Anthropic's $7 Billion Chip Bet, and the Rival Quietly Paying It $10 Billion a Year

Startups · Infrastructure & Funding

Anthropic went far enough into buying a chip startup to attach a $7 billion price tag to it, then walked away. Around the same time, the company publicly criticizing Anthropic's approach to AI turned out to be one of its biggest customers, projecting up to $10 billion a year on Claude. Two separate stories, one obvious theme: nobody in this industry can afford to only be somebody else's rival anymore.

Wireframe 3Sixty Team · Published August 29, 2026 · 13 min read · Startups

Reuters broke this one the way it breaks most things in this beat now — a Thursday afternoon story, two unnamed sources, and a dollar figure specific enough that it stopped anyone from waving it off as speculation. Anthropic, the story said, had discussed buying an AI chip startup called MatX for roughly $7 billion. By the time the report published, the talks weren't even active anymore. They'd already cooled into something looser: a conversation about a partnership instead of an outright purchase. That alone would have been a solid week's worth of chip-industry gossip. What made it a genuinely interesting story is what surfaced two days later, in a completely separate report, about who's actually paying Anthropic's bills right now.

What actually happened with MatX

MatX is a small company with an unusually specific résumé behind it. It was founded in 2023 by Reiner Pope and Mike Gunter, two engineers who previously worked on Google's Tensor Processing Units — Pope on the software side, Gunter as a lead hardware designer. That pedigree matters more than it might for a typical chip startup, because it means MatX isn't starting from zero on a genuinely hard problem: building processors purpose-built for training large language models, rather than general-purpose GPUs repurposed for the job.

According to Reuters, which cited two people briefed on the discussions, Anthropic and MatX got far enough into acquisition talks to land on a number: roughly $7 billion. A third source told Reuters the conversation has since shifted toward a partnership rather than a full buyout, and nobody involved would explain why the more ambitious version of the deal fell apart. Anthropic declined to comment. MatX didn't respond to a request for comment either. What is confirmed independently is that MatX is now out raising fresh capital of its own, at a valuation of roughly $4 billion — a number worth sitting with for a second, because it means Anthropic was reportedly prepared to pay nearly double MatX's own fundraising valuation to bring the team in-house entirely, rather than just invest in it.

The MatX story, at a glance

Reported price ~$7B for a full acquisition — talks have since cooled into partnership discussions instead

MatX's own valuation Currently raising new capital at roughly $4B, per Reuters' sourcing

Founders Reiner Pope and Mike Gunter, both former Google TPU engineers, founded the company in 2023

Why it fell apart Not disclosed — Reuters could not determine why the active talks ended

What Anthropic is doing instead Building an in-house silicon team, actively hiring engineers with chip-tapeout experience

Why a $7 billion acquisition looks almost modest right now

Here's the context that makes this number easier to make sense of: Anthropic's total infrastructure spending this year makes a $7 billion chip acquisition look like a rounding error rather than a headline figure. The company has committed $36 billion to Google for access to TPU capacity, agreed to pay roughly $45 billion to rent cloud computing power from a provider called Nscale, and locked in $1.25 billion a month to SpaceX through May 2029. Separately, Broadcom has reportedly been seeking more than $60 billion in debt financing specifically to fund chip production destined for Anthropic, and AMD has put $5 billion directly into the company alongside a two-gigawatt hardware deployment commitment.

Designing a chip from scratch is slow and genuinely expensive — a single generation can take well over a year and run into the hundreds of millions of dollars before a single unit ships. Buying a team that has already taped out working silicon designs, the way MatX reportedly has, is considerably faster than building that same capability from a standing start. That's the logic behind the price tag: Anthropic wasn't just buying a company, it was buying time, in an industry where the companies that get their own custom silicon working first get a real, compounding cost advantage over everyone still renting someone else's chips.

Against this backdrop, $7 billion for a chip team isn't the biggest number on Anthropic's ledger this year — it's arguably one of the smaller ones.

The other story: Meta's $10 billion problem

This is where the week got genuinely interesting, because it involves a different company entirely. Just days after the MatX story broke, The New York Times reported that Meta had internally projected it could spend up to $10 billion a year on Anthropic's AI tools and models — Claude Code adopted at scale by Meta's own developers, and Anthropic's models reportedly running underneath internal testing and parts of Meta's new AI agent product, Hatch. To put that number in some perspective, Anthropic itself had projected in July that its total annualized revenue would exceed $65 billion. If Meta's internal projection holds, one company's spending alone could account for a meaningful slice of Anthropic's entire business.

The awkward part is the timing relative to Meta's own public messaging. Mark Zuckerberg published a lengthy public essay on August 10 arguing against AI power concentrating inside a small handful of closed labs — commentary widely read across the industry as aimed squarely at Anthropic and its CEO, Dario Amodei. According to the Times, Meta was quietly one of Anthropic's largest customers at the exact same time that essay was circulating. Meta's head of AI products, Nat Friedman, reportedly told employees internally that leaning harder on Meta's own coding tools instead of Anthropic's could reduce Anthropic's revenue ahead of its anticipated public offering — a comment that only makes sense in a world where Meta genuinely sees Anthropic as both a rival to undercut and a vendor worth paying.

Zuckerberg spent thousands of words warning about AI power concentrating in a few closed labs. His own company was quietly on track to hand one of them ten billion dollars a year. Editorial analysis — Wireframe 3Sixty

There's a second Meta–Anthropic thread, and it's easy to mix up with this one

Worth being precise here, because two genuinely different Meta–Anthropic stories have been circulating this year and they get conflated easily. Back in July, Reuters and CNBC reported that Anthropic had proposed leasing up to $10 billion of computing capacity from Meta over a two-year period — chips, power, and data-center infrastructure, paid in monthly instalments, to help Anthropic keep up with Claude's surging usage. That deal runs in the opposite direction: Anthropic paying Meta for compute, not Meta paying Anthropic for software. Those talks began in June, remained described as early-stage as of the July reporting, and either side reportedly retains the option to exit before the two-year term ends.

