to-written form converters, which AFAIK are still using WFSTs, although at this point I wouldn't be shocked if both of those systems were converted to DNNs). So let's take an estimate of 10 DNNs for their speech systems. Per language, so make that 200 DNNs to support 20 languages. This ignores that they have separate models for YouTube, voice search (one model for on-device and a cloud-side model), voicemail.
Their machine translation system probably has a similar # of DNNs, and there you have to deal with language pairs, rather than single languages. Let's call it another 400.
That's two side-projects. Then you pull in query prediction, driverless cars, all kinds of infrastructure modeling, spam detection, all of the billions of things that are happening in ads, recommendations, I haven't really even mentioned search yet... Honestly, if I'm right in assuming that the cited figure is really "# of DNNs that do different things", then I'm surprised it's not higher.
Their machine translation system probably has a similar # of DNNs, and there you have to deal with language pairs, rather than single languages. Let's call it another 400.
That's two side-projects. Then you pull in query prediction, driverless cars, all kinds of infrastructure modeling, spam detection, all of the billions of things that are happening in ads, recommendations, I haven't really even mentioned search yet... Honestly, if I'm right in assuming that the cited figure is really "# of DNNs that do different things", then I'm surprised it's not higher.