Extropic, a hardware startup led by former members of Alphabet Inc.’s quantum computing research team, today announced that it has raised $14.1 million in seed funding. 

Kindred Ventures led the investment. It was joined by HOF Capital, Julian Capital, Marque VC, OSS Capital, Valor Equity Partners and Weekend Fund. Extropic says the round also saw the participation of more than a dozen other backers, including executives from Adobe Inc., Shopify Inc. and several venture-backed artificial intelligence startups.

Extropic was founded last year by Chief Executive Officer Guillaume Verdon, who previously led a quantum computing team at Alphabet’s X research unit. The company’s chief technology officer, Trevor McCourt, was also a researcher at the search giant. While at Alphabet, Verdon and McCourt led the development of a TensorFlow library that can be used to run AI models on quantum computing chips.

It’s believed Extropic is building a chip optimized to run large language models, or LLMs. In a blog post today, Verdon described the company’s technology as a “novel full-stack paradigm of physics-based computing” and detailed that it harnesses “out-of-equilibrium thermodynamics.” This hints that Extropic’s chip design incorporates concepts from non-equilibrium thermodynamics, an emerging branch of physics focused on studying phenomena such as chemical reactions.

Verdon’s blog post indicated the company’s product is not a quantum computing chip. “As the timelines to scalability for quantum physics-based computers grew endlessly longer and longer, many of our team sought a different path to practical physics-based computing,” he wrote. 

One of the main reasons researchers have not yet succeeded in building a commercially useful quantum chip is that such processors are highly prone to computing errors, or noise. Those errors make it impossible to carry out complex calculations reliably. Extropic is seeking to build a system in which “noise is an asset rather than a liability.”

The company hasn’t shared more detailed technical information about its technology. However, Verdon did reveal that one of Extropic’s goals is to reduce the amount of electricity needed to run AI models. He also suggested the company’s technology will automate certain coding tasks, writing that “one could imagine a computer which, instead of being imperatively programmed, naturally finds a way to program itself to learn representations of the world.”

It’s possible to infer additional details about Extropic’s technology from the fact that it’s to run AI models.

Neural networks crunch data using mathematical calculations called matrix multiplications. Those are calculations that are carried out on matrices, mathematical structures comprising numbers organized into rows and columns like a spreadsheet. Because AI models use matrix multiplications extensively while processing data, practically all AI-optimized chips include circuits optimized to carry out such operations.

AI chips typically also include a large amount of high-speed memory. The reason is that AI models frequently move data to and from memory during processing. The faster the data can complete this round trip, the sooner a neural network can generate results for users.

If the report that Extropic is developing an LLM-optimized processor is accurate, the company can expect to face competition from Nvidia Corp. The chipmaker recently introduced a new data center processor, the H200, that is specifically built to run LLMs. It features twice as much onboard memory as Nvidia’s previous flagship graphics card. 

Image: Unsplash