Two people familiar with the matter said that AI chip startup DensityAI, founded a year ago by former core members of Tesla's Dojo supercomputer program, is in late-stage funding talks to raise several hundred million dollars in a round that would value the company at $10 billion.

That valuation is unusually high for a company whose chips are still in early development and may be years away from mass production.

The people said management has told potential investors that it has an agreement under which Amazon Web Services (AWS) will purchase its chips if DensityAI's chips meet specified performance targets.

One of the people said Andreessen Horowitz (a16z) is in talks to lead the round.

The negotiations reflect strong investor interest in computing power startups, even when such chip products are far from mature.

Etched, a four-year-old company also developing AI-specific chips, closed a $700 million round led by Jane Street Capital in August at a $21 billion valuation.

The company said it began shipping chips this summer.

Andreessen Horowitz also set up an $1.1 billion fund last month dedicated to investing in chip, memory and other AI hardware component companies.

DensityAI's previous valuation is unknown.

The company had previously raised an undisclosed amount from investors including Dolby Family Ventures, South Park Commons and Firestreak Ventures.

Spokespeople for Amazon and Andreessen Horowitz declined to comment.

Representatives for DensityAI did not respond to requests for comment.

The startup was co-founded a year ago by Ganesh Venkataramanan, former head of Tesla's Dojo program, Bill Chang, former chief systems engineer for Dojo, and Ben Floring, former head of Dojo's AI infrastructure team.

Tesla had hoped the Dojo supercomputer would support the AI software behind autonomous driving, but the Elon Musk-led electric vehicle company temporarily disbanded the team last summer.

A person familiar with the funding said DensityAI has told potential investors that its chips use a unique memory layout that can improve AI computing speed and reduce energy consumption.

The company plans to use three-dimensional dynamic random access memory stacking technology, an emerging approach that stacks memory cells or dies directly on top of a chip's compute cores rather than placing them side by side.

This structure shortens the data transmission distance during inference operations.

Chip startup d-Matrix has already adopted 3D DRAM stacking technology in its second-generation chips and plans to further optimize it in subsequent product iterations.

Media reports said the company, which already sells chips externally, raised $275 million in its previous round at a valuation of about $2 billion and is in talks for a new round.

Nvidia is also exploring new memory technologies, including a $500 billion cooperation with memory manufacturer SK Hynix.

3D DRAM stacking technology has clear shortcomings.

The most prominent challenge is that stacking temperature-sensitive DRAM above high-heat compute units is extremely difficult to engineer, and the fragile stacked layers are highly prone to deformation or damage.

If DensityAI can overcome this problem, the technical approach will help AI developers lower model operating costs and improve computing speed.

Amazon has developed custom AI chips such as Trainium, aiming to offer better cost performance than Nvidia GPUs in AI model training and inference scenarios.

But the cloud giant has also signed deals with several hardware startups to supplement its in-house chip business.

In February, Amazon signed with Astera Labs to purchase the latter's specialized semiconductor interconnect hardware for Trainium chips.

In March, Amazon reached a partnership with chip design company Cerebras, under which the two sides will interconnect their respective server chips and execute AI inference tasks in segments across multiple hardware types, an approach known as distributed inference.

It is not yet clear whether that partnership has been implemented.