Anthropic's Abandoned $7 Billion Chip Buyout Is a Margin Story, Not an M&A Story

Anthropic was prepared to pay roughly $7 billion for MatX, an AI chip startup, and then it wasn't. Reuters reported that the two companies discussed a buyout meant to accelerate Anthropic's custom-hardware effort; the talks are no longer active and have evolved into a possible partnership, while MatX now looks to raise new capital at a valuation of about $4 billion — roughly half the price Anthropic was willing to pay to own it.

That gap is the first and most instructive number in the story. A $7 billion price for a startup selling at $4 billion is a 75% premium on disclosed figures — and it is not a premium for revenue, because MatX has essentially none to sell yet. It is a premium for control: exclusive access to a chip roadmap, ownership of the engineering team, and leverage over the cloud providers whose hardware runs Claude. Ownership would have let Anthropic point MatX's silicon at its own models and keep rivals out. Anthropic got none of that.

Why would a software lab be willing to pay so much for a hardware company that has yet to ship? Because its single largest expense is renting other people's silicon. Analysts estimate compute consumes roughly 56 to 71 cents of every dollar of Anthropic revenue — the exact figure swings with how training spending is amortized — while the company's revenue run rate has blown past $65 billion annualized. Nvidia, which supplies much of that fleet, runs a gross margin near 74% and has said supply stays short through 2027. Every dollar of chip rent is a dollar that never reaches Anthropic's own margin, which most estimates put in the 44–60% range today.

That is where custom silicon enters. Building a chip for how Claude actually serves tokens can cut the total cost of operating inference by as much as 65% versus renting Nvidia GPUs, according to reporting on Anthropic's in-house co-design effort. A chip that makes each token cheaper to produce is not an R&D sideshow; it is the difference between a gross margin in the 50s and one that approaches the 70s that a $2 trillion valuation assumes.

Set the $7 billion next to what Anthropic has already signed up to spend on other people's computing, and the buyout stops looking like money and starts looking like leverage. The announced figures include $45 billion for cloud capacity from Nscale, a planned $36 billion purchase of Google's AI chips, $1.25 billion a month paid to SpaceX infrastructure through May 2029, and more than $100 billion on Amazon Web Services over a decade. These run across different horizons and some are options, but a simple sum of the disclosed amounts exceeds $200 billion. On that scale a $7 billion bid was never a meaningful outflow — it was a bet that owning the chip design would bend the one cost line that decides whether the valuation holds.

MatX's own investors help explain why the buyout stalled. In February the startup raised $500 million from Jane Street and Situational Awareness, the fund of former OpenAI researcher Leopold Aschenbrenner, and counts Marvell — itself a custom-chip designer — as an investor. Founders who take money on those terms typically want to sell chips to every lab with a large model, not become a captive subsidiary of one. A partnership would still hand Anthropic optimized silicon and a seat in the design conversation, but without exclusivity: the same wafer could end up powering a rival's product line. Reuters reported it could not determine why the talks closed; the shape of the outcome — MatX remaining independent, Anthropic declining to pay the premium — is consistent with founders who wanted the open market.

This is also why the buyout is connected to Anthropic's reported profitability, and the connection is uncomfortable. The company's march to its first quarterly operating profit has a narrative layer and a numbers layer. A skeptical analysis of the second quarter's projected $559 million operating profit points to timing: SpaceX fees for May and June were temporarily discounted, and customers' prepaid token contracts were booked as revenue ahead of the compute cost being recognized. When the full bill returns, the argument goes, costs snap back toward trend. Custom silicon is the structural answer to an accounting problem. Reduce cost per token permanently and you no longer need a one-time discount to print an operating profit — which is exactly why Anthropic has been hiring chip engineers — Reuters reported a veteran of Google's chip team and an engineer who worked on OpenAI's in-house "Jalapeno" chip.

The in-house inference chip and the MatX discussions are the same strategy from two angles. Anthropic is co-designing a custom ASIC for inference — running Claude's answers — with Samsung reportedly in talks to manufacture it, per The Information, while MatX, founded by former Google TPU engineers, offers the training route. Move the design in-house and the supply chain hardly flattens: Broadcom and Marvell still design roughly 95% of custom AI chips, and the silicon still gets fabricated at TSMC or Samsung and wrapped in high-bandwidth memory. Anthropic would be moving from renting Nvidia's product to renting the foundry that makes its own — the same logic OpenAI and Broadcom used in June, when they unveiled their first custom inference chip.

The public-company trail follows that money. Anthropic itself is not tradeable yet — it filed confidentially for an IPO on June 1, and the valuation case rests on reaching $190 billion to $200 billion of revenue in 2028 to justify a listing in the neighborhood of $2 trillion. For investors who cannot buy the company directly, the chain that carries its chip spending is:

• Amazon — the closest public proxy. It deepened its stake in April with the potential for $25 billion of new investment, will host Claude at scale on its Trainium chips, and is tied to more than $100 billion of Anthropic's AWS spending over a decade. You are trading Amazon's own thin-margin, enormous-capex profile (margins near 50%, capital spending running into the hundreds of billions) for the growth of its largest AI customer.
• Broadcom and Marvell — the designers who would actually build a custom Anthropic chip. Broadcom's gross margin is near 68% and it has reported AI revenue compounding fast; Marvell, which already backs MatX, is the more speculative lever.
• TSMC and Samsung — the foundries every custom chip ultimately passes through.
• Nvidia — the incumbent whose inference margins are the target of the whole exercise, even as it retains dominant control of training, 74% gross margins, and supply that stays tight through 2027.

Now the other half of the evidence, because it is the part that keeps the story honest. None of this hardware exists yet. MatX's chip, the MatX One, targets roughly 10x the throughput of Nvidia's silicon for large models — and it is a design goal, not a product, projected to start shipping in 2027 through TSMC, assuming foundry capacity bends its way. A custom chip generation costs hundreds of millions of dollars and takes more than a year to develop. And the Anthropic partnership is a discussion, not a contract; at this stage there is no exclusivity to pay for, and no guarantee MatX's silicon will be friendly to just one lab.

What is real is the direction. A company that has signed up for hundreds of billions of dollars of third-party compute has concluded that the only durable way to fix its cost structure is to control the design of the chips that compute runs on. The number to watch is not a deal size. It is Anthropic's cost per revenue dollar and gross margin as its silicon effort ramps, and which node of the chain — designers, foundry, or the squeezed incumbent — collects the flow if the plan works.