Machine learning demands computation so massive it drove Google’s data center investments.

Evidence

Google is known as a “full stack AI company” that uses its own data stores “to train its own algorithms running on its own chips deployed on its own cloud.” Its dominance is further strengthened by the fact that machine learning is only as intelligent as the amount of data it has to train on, and Google has the most data. By 2013, the company understood that its shift into the “neural networks” that define the current frontier of artificial intelligence would substantially increase computational demands and require a doubling of its data centers. As Urs Hölzle, Google’s senior vice president of technical infrastructure, put it, “The dirty secret behind [AI] is that they require an insane number of computations to just actually train the network.” If the company had tried to process the growing computational workload with traditional CPUs, he explained, “We would have had to double the entire footprint of Google—data centers and servers—just to do three minutes or two minutes of speech recognition per Android user per day.” — Shoshana Zuboff, The Age of Surveillance Capitalism, 591

Categorization

• SHARE: The specific number — doubling Google’s whole data center footprint for two minutes of daily speech recognition per Android user with traditional CPUs — is a striking concrete detail to send to someone working in infrastructure or ML.
• SEED: The tension between AI’s ‘dirty secret’ of insane computation and the environmental/silicon costs of scaling is a usable angle: ‘the hidden energy ledger of the AI revolution.’
• XREF: connects to: Machine learning underlies Big Data and drives today’s massive investment. — Nick Couldry & Ulises A. Mejias, The Costs of Connection: How Data Is…, 392