Anthropic Planned, Then Walked Away From a $7 Billion Acquisition of Chip Startup MatX

Anthropic, the company behind Claude, seriously considered buying its way into the silicon business. According to a Reuters exclusive published August 27, 2026, the AI lab held discussions to acquire MatX — a startup founded by former Google TPU engineers — for roughly $7 billion, before ultimately abandoning the talks. The revelation offers a rare look at how aggressively frontier AI companies are weighing vertical integration into chips as compute costs become the dominant line item in their P&L.

What Happened

Citing sources familiar with the matter, Reuters reported that Anthropic explored acquiring MatX as a way to accelerate its efforts to develop custom hardware for training and inference. The reported price tag was approximately $7 billion — a striking figure for a startup that had raised a little over $600 million to date. Talks eventually stalled, and no deal was reached. Rather than a full acquisition, the relationship has reportedly shifted toward potential partnership discussions.

MatX, for its part, is now seeking fresh funding at a valuation of around $4 billion, according to the same Reuters report. That detail is notable: Anthropic’s would-be offer would have represented a substantial premium over where the startup’s next round is likely to clear.

Who Is MatX?

MatX was founded in November 2022 by two Google alumni with deep TPU pedigrees. CEO Reiner Pope previously led AI software development for Google’s Tensor Processing Units, while CTO Mike Gunter was a lead designer of the early TPU silicon itself. Their thesis is unusually focused: build chips that only know transformers. Rather than general-purpose accelerators that handle many workloads adequately, MatX designs its silicon exclusively around the transformer architecture that underpins virtually every frontier large language model.

The company raised a $500 million Series B in February 2026, led by Jane Street with participation from Situational and others, on top of earlier rounds backed by Lightspeed Venture Partners, Bessemer Venture Partners, Cisco Investments, and D1 Capital. Total funding sits north of $600 million — pocket change compared to the nine-figure gigawatt commitments hyperscalers are signing, but enough to bring a first-generation product toward tape-out.

Why Anthropic Would Want Its Own Silicon

The strategic logic behind the flirtation is straightforward: compute is the binding constraint on Anthropic’s business. The company’s annualized revenue run rate surpassed $65 billion in August 2026, according to Bloomberg and Axios reporting, up from $9 billion at the end of 2025 — and nearly all of that revenue is spent, directly or indirectly, on compute.

Anthropic has already made enormous bets on someone else’s silicon. In April 2026 it signed a landmark agreement with Google and Broadcom securing 3.5 gigawatts of next-generation TPU capacity beginning in 2027, on top of an earlier commitment to one million TPUv7 Ironwood chips through Google Cloud. Those deals give Anthropic scale — but not control. Renting capacity on Google’s TPUs means Anthropic’s marginal cost curve, its roadmap timing, and even aspects of its competitive positioning are tied to a rival’s infrastructure decisions. Google is, after all, building Gemini in direct competition with Claude.

Owning a chip design team outright would change that calculus. It would give Anthropic the kind of vertical integration Google itself enjoys with the TPU program, and that Amazon has with Trainium and Microsoft has pursued with Maia — plus, of course, the model NVIDIA has turned into the most valuable semiconductor franchise on Earth. Reuters notes Anthropic has already bolstered its internal chip team with strategic hires, suggesting the company’s silicon ambitions did not die with the MatX talks.

Why the Deal Died

Neither party has commented publicly, so the reasons for the breakdown can only be inferred. But the backdrop offers clues.

First, there is the price asymmetry: a $7 billion offer against a startup now raising at ~$4 billion suggests the sides may have disagreed on timing and risk — MatX’s chips are still pre-scale, and paying a near-doubling premium for unproven silicon is the kind of bet that looks brilliant in hindsight or reckless in a downturn.

Second, Anthropic is widely reported to be preparing for an IPO, with market chatter about valuations approaching $900 billion in upcoming funding rounds. Pre-IPO companies tend to avoid absorbing $7 billion of integration risk in a category — semiconductor design — where execution timelines run in years, not quarters. A partnership structure lets Anthropic buy optionality without the balance-sheet drama.

Third, competition for chip talent and IP is intense, but so is the alternative path: Anthropic already works with Broadcom, one of the most sophisticated custom-silicon design houses in the world. Building on that relationship may deliver 80% of the benefit at a fraction of the risk.

The Bigger Picture

Whatever the specific reasons, the story confirms a structural trend in the AI industry: the frontier is consolidating around whoever controls compute. OpenAI has its multi-gigawatt deals with Oracle, Microsoft, and AMD, plus its own networking ambitions. Google, Meta, and Amazon design their own accelerators. xAI builds its own datacenters. And Anthropic — the lab that once positioned itself as the compute-humble safety alternative — is now negotiating billion-dollar silicon acquisitions and locking in gigawatts of capacity years in advance.

For MatX, the leaked talks are arguably a gift: being courted at $7 billion by Anthropic validates the transformer-only thesis and will anchor its upcoming round. For the rest of the industry, it is one more data point that in 2026, the moat isn’t just the model — it’s the metal.

The deal didn’t happen. The signal did.