Yann LeCun once designed an AI chip called ANNA in 1992 while working as a researcher at Bell Labs outside New York City. He and several other researchers built this chip to run deep neural networks—complex mathematical systems that can learn tasks by analyzing vast amounts of data—but it never reached the mass market. Neural networks were effective at recognizing letters and numbers on personal checks and envelopes, but they didn't work practically for other tasks at the time. Today, however, neural networks are rapidly transforming the internet's biggest players, including Google, Facebook, and Microsoft. LeCun, who now oversees the central artificial intelligence lab inside Facebook, where neural networks identify faces and objects in photos and translate between languages, notes that twenty-five years later, the market very much needs chips like ANNA, and these chips will soon arrive in large numbers.

The challenge is that while Google, Facebook, and Microsoft can run their neural networks on standard computer chips known as CPUs, this approach is terribly inefficient. Neural networks run faster and consume less power when paired with chips specifically designed to handle the massive mathematical calculations these AI systems require. Google's own AI chip, the TPU, is widely deployed inside the massive data centers that underpin the company's online empire, where it helps with everything from identifying commands spoken into Android smartphones to choosing results on the Google search engine. By rolling out the TPU chip, Google says it saved the cost of building about 15 extra data centers. Now, as companies like Google and Facebook push neural networks onto phones and VR headsets to eliminate the delay from shuttling images to distant data centers, they need AI chips that can run on personal devices too. According to LeCun, "There is a lot of headroom there for even more specialized chips that are even more efficient."

Multiple major technology and chip companies are racing to build specialized AI hardware. As CNBC revealed, several of the original engineers behind the Google TPU are now working at a stealth startup called Groq. Intel, after acquiring the startup Nervana, is now building a chip specifically for machine learning. IBM is also developing hardware architecture that mirrors the design of a neural network. Qualcomm has started building chips specifically for executing neural networks, a project on which Facebook is helping the chip maker develop technologies related to machine learning, as confirmed by Qualcomm vice president of technology Jeff Gehlhaar, who says, "We're very far along in our prototyping and development." Meanwhile, nVidia recently hired Clément Farabet, who studied under LeCun at NYU, explored this kind of chip architecture, and went on to found the deep learning startup Madbits, which was acquired by Twitter in 2014.

Currently, nVidia dominates in AI through GPU chips used for training neural networks. Before companies like Google and Facebook can use a neural network to perform tasks like translating languages, they must first train it by feeding it enormous collections of existing data. As LeCun explains, "For training, GPUs basically have cornered the market, particularly nVidia GPUs." However, GPUs were not originally designed for AI—they were designed for rendering graphics. About five years ago, companies like Google and Facebook started using them for neural network training simply because they were the best available option. LeCun believes GPUs will continue to play this training role because "[GPUs] are going to be very hard to unseat, because you need an entire ecosystem." Yet he also believes that a new breed of AI chips will significantly change the way big internet companies execute neural networks both in data centers and on consumer devices, from phones to smart lawn mowers and vacuum cleaners.

Dedicated AI chips promise substantial improvements in efficiency as demand for AI services increases. In executing neural networks, they can burn less electrical power and generate less heat than general-purpose processors. As LeCun quips, "If you don't want to boil a small lake, you might need specialized hardware." The need for such efficiency extends beyond data centers to consumer devices as virtual and augmented reality become more pervasive. Phones and headsets will need similar specialized chips to handle the neural networks required for augmented reality systems that recognize the world around users, technology that Facebook has begun unveiling with its new augmented reality tools.