Apple bought the same founder's company twice, twelve years apart

In July we wrote about the offices where people dictate prompts into foam masks, and called it the visible end of a $22 billion voice AI market. We left out the largest deal in that market. Apple had already spent $1.6 billion on making dictation silent, six months earlier.

How the technology got here

• November 2013. Apple pays about $360 million for PrimeSense, an Israeli company whose depth-sensing work became Face ID in 2017.
• 2018. Arnav Kapur demonstrates AlterEgo at the MIT Media Lab, a headset that reads the neuromuscular signals your jaw and throat make when you say a word to yourself, which a neural network then decodes.
• 29 January 2026. Apple confirms the purchase of Q.ai for $1.6 billion. The Israeli startup, founded in 2022 and never out of stealth, reads silent speech with an optical head aimed at the face. Around 100 employees move across.
• 23 April 2026. POSTECH in South Korea publishes a silicone neckband that turns mouthed words into synthesised speech.
• June and July 2026. Developers start wearing sound-muffling masks in open-plan offices, and we wrote about it.

The founder Apple has now bought twice

Aviad Maizels founded PrimeSense, sold it to Apple in 2013, and founded Q.ai in 2022. Apple has bought his company again, at 4.4 times the first price. Both purchases do the same thing: read a face and turn what it does into data.

The first one became Face ID and shipped in every iPhone from 2017. The second is expected in Vision headsets, AirPods stems and the smart glasses due late this year.

The number that measures the gap

Against the $22.5 billion voice AI market our July piece described, $1.6 billion is 7.1%. Against the more than $7 billion investors put into voice AI startups in the first quarter of 2026, it is 23%. One deal came to roughly a quarter of what the whole sector raised in three months.

We built that article around a mask made of foam. Apple was solving the identical problem with cameras, and had signed for it in January.

What the science can do

The POSTECH neckband reached 85.8% accuracy across 26 words from the NATO alphabet. Walking or moving the head dropped it to 39.72%.

“We hope this technology will accelerate the day when patients with speech disorders can reclaim their voices.”

— Sung-Min Park, New Atlas, 23 April 2026

Quote source: Professor Sung-Min Park, who led the POSTECH research, via New Atlas, 23 April 2026

A neural network trained on the wearer does the decoding in each of these designs. The second figure is the one that describes the field. Sit still and the device gets about six words in seven. Stand up and walk and it gets two in five, on a vocabulary of 26 words.

What $1.6 billion bought is the team that sold Apple Face ID twelve years earlier, and a patent on pointing light at a jaw. The finished product comes later.

None of this should be read as personalized investment advice.