Loading today’s stories
Product1 distinct publisher2 min readPublished
Reuters reports the acquisition talk has cooled into a partnership, but the $7 billion figure and two senior chip hires say Anthropic treats silicon as a cost it has to own rather than a vendor it can pick.
The Product Desk · Product desk
Compiled by The Product DeskSomething wrong?How this is made
Follow any of these and your For You feed starts watching them — no settings page required.
security
The nationalization argument is really a vendor-continuity memo1 distinct publisher
security
Washington names industrial-scale distillation, then hands the detection bill to abuse teams1 distinct publisher
build
Hugging Face's $13B process puts most teams' model pipeline under a single owner2 distinct publishers
invest
Scalable Capital puts ChatGPT, Claude and Grok inside the European order ticket2 distinct publishers
Whether this touches you depends on whether it's training or inference. MatX has been working on a chip for training large models [8], which sets what it costs Anthropic to build the next Claude rather than what it costs anyone to call the current one. OpenAI's claim at a conference this week that its Jalapeno part beat a comparable Nvidia processor rested on energy efficiency in inference calculations [15], the side of the line a buyer sees on a rate card. Anthropic's sources say it could elect to produce an inference chip as well [18].
The case for owning the design is a cost-of-goods case, and the people briefed on the talks put it plainly: buying a start-up like MatX would give Anthropic in-house design expertise and could help lower costs over the longer term [20]. Set that against what renting costs. The Google chip purchase and the Nscale deal alone come to $81 billion in named commitments [4], and the SpaceX arrangement annualises to roughly $15 billion a year [2], more than twice the price discussed for MatX [3]. That discussed price also sat about 75 percent above the valuation MatX is now asking investors for [1]. Seen from the rent side of the ledger, a design team is cheap to buy. Anthropic declined to comment on the deal talks and MatX did not respond [5], and Reuters could not learn why the talks are no longer active [3].
For the person who has to sign a commitment, two questions sort any lab's chip news: whether it's training or inference, and whether it's money committed or people committed. The chip purchases and cloud deals are money, with dates attached. The MatX outcome so far is people: a partnership discussion in place of a purchase, plus Google chip veteran Amir Salek hired this week and former OpenAI chip engineer Clive Chan hired in June [14]. Only the inference-and-money quadrant should move your assumptions about per-token pricing over the next two years. The people column tells you something slower and still worth knowing, which is where a vendor expects its margin to come from at the end of the decade.
The tradeoff in that rule is that it makes you late. Google builds TPUs and Amazon builds Trainium and Inferentia [19]. Custom silicon from a model lab is not a new idea, and if Anthropic's own silicon ever does for inference what OpenAI says Jalapeno does, customers will most likely notice it first in a quieter rate card, not a keynote.
Ranked by verification strength, evidence, and original report placement.
Anthropic said it is expanding an in-house silicon team to design custom chips that will let Claude models run faster and more efficiently, and plans to keep a multi-chip approach by working with providers including Nvidia and Google.
Designing chips is expensive and time-consuming: it can take a year or more to produce a viable piece of hardware, and design costs for a single generation run in the hundreds of millions of dollars.
Anthropic plans to buy US$36 billion worth of Google's AI chips.
Anthropic signed a US$45 billion deal to rent AI cloud computing power from Nscale.
Anthropic agreed to pay SpaceX US$1.25 billion per month through May 2029 for computing capacity across its data centre clusters.
Anthropic is expected to list in 2026, and Reuters reported earlier in August that the IPO will chase a US$2 trillion valuation hinging on a 2028 revenue figure of as much as US$200 billion.
Distinct publishers with included, body-backed reporting in this cluster.
Evidence-backed comparisons of source perspectives and observed adoption signals. Read the methodology
Which Builder, Operator, and Investor concerns the observed source mix emphasized—not a truth score.
Evidence, demonstrated adoption, hype gap, incentives, and confidence are assessed independently, each on its own current evidence. How these are measured.
One wire scoop, mostly unnamed
Split the story in two and it grades very differently. The contracts, the hires, Nvidia's Aug 26 shortage guidance and Anthropic's own silicon-team statement are checkable and dated. The part that gives the story its headline -- a roughly US$7 billion price, a US$4 billion fundraise, a merger that became a partnership -- comes from three anonymous people at a single desk, with Anthropic declining to comment and MatX not replying. Reuters is unusually candid that it cannot say why the talks stopped, which is honest and also a gap.
Renting at scale, designing from zero
Anthropic's compute consumption is enormous and documented -- US$81 billion in named Google and Nscale commitments, about US$15 billion a year to SpaceX, models already running across Nvidia, Google and Amazon parts. Its own silicon, by contrast, is at the staffing stage: two hires, a survey of start-ups, no acquisition, no part. The pattern around it is real, with Google shipping TPUs, Amazon shipping Trainium and Inferentia, and OpenAI showing Jalapeno, which is why this scores above the floor rather than near it.
A deal that stopped being a deal
The framing runs slightly ahead of the facts. What actually happened, on this account, is that a US$7 billion approach cooled into an unspecified partnership and Anthropic bought nothing -- yet the number anchors the whole story, and MatX's own ask is 75 percent lower. The strategic reading is not wrong: the hires are real, and Reuters' own arithmetic on design cost and timeline explains why buying beats building. But a chip that takes a year-plus per generation is being narrated at the speed of a term sheet, and the piece leaves OpenAI's claim that Jalapeno beat an Nvidia part standing unmeasured.
A number that helps a fundraise
Follow who gains from this leak. MatX is in the market at about US$4 billion, and a story reporting that Anthropic weighed roughly US$7 billion is the single most useful data point a founder could put in front of investors. Anthropic, months from a listing that chases US$2 trillion, benefits from a public narrative that it is pulling its largest cost line in house. Nvidia's shortage guidance gives everyone a respectable reason to talk about second sources. Nobody attaches their name to any of it, and both companies chose not to speak on the record.
Direction firm, specifics soft
That Anthropic is serious about owning chip design is about as solid as a single-source story gets: an on-record statement, two named senior hires, and a documented reason to hedge Nvidia. Everything transactional -- the price, the partnership, whether MatX ends up involved at all -- would move on one on-record confirmation or one competing account, and our coverage contains neither.