A VC-Led Fab Fund: Coatue and MatX Near a Multibillion-Dollar Joint Venture to Finance AI Chips

The most interesting deal in AI hardware right now does not come from Nvidia, and it does not even come from a chip company. It comes from a hedge fund.

On September 4, The Information reported that Coatue Management — the technology-focused investment firm with roughly $50 billion under management — is in early talks to form a multibillion-dollar joint venture with MatX, a two-year-old AI chip startup founded by former Google TPU engineers. The structure is unusual enough to be worth pausing on: rather than simply buying equity, Coatue would help stand up a dedicated vehicle whose job is to finance chip purchases, memory supply, and manufacturing capacity on MatX’s behalf, ahead of the startup’s first commercial shipments planned for 2027.

If the talks conclude, the result would be one of the first times a venture-style investor has applied infrastructure-scale project finance to a pre-revenue semiconductor startup. It is a deal that says as much about the bottlenecks of 2026 as about anyone’s technology.

What the joint venture would actually do

According to the report, the venture’s core purpose is to help MatX “overcome fierce component shortages.” That phrase deserves unpacking, because in 2026 it is the defining constraint of the entire AI economy.

The global memory market — high-bandwidth memory (HBM) above all, but increasingly conventional DRAM as well — has been sold out for quarters, with hyperscalers and AI labs locking in allocations years in advance. Advanced packaging capacity at TSMC is similarly oversubscribed. For a startup shipping zero revenue today but planning to deliver silicon in 2027, the question is not whether its design is good; it is whether there will be memory attached to its chips and a packaging line available when they come back from the fab.

A joint vehicle capitalized by Coatue could do several things equity cannot. It can sign multi-year offtake agreements for memory supply, effectively reserving HBM allocation the way a utility reserves fuel. It can pre-pay for manufacturing capacity at foundries and packaging partners, moving MatX up the queue. And it can carry that spending off MatX’s own balance sheet, so the startup’s reported burn stays focused on engineering rather than supply-chain hoarding.

Multiple outlets — including Crypto Briefing and KuCoin’s news desk, citing the same reporting — described the venture as financing chip purchases “ahead of its 2027” ship date, with The Information’s own summary noting it “aims to finance memory and chip manufacturing capacity.” Talks are described as early, and no dollar figure beyond “multibillion” has been fixed.

Why MatX keeps attracting this kind of attention

MatX is not a normal early-stage chip bet. Founded in 2022 by Reiner Pope and Mike Gunter — both veterans of Google’s TPU team — the company raised a $25 million seed in 2024, a Series A at a reported $300 million valuation later that year, and then a $500 million Series B in February 2026 led by Jane Street and Situational Awareness, with Spark Capital and others joining. Total disclosed funding now stands north of $600 million, at a valuation reported in the “several billion” range.

What distinguishes the company is its scope of ambition: MatX is building chips designed to handle both training and inference for large language models — a “dual-purpose” architecture, as The Information characterized it — optimized specifically for transformer workloads rather than adapted from general-purpose GPU designs. That focus on LLM-native silicon is precisely why the company keeps surfacing in strategic conversations.

It surfaced most prominently two weeks ago, when Reuters reported that Anthropic had discussed acquiring MatX for roughly $7 billion before abandoning the plan and pivoting toward a partnership on chip design. MatX has also reportedly been seeking fresh funding at around a $4 billion valuation while engaging with top AI labs to test real-world use cases. The Coatue talks are the third distinct strategic approach to the company in a month — a trajectory that mirrors how Nvidia’s early customers once lined up, except compressed into a single funding cycle.

The financing innovation is the story

Strip away the names and the structural novelty is clear. The AI buildout has hit the point where the binding constraint is not capital in general, but specific, committed capital aimed at specific physical inputs: memory, packaging, power, and fab slots.

Hyperscalers solve this with balance sheets — Google pre-books TPU capacity internally, Microsoft and Meta negotiate directly with SK hynix and TSMC. Startups cannot. A seed-stage-style equity round, even a $500 million one, buys engineering headcount and tape-outs but does not comfortably buy years of reserved HBM allocation. That is the gap a project-finance-style joint venture fills.

Coatue is, in effect, behaving like an infrastructure investor — closer in spirit to the funds behind data-center joint ventures and power projects than to a typical Series B lead. Reports of similar structures elsewhere (Anthropic itself was recently linked to a $1.5 billion joint venture with Blackstone for compute infrastructure) suggest this is becoming a template: when the scarce asset is physical capacity, the financing instrument migrates from venture equity to project finance.

Why it matters

For the chip supply chain, a Coatue–MatX deal would be a signal that allocation warfare has trickled down from trillion-dollar hyperscalers to Series-B startups. If a two-year-old company can lock memory and fab capacity through a multibillion-dollar side vehicle, expect every credible accelerator startup to seek the same — further tightening an already brutal supply environment for everyone outside these deals.

For investors, it reframes what “AI infrastructure” means as an asset class. The money is no longer just chasing GPUs and data centers; it is chasing the inputs to the inputs, with financing structures borrowed from oil-and-gas-style offtake agreements.

And for Nvidia, the quiet subtext of every story like this one: the more that labs, clouds, and now financiers organize to secure non-Nvidia silicon — even silicon that will not ship until 2027 — the more the frontier of competition shifts from chip design to supply-chain command. MatX’s chips may or may not be great. The market is already betting that in 2026, whoever controls allocation controls the game.

The talks may still fall apart — early-stage joint-venture discussions often do. But the fact that they are happening at all marks the newest entry in 2026’s most important trend: the financialization of AI’s physical supply chain.