AI Chip Startup Positron Raises $230 Million From Arm, #Qatar to Compete With Nvidia - Bloomberg

AI Chip Startup Positron Raises $230 Million From Arm, Qatar to Compete With Nvidia - Bloomberg

AI chip startup Positron, which is trying to compete with Nvidia Corp., raised $230 million in a funding round from investors including Arm Holdings Plc and the Qatar Investment Authority.

The round values the company at more than $1 billion, including funds raised, Chief Executive Officer Mitesh Agrawal said in an interview. The round was co-led by Arena Private Wealth, formerly OCM Private Wealth, Positron customer Jump Trading, and Unless. Helena, as well as previous investors including Valor Equity Partners, Atreides Management and DFJ Growth, also participated.

Positron, based in Reno, Nevada, is trying to compete with Nvidia by offering energy-efficient AI chips for inference, or running AI models.

The company, which also has listed Cloudflare Inc. and Parasail as customers, is currently selling a first version of its product based on existing reprogrammable chip technology. The second version is being designed from the ground up, and the company expects to have that done by September or October.

“In an ideal world we want to match Nvidia on product cadence,” Agrawal said. “That’s one of the things that AI silicon companies haven’t done in the past.” Positron’s chip will have more memory attached than Nvidia’s Rubin chip, which comes out later this year, he said. That will make Positron’s product perform well when reasoning and video models answer queries, he said.

Agrawal said the company wasn’t trying to raise money until Jump, a proprietary trading firm, tested Positron’s first-generation product and became interested in its plan for the next chip. Jump’s chief technology officer suggested the company invest. Arena knew Agrawal from its investment in Lambda, a so-called neocloud company, where Agrawal served as chief operating officer.

Nvidia’s chips are the dominant products to training and run AI models, but that will change, said Alex Davies, Jump’s CTO.

“We don’t think there’s going to be one winner,” Davies said. “I don’t think in five years’ time there’s going to be one company making a piece of hardware that everyone does inference on.” Jump often has tasks where it needs access to more memory and where it can’t get endless amounts of power, making Positron’s design appealing. “We look a bit more like a customer that has constraints, rather than an OpenAI or an Anthropic who are just putting infinite power into buildings.”

Recent interest in companies that make AI chips that can perform some inference tasks better, faster or cheaper is boosting enthusiasm for AI chip makers like Positron. There’s also so much demand for semiconductors to power AI that companies like OpenAI are signing deals with various vendors. In December, Nvidia signed a license with AI chip startup Groq for a reported $20 billion and hired most of that company’s chip engineers, while last month OpenAI reached a multiyear deal to use hardware from Cerebras Systems Inc. for 750 megawatts’ worth of computing power.

The deals signal an opening for products other than Nvidia’s chips if they can offer an advantage in performance per dollar or performance per watt of energy. Buyers are now more willing to have AI data centers that aren’t all based on Nvidia’s chips.

“What folks are saying is, especially on the inference side, because the demand is growing so much, we are OK with heterogeneous architectures,” Agrawal said. “The workloads are so large that, if you can save whatever percentage, it’s worth it.”