Startup Founded by ex-Tesla Dojo Leaders Nears $10 Billion Valuation

DensityAI, an AI chip startup founded just a year ago by former leaders of Tesla’s Dojo supercomputer program, is in advanced talks to raise hundreds of millions of dollars in a round that would value it at $10 billion, according to two people with knowledge of the discussions.

The valuation is high for a company whose chips appear to be early in their development, meaning they could be years away from mass production. But the company’s leaders have told potential investors they’ve secured an agreement in which Amazon Web Services would purchase eventual chips from DensityAI if they meet certain performance requirements, the people said.

Andreessen Horowitz has been in talks to lead the round, the company’s leaders have said, according to one of the people.

The funding talks highlight investor interest in startups whose products aim to satisfy voracious demand for compute, even if the chips are far from ready.

Four-year-old Etched, which is also designing a specialized chip for AI, raised $700 million in August at a $21 billion valuation in a round led by Jane Street. Etched began shipping its chips over the summer, the company has said. And Andreessen Horowitz last month raised $1.1 billion for a fund to back companies building chips, memory and other physical components of AI.

DensityAI’s last valuation couldn’t be learned. It previously raised an undisclosed amount of money from Dolby Family Ventures, South Park Commons, Firestreak Ventures and other investors.

Spokespeople from Amazon and Andreessen Horowitz declined to comment. A representative from DensityAI didn’t respond to a request for comment.

The startup was founded a year ago by Ganesh Venkataramanan, the former head of Tesla’s Dojo team; former Dojo chief system engineer Bill Chang; and former Dojo AI Infrastructure team lead Ben Floering. Tesla originally hoped that its Dojo supercomputer would help to run the AI software behind Tesla’s self-driving vehicles, but Musk’s electric car company temporarily disbanded the group last summer.

DensityAI has told potential investors it’s taking a unique approach to arranging memory in its chips to make them faster and more energy-efficient for running AI, one of the people with knowledge of the fundraise said. It’s aiming to use 3D dynamic random-access memory stacking, an emerging approach that stacks memory cells or dies directly on top of the part of the chip that performs the calculations rather than next to it, one of the people said. This arrangement reduces the distance that data need to travel during inference.

Chip startup d-Matrix also uses this 3D DRAM stacking in its second-generation chip and plans to improve upon it in future generations. That startup, which has been selling its chips to customers, last raised $275 million at an approximately $2 billion valuation but is in talks to fundraise again, The Information has reported. Nvidia is also exploring new memory technologies, in part via a $500 billion collaboration with memory maker SK Hynix.

There are major downsides to 3D DRAM stacking. Most notably, stacking extremely temperature-sensitive DRAM on top of ultra-hot compute is very difficult to accomplish, because the fragile layers could then warp or break. If DensityAI is able to work around this issue, its approach could help AI developers make running their models cheaper and faster.

Amazon has developed custom AI chips including Trainium, which aim to be more cost-effective for running and training AI models than Nvidia’s GPUs. But the cloud giant has also signed deals with multiple hardware startups to complement its in-house silicon efforts. In February, the cloud provider signed a deal with Astera Labs to purchase its specialized semiconductor-connectivity hardware for its Trainium chips.

In March, Amazon and chip designer Cerebras signed a deal for the two companies to connect their server chips to each other to run different parts of AI inference across the hardware, an approach known as disaggregated inference. It’s not clear whether that has happened yet.