Anthropic's $7 Billion MatX Buyout Talks Cool Into a Chip Partnership
The AMW Read
Adds a named ex-Google TPU hire (Amir Salek) and multi-startup chip-shopping detail to the already-known Anthropic-MatX acquisition-to-partnership pivot, meaningfully updating Anthropic's in-house training-chip strategy without a new top-tier entrant.
Named counterparties: Anthropic
Anthropic's $7 Billion MatX Buyout Talks Cool Into a Chip Partnership
Anthropic discussed acquiring AI chip startup MatX for roughly $7 billion before the talks stalled and shifted toward a hardware partnership instead, per a Reuters report. MatX, co-founded in 2023 by ex-Google TPU engineers Reiner Pope and Mike Gunter, designs chips aimed specifically at large-model training rather than inference, and closed a $500 million Series B in February 2026 backed by Jane Street and Situational Awareness Fund. Anthropic has met with several other chip startups in recent weeks without settling on an acquisition target or a technical architecture, and has hired Amir Salek — who led Google's TPU program through its first seven generations before an eight-year run at Nvidia and a stint at Cerberus Capital — into its computing organization under compute lead James Bradbury.
The move puts Anthropic on a training-chip track just as OpenAI's newly announced inference chip, Jalapeño, targets serving cost rather than training cost, splitting the two leading labs' custom-silicon bets along opposite ends of the model lifecycle. As pretraining, post-training and reinforcement learning runs scale to tens of thousands of accelerators, small efficiency gains compound into large cost deltas, pushing frontier labs toward Google TPU- and Amazon Trainium-style co-design of chips, model architecture and training systems. MatX is tracked in the AI Market Watch index with $620 million in total funding (coverage spans roughly 5,000 companies, not a census), underscoring how much capital training-silicon startups are attracting even without a completed acquisition.
For chip-startup investors, an abandoned buyout followed by a partnership conversation is a real validation signal without the dilution or control loss of a sale, and Anthropic's parallel courtship of multiple chip startups suggests more such deals are coming before any one architecture wins. For enterprises building on Claude, an in-house training-chip roadmap is a multi-year bet whose near-term effect is organizational, not pricing — the signal to watch is whether Salek's team ships a working architecture, not whether unit costs move soon.