The story that broke in late August is the mirror image of that one: Meta as Anthropic's customer for AI tools and models, not its compute supplier. Both figures happen to land at roughly the same $10 billion ceiling, which is almost certainly why they've been getting tangled together in casual conversation about this story. But they describe two different possible flows of money, in two different directions, tied to two different parts of each company's business. It's entirely possible both arrangements could exist simultaneously — Meta selling Anthropic spare data-center capacity while Anthropic sells Meta's own developers access to Claude — which would make the two companies each other's customer and supplier at the same time, on top of already being direct competitors in foundation models.

Which deal	Money flows	Status as of late August
Compute lease (reported July)	Anthropic pays Meta, for GPU and data-center capacity	Early-stage talks, initiated by Anthropic in June; either side can exit before the 2-year term ends
Tool spending (reported August)	Meta pays Anthropic, for Claude Code and model access	Internal Meta projection, not a signed contract — but already reflects real, current usage
MatX chip talks	Anthropic would have paid MatX roughly $7B for the team and its designs	Acquisition talks cooled; discussion has shifted to a partnership instead

Why none of this is actually contradictory

It's tempting to read all of this as hypocrisy — a company criticizing a rival in public while quietly funding it in private, or a chip acquisition falling apart right as the buyer's need for chips keeps growing. The more useful read is less cynical and, honestly, more revealing about where the entire AI industry actually stands right now: the amount of compute every major lab needs has grown faster than any single company's ability to build or buy it alone, and that scarcity is forcing direct competitors into commercial relationships that would have looked absurd two years ago. SpaceX already sells computing capacity to both Anthropic and Google. Google sells access to its own Gemini models to Meta, even while rationing its own internal capacity. Meta and Anthropic compete head-to-head for the same enterprise AI customers through Llama and Claude, and may simultaneously end up as each other's supplier and customer on two entirely separate deals.

The MatX story fits the same underlying pattern from a different angle. Anthropic doesn't actually need to own a chip company to reduce its dependence on Nvidia — it needs the expertise and the head start that owning one would provide, and if a partnership gets it most of that value without the full $7 billion price tag and the complexity of an acquisition, that's arguably the more disciplined outcome, not a failure. Anthropic is heading toward an IPO reportedly targeting a valuation near $2 trillion, built on revenue projections that assume it can keep scaling Claude affordably. Every one of these deals — the abandoned MatX buyout, the in-house chip team, the Broadcom debt financing, the Meta compute lease, and Meta's own Claude spending — is a different piece of the same underlying problem: getting enough affordable compute to justify that valuation before the public markets get to judge it directly.

What to watch from here

Whether the MatX partnership actually materializes

A "discussion about a partnership" is not a signed agreement. Watch for whether Anthropic announces any formal hardware collaboration with MatX, licenses its designs, or simply hires away key engineers instead — each outcome tells a different story about how much of MatX's value Anthropic actually needed to own outright.

Whether Meta's internal $10B figure becomes public spending

Internal projections shift constantly and aren't commitments. The real signal will be whether Anthropic's own revenue disclosures, expected as part of its IPO process, show Meta as a named or clearly identifiable large customer.

Whether the Meta compute-lease deal actually closes

Reuters' July reporting described those talks as complex and early-stage, partly because Meta has never run a business selling computing power before. A signed deal here would be a much bigger structural shift than the MatX story — it would make Meta a genuine infrastructure vendor to a direct AI competitor.

Frequently asked questions

Did Anthropic actually buy MatX?

No. Reuters reported that Anthropic discussed acquiring MatX for roughly $7 billion, but those talks are no longer active and have reportedly shifted toward a partnership discussion instead. No acquisition has closed.

Is Meta really spending $10 billion a year on Anthropic?

That figure is an internal Meta projection reported by The New York Times, not a confirmed, finalized annual spend. It reflects Meta's growing use of Claude Code internally and Anthropic models powering parts of Meta's Hatch product.

Is this the same as the Meta compute deal reported in July?

No — those are two separate arrangements. The July story described Anthropic paying Meta for computing capacity. The August story describes Meta paying Anthropic for AI software and models. Both happen to involve figures near $10 billion.

Why does Anthropic want its own chips if it already has deals with Google, AMD, and others?

Diversifying suppliers reduces dependence on any single one, but owning custom silicon designed specifically for its own models could let Anthropic run Claude faster and cheaper over time than renting general-purpose hardware, even from multiple vendors.

The bottom line

Neither of these stories is really about hypocrisy, even though that's the easy headline for both. They're about an industry where the compute bill has grown large enough that competitive rivalry and commercial dependency now routinely sit on top of each other inside the same relationship. Anthropic was willing to pay nearly double a startup's own valuation just to shortcut its way to custom silicon, then walked that back the moment a lighter-weight partnership looked like it could get most of the same result. Meta spent the summer publicly warning about AI power concentrating in a handful of labs while privately becoming one of those labs' largest customers. Both companies are behaving rationally by the logic of their own balance sheets. It just doesn't look that way from the outside, where the public statements and the private invoices are telling two different stories at the exact same time.

The number worth remembering out of all of this isn't $7 billion or $10 billion specifically — it's how unremarkable both figures have become in a single week of AI infrastructure reporting. That, more than any single deal, is the actual story of where this industry's money is going in 2026.