An AI chip startup, DensityAI, founded a year ago by former core executives of Tesla Motors (TSLA)'s Dojo supercomputer project, is in late-stage financing talks to raise several hundred million dollars at a post-money valuation of $10 billion, according to two people familiar with the matter.
The company's chips are still in early development and may be years away from mass production, yet it has secured such a high valuation.
The people said management has told potential investors that a deal has been finalized: if DensityAI's chips meet specified targets. One of the people said Andreessen Horowitz (a16z) is in talks to lead the round.
The financing negotiations reflect that even when chip products are far from mature, investors remain bullish on chip startups that can meet enormous computing power demand.
Etched, which is also developing specialized AI chips and was founded four years ago, completed a $700 million funding round in August this year led by Jane Street Capital, at a post-money valuation of $21 billion, and the company said it began shipping chips this summer.
In addition, a16z completed fundraising for an $1.1 billion fund last month dedicated to investing in chip, memory, and other AI hardware component companies.
DensityAI's previous valuation was not disclosed. The company had previously received investment from Dolby Family Ventures, South Park Commons, Firestreak Ventures and other institutions, with the funding amount undisclosed.
Spokespeople for Amazon and a16z declined to comment; a representative for DensityAI did not respond to a request for comment.
The startup was co-founded a year ago by Ganesh Venkataramanan, former head of the Dojo project at Tesla Motors (TSLA), Bill Zhang, former chief systems engineer for Dojo, and Ben Floering, former head of the Dojo AI infrastructure team.
Tesla Motors (TSLA) originally hoped the Dojo supercomputer could support the AI software behind autonomous driving, but the Musk-owned automaker temporarily disbanded the project team last summer.
A person involved in this funding round said DensityAI has told potential investors that its chips use a unique memory layout that can improve AI operating speed and reduce power consumption.
The company plans to use 3D DRAM stacking technology: stacking memory cells or memory dies directly on top of compute units, rather than the traditional side-by-side layout. This architecture can shorten the data transmission distance during inference.
Chip startup d-Matrix's second-generation chip also uses 3D DRAM stacking and plans to continue optimizing it in subsequent iterations. The Information previously reported that the company, which has already shipped chips to customers, raised $275 million in its previous funding round at a valuation of about $2 billion.
Nvidia is also exploring new memory technologies, one of which is a $500 billion-class AI partnership with major memory manufacturer SK Hynix.
But 3D DRAM stacking technology has obvious shortcomings: DRAM is highly sensitive to temperature, and stacking it on top of compute chips that generate enormous heat is extremely difficult, with fragile stacked layers highly prone to deformation and damage.
If DensityAI can overcome this challenge, the approach will help AI developers lower model operating costs and improve computing speed.
Amazon has already developed its own AI chips such as Trainium, aiming to achieve higher cost-performance in AI training and inference compared with Nvidia GPUs. At the same time, the cloud provider is also working with multiple hardware startups as a supplement to its in-house chips.
In February this year, Amazon signed an agreement with Astera Labs to procure the latter's specialized semiconductor interconnect hardware for Trainium chips. In March, Amazon reached a partnership with chip manufacturer Cerebras, under which the two sides will interconnect server chips to run different stages of AI inference on different hardware, known as disaggregated inference.
It is not yet clear whether this approach has been put into production.
Comments