In 2017, Apple shipped the A11 Bionic with a 2-core Neural Engine capable of 600 billion operations per second: the first dedicated on-device AI processor in a consumer smartphone. It was announced as a performance feature. Face ID. Animoji. Photo processing. Tasks that need fast inference but not a server farm.
The cost calculation behind that decision has never been publicly stated. Apple has 2.5 billion active devices. If every AI inference required a cloud round trip, the infrastructure cost would be extraordinary. On-device inference has no per-query cost. The compute is paid for when the customer buys the device.
The Neural Engine went from 2 cores in 2017 to 16 cores by 2020. Performance went from 600 billion operations per second to 38 trillion in the M4 in 2024: a 63x improvement in seven years.
Once the infrastructure existed, Apple turned it into a marketing asset. At CES 2019 they ran a billboard: "What happens on your iPhone, stays on your iPhone." It wasn't a privacy policy. It was a product differentiation strategy made possible by hardware that already existed.
The average iPhone sells for $940. The average Android device sells for $380. Apple captures 46% of global smartphone revenue while shipping roughly 25% of units. Privacy is part of how that premium gets justified.
The largest on-device AI deployment on earth runs inside a device that has never been marketed as an AI product, by a company that has never needed a term for what they built.
Pete Bernard
xSilicon Valley, xMicrosoft, and an independent leader in edge AI
Visit my website
https://www.linkedin.com/in/bernardpete/
#edgeAI