Ten Years with a Single Thesis
Around 2015, Connor Zwick wrote a thesis to himself: AI would eventually replace the need for a human language tutor. He was wrong about the timing - but not by much. Eight years later, Speak processes over a billion spoken sentences a year, counts OpenAI as both investor and technology partner, and carries a $1 billion valuation. Zwick is mid-stride, and the finish line keeps moving.
This is not a rags-to-riches story. It's weirder than that. Zwick grew up in Waukesha, Wisconsin, and by age 13 was writing code well enough that an Australian tech publication called Nettuts.com tried to hire him as a writer - without realizing he was in middle school. They found out. They tried to hire him anyway.
In high school, he built Flashcards+, a study app that hit nearly a million downloads before he graduated. The app did well enough that both Harvard and Y Combinator accepted him in the same year - two institutions that have never been known to overlap. He chose both, briefly, then chose neither: he took a Thiel Fellowship in 2012, one of about 20 spots given annually to people under 20 who agree to pursue entrepreneurship instead of college.
He left Cambridge with a leave-of-absence letter in his pocket - his exact words: it was "like a safety net." He would later say he probably learned more after leaving than he ever did inside the classroom.
His first post-Harvard venture was the Coco Controller - a $35 gaming device that slid over smartphones, interfacing through the audio jack. It was thoughtful, technically interesting, and went nowhere. He sold Flashcards+, by then at 5 million users, to Chegg for an undisclosed sum in 2013 and became an advisor to the company.
Then he started learning. Literally. He audited machine learning courses at UC Berkeley and Stanford - sitting in on classes no one expected him to take, building the mental models that would eventually become Speak's core technology. This is worth pausing on: a teenager who'd already sold a startup was sitting in lecture halls, taking notes, not because anyone asked him to.
Seoul Has a Problem. Zwick Has a Company.
Speak was founded in 2016 with Andrew Hsu, a neuroscientist and fellow Thiel Fellow Zwick had met through the fellowship network. The two shared something unusual: both had been interested in artificial intelligence from the start, back when that interest made you look like a science fiction fan rather than a venture capitalist.
The launch strategy was deliberately counterintuitive. Instead of entering saturated Western markets, Zwick moved to Seoul. He spent years shuttling between San Francisco and South Korea, learning a market that was unlike anything he'd seen in the US. Korean students studied English for 15-plus years, typically. They could read it. They could pass tests. But they couldn't speak it. The gap between instruction and ability was vast and painful - a national frustration Zwick later called "the silent English problem."
The solution wasn't a better flashcard app. It was a different philosophy entirely. Speak uses AI to simulate real conversation - not grammar drills, not multiple-choice vocab tests, but the closest thing to sitting across from a native speaker that technology has yet produced. The platform processes voice, provides instant feedback, and adapts to each learner's specific gaps and patterns. Zwick's three-step model: immersive listening and speaking, targeted repetition, and real-world contextual application.
Speak launched in Korea in 2019. The traction was immediate. OpenAI's Startup Fund led a $27M Series B in 2022 - an unusual arrangement in which the company powering your AI is also your investor, and would later become a customer. By December 2024, Accel led a $78M Series C at a $1 billion valuation. Total raise: $162 million. Speak had become a unicorn by doing something unglamorous: actually teaching people to talk.
The Billion-Sentence Engine
The numbers are instructive. Speak processes over 1 billion spoken sentences per year - roughly 32 sentences every second, day and night, across 40+ countries. Average daily usage runs 10 to 20 minutes per session, which is long for a consumer app. The company has crossed $100 million in annualized revenue and was named to Forbes' AI 50 list in 2025.
Enterprise came later but grew fast. Speak now serves more than 500 companies - including KPMG and HD Hyundai - through a B2B product that provides employee language training at scale. The platform teaches English, Spanish, Japanese, French, Italian, and Korean. A US consumer market launch was planned for mid-2025.
Zwick's philosophy on AI product development is worth quoting directly: "You can't outsource a core competency of your company." And on building for mass consumers: "Productizing consumer AI is super hard." Both statements carry the weight of someone who built the thing, not someone who read about it.
He does not appear to be slowing down. At the ASU+GSV Summit in April 2026, he sat on a panel titled "At the Speed of AI - Personalizing Knowledge for 8 Billion People." The panel title is essentially Speak's mission statement in slightly more formal clothes.
The ambition is not modest. Zwick wants to build a superhuman tutor - not a supplement to existing education, but a replacement for the model that has left 1.5 billion English learners fluent in grammar exercises and speechless at the dinner table. The company has a long way to go. So does everyone else in this race. Zwick has been running it for a decade longer than most of them.