Yes, that's correct (although I mostly built the infrastructure for exacycle rather than doing the actual research). At one point I personally had control over more CPU cycles (choosing which jobs to run) than any other person on the planet. I also worked on the Ads Database and its related data processing systems, plus the Making in Science team (we built and ran the booth at the Maker Faire), created the predecessor to Google Accelerated Science, and spent years working on ML software and hardware.

My last work, unpublished (and probably gone forever) was using the idle time on TPUs to establish the using multiple large TPU and standard clusters to do multitask training on large multimodal corpus (the docjoins, labelled youtube) with the goal of creating a model that generated a image of a human head that could be interacted with verbally, and had enough context/state to pass any test humans give it (my prior is that you can make a non-sentient system that fools experts, not Blake, although I would be delighted to learn that merely making such a system generated a truly sentient life-form, it seems like that would be truly impossible to demonstrate convincingly).