Anthropic planned a $7bn takeover of chip startup MatX, then walked away

Anthropic got far enough into a plan to buy the AI chip startup MatX to have a number attached to it, roughly $7bn, before the deal stopped moving. Reuters reported the talks on Thursday, citing people familiar with them, and said the discussions have since evolved into a conversation about a partnership rather than an acquisition.

Why the purchase was abandoned is the part nobody has explained. Reuters said it could not determine what ended the active negotiations, and the reporting rests on unnamed sources rather than any filing or announcement.

Anthropic declined to comment to Reuters on the talks, and MatX did not respond to Reuters’ request for comment. TNW has not independently verified the report.

MatX is a small company with an unusually specific pedigree. It was founded in 2023 by Reiner Pope, who led AI software work for Google’s tensor processing units, and Mike Gunter, a lead designer on the TPU hardware, and it builds processors aimed squarely at large language model training rather than at general-purpose computing.

The startup raised a $500m Series B in February, led by Jane Street and Leopold Aschenbrenner’s Situational Awareness fund, with Marvell, Spark Capital and the Collison brothers also on the cap table. Its pitch is that a chip designed for one workload can beat Nvidia’s GPUs by a wide margin on that workload, with volume shipments planned from 2027.

The valuation arithmetic is worth pausing on. Reuters reported that MatX is now raising at about $4bn, which puts the discussed purchase price roughly three-quarters above the price at which private investors are currently being asked to buy in.

That gap is the sort of premium a buyer pays for control and for time. Anthropic has been assembling silicon capability at speed, and buying a team that has already taped out designs is considerably faster than hiring one.

The company confirmed an in-house chip team earlier this year, advertising for engineers who had shipped silicon, and it is separately in talks with Samsung about manufacturing a custom part.

The financing around all of this has grown to a scale that makes a $7bn acquisition look almost modest. Broadcom has been seeking more than $60bn in debt to fund chips destined for Anthropic, and AMD has put $5bn into the company alongside a two-gigawatt deployment commitment.

The wider market has moved in the same direction, and fast. Etched, which builds along similar lines, raised $500m at a $5bn valuation, and a cluster of accelerator startups took in about $1.6bn between them across five rounds this year.

Buyers at Anthropic’s scale change what those companies are worth, which is part of why a reported $7bn figure drew attention without a signed deal behind it. It sets an informal ceiling that the next negotiation in the sector will be measured against.

Against that backdrop, a partnership rather than a purchase is not obviously a retreat. An acquisition would have given Anthropic a design team and an integration problem at the same moment it is trying to get chips into data centres, and a supply agreement delivers the silicon without the org chart.

It also leaves MatX free to sell to everyone else, which matters for a company whose Series B investors presumably did not back it to become a single customer’s internal department. A $4bn round is easier to raise as an independent supplier than as an asset with one plausible buyer who has already passed.

What the episode confirms is the direction of travel. The frontier labs are no longer content to be Nvidia’s customers, and the fastest route to an alternative runs through the small number of teams who have designed accelerators before.

Neither company has said anything publicly since the report, and there is no indication that a partnership has been signed. For now, the only firm fact is the one Reuters put on the record: a $7bn deal was planned, and then it was not.