Coatue, MatX in Talks on New Multibillion Chip-Financing Venture

Investment firms and tech companies are devising increasingly novel structures to get around the financial and hardware supply constraints of the AI boom.

In the latest example, Coatue Management is in talks to form a joint venture involving MatX, a young AI chip startup, to help it overcome fierce component shortages, according to people with knowledge of the deal. MatX has developed a chip that can handle both training of AI models and inference—running the models—and has attracted interest from big tech firms, including Anthropic, which briefly considered buying it.

The joint venture would aim to finance purchases of memory and logic dies as well as capacity at chip manufacturers such as Taiwan Semiconductor Manufacturing Co., the same way large, established customers like Broadcom and Marvell Technology do, one of the people said. Another person familiar with the joint venture said it is in discussions to be funded with billions of dollars.

The specifics of the arrangement will ultimately depend on negotiations with memory and chip manufacturers. It has not yet been decided which manufacturers will participate in the venture.

Coatue is discussing the joint venture because it’s looking for creative new ways to finance AI infrastructure, a person familiar with the deal said. The deal talks are early and could change.

The investment firm, which managed more than $70 billion in assets as of the end of last year, previously led or co-led financings of data center and AI startups including Anthropic, CoreWeave and Cursor. Coatue also set up a data center venture in 2025 called Next Frontier to buy land for developing AI data centers. Coatue also holds stakes in TSMC and Cerebras.

Separate from the chip capacity venture, Coatue has discussed participating in a new funding round for the Mountain View, Calif.–based chip firm that would at least double its approximately $4 billion valuation set in a financing this past February, according to multiple people with knowledge of the talks. The funding plans are in their early stages, and terms have not been set.

As the cost of renting servers to train AI rises beyond the reach of many startups, private equity and investment firms as well as rich incumbents like Nvidia, Broadcom and Google are doing everything they can to give startups enough capital or debt to stay in the game.

Nvidia, for instance, has been on an investment spree for the past three years, providing capital to dozens of companies that buy or rent its chips, and even to firms that develop land and power for data centers. Meanwhile, investment firm Andreessen Horowitz for years has maintained a cluster of AI servers to help its young portfolio companies access compute.

The Coatue joint venture involving MatX would be innovative in opening up more opportunities for hardware startups trying to cope with the intense competition for chip components and chip manufacturing capacity.

Even large chip designers are struggling to get memory and other components or capacity from the likes of TSMC, SK Hynix, Samsung and Micron. Companies that can show they can afford the capacity and have customers lined up can get allocation from those suppliers more easily.

In MatX’s case, supply constraints are some ways out: The four-year-old company expects to reach a milestone known as tape-out, when the physical designs of the chips are complete and ready for trial production, in the first half of next year, according to three people familiar with its plans. But Coatue is looking to get ahead of looming problems.

Though the specific structure of the joint venture hasn’t been set, MatX has discussed having something like a right of first refusal to use the manufacturing capacity, according to one of the people familiar with the arrangement. It’s not clear what exactly would happen if MatX decides not to use the capacity reserved under the agreement. The joint venture could potentially work with other startups, or return the capacity to the supplier to then resell to another customer.

Dual Purpose

MatX was founded by Reiner Pope and Mike Gunter, who previously worked at Google’s AI research group and chip design teams, respectively. The startup says it’s developing AI chips that can process a lot of data quickly.

MatX’s prototype chip can handle both AI training and inference, but the interest it received from major AI developers has focused primarily on inference, according to a person familiar with the discussions. Inference chips run AI models.

The chip can be both fast and powerful because it combines an amount of high-bandwidth memory similar to that in Nvidia’s Rubin graphics processing units with another, cheaper type called static random-access memory. The chip can also multiply large amounts of numbers—the core calculations that make chips run—in a more efficient way than existing chips, the company’s founders have said.

This technology has attracted interest from larger tech companies. In recent months, Anthropic discussed a potential acquisition of MatX, according to two people familiar with the talks, but it didn’t proceed. Reuters earlier reported that Anthropic considered paying $7 billion for MatX.

MatX last raised $500 million in February in a round led by quant trading giant Jane Street Capital that included investments from Spark Capital and Leopold Aschenbrenner’s now-troubled investment fund, Situational Awareness.

Venture capitalists are no longer shying away from capital-intensive, high-risk hardware investments. In recent months, investors have shown a willingness to value chip startups richly, even before they have significant revenue. Nvidia kicked off the interest by agreeing in December to license technology from chip startup Groq for $20 billion.

Rival Etched, which is also designing a specialized chip for AI, just raised $700 million last month at a $21 billion valuation in a round led by Jane Street. Etched said it has $1 billion in chip orders from customers including Jane Street.