TORONTO — Tenstorrent has developed a surprising side-hustle licensing its AI compute technology to countries and companies that are trying to avoid getting locked into chip giants like Nvidia and Arm. 

Customers using the Canadian AI hardware startup’s intellectual property include South Korea’s Hyundai Motor and LG Electronics, Swiss consortium SingularityNet as well as a research centre backed by the Japanese government. They’re turning to Tenstorrent in pursuit of more control, better performance and savings, according to the company. “I’m targeting organizations that want to own their own computing roadmap,” said chief customer officer David Bennett.

The AI boom has driven huge demand for the hardware that fills the data centres and devices on which the technology runs. Many developers favor Nvidia’s graphics processing units (GPUs) to train their models, driving the firm from Santa Clara, Calif., to a US$3.5-trillion market capitalization. Intel and British firm Arm control the architectures—sets of instructions that guide where processors go and data flows—that chip designers most widely use.

Tech giants Amazon, Google and Microsoft have all built their own hardware for their cloud arms, which dominate the market for processing power. Apple and Tesla have also developed custom chips for laptops and autonomous vehicles, respectively. Tenstorrent CEO Jim Keller led both those efforts.  

The startup, founded in April 2016, now sees an opportunity to work with firms and governments taking a similar approach. Chip shortages and growing geopolitical risks have alerted firms to the issues in their hardware supply chains, while countries like Canada are trying to build “sovereign compute” capacity within their borders to retain companies and sensitive data within their borders.

Tenstorrent sells clients the IP for its high-performance computing and AI chips, as well as its architecture, based on an open-source system called RISC-V. The startup earns royalties as customers produce hardware using its technology. Tenstorrent also sells design services. “We’re going to bring down the time-to-market for new custom silicon, and the cost,” said Bennett. 

Tenstorrent’s chips are specifically designed to train and run AI—many hardware startups focus on one or the other—and can be strung together to handle large workloads. Firms can get better performance than off-the-shelf chips by designing custom hardware specifically for the kinds of AI they sell, said Gaurav Gupta, a vice-president of research and advisory at Gartner.

Buying technology from Nvidia and Arm is also expensive. The savings from developing chips  in-house can be “huge,” said Gupta, noting that the chip giants’ prices include significant margins. But he argued custom hardware only makes sense for firms that need huge volumes to fill, say, massive cloud data centres or fleets of self-driving cars. That may limit the market for companies looking to license their technology.

Tenstorrent didn’t set out to be in the custom chip business. “We were building this IP for our own products,” said Bennett. The company has sold complete desktop machines and servers to firms in the data-centre and automotive sectors in Asia, Europe, and North America. It’s now receiving interest from the Middle East. 

Tenstorrent eventually saw an opportunity to turn its IP into another revenue stream because customers were asking for it. It also offers its software under open-source terms.

The firm is currently in technology-licensing discussions with both major firms and government-backed organizations, according to Bennett although he did not provide specifics. “They’re looking for a North American company that has experience with high-performance compute and AI,” he said. Tenstorrent, which has over 550 employees, most recently raised US$100 million last August.

Hardware developers using RISC-V are pushing into the server business, but it’s unclear whether they’re gaining much market share from the widely used Intel or Arm architectures, Gupta said. And lots of startups are lining up to sell chips to governments building sovereign compute capacity.

Tenstorrent isn’t just trying to be a cheaper substitute for dominant hardware providers like Nvidia, but to give customers a chance to control their own processing power, Bennett said. “You have a single company who’s monopolizing the market and taking the vast majority of the profit. That’s ripe for disruption.”