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

Reuters reports Anthropic previously discussed acquiring AI chip startup MatX, with a ballpark figure of roughly $7B cited before the talks stalled. The discussions have reportedly shifted from an outright acquisition toward a partnership rather than a deal structure. Net impact is unclear until deal terms are finalized.

Analysis

This reads less like a classic M&A event and more like a signal that frontier AI buyers are putting a strategic value on silicon optionality: owning or partnering on chip IP is a bargaining tool against concentrated supply, not just a growth asset. The immediate equity read-through is limited, but the negotiating leverage shifts toward any vendor that can design, fabricate, or package differentiated AI silicon faster than a startup can scale alone.

The first-order winners are the enablers of custom accelerators and the supply chain around them: AVGO and MRVL on design wins, TSM on advanced-node capacity, and AMAT/ASML on the tooling that makes each new tape-out possible. The less obvious loser is the assumption that all AI spend flows linearly into merchant GPUs; if large model companies keep pushing custom silicon, the long-run mix shifts from pure compute units toward higher-margin inference efficiency, which can cap terminal multiples for NVDA even if near-term demand stays strong.

Contrarian view: the market may misread this as "more AI capex," when the deeper signal is "cost discipline and vendor concentration risk." The real catalyst is not the rumor itself but the next 1-2 earnings cycles, where commentary on inference economics, custom ASIC adoption, and supply commitments will confirm whether this is a structural share shift or just talent acquisition. If NVDA/AVGO/TSM all continue to post unchanged data-center demand and no change in customer mix, the thesis is likely over-interpreted.

Key Decisions for Investors

• Lean long AVGO on any 2-4% pullback over the next 1-3 months; best expression of custom-AI monetization with lower binary risk than a pure startup proxy. Stop if hyperscaler/custom-chip commentary fails to improve by the next earnings cycle.
• Add TSM as a 6-18 month structural beneficiary on weakness; custom silicon still needs leading-edge wafers and packaging, and that demand is harder to displace than individual chip architectures.
• Relative-value idea: long AVGO / short NVDA into the next 1-2 quarterly reports only if channel checks show custom ASIC traction; thesis is mix shift, not absolute AI demand collapse. Falsify if NVDA re-accelerates data-center growth materially above consensus.
• Use SMH as a tactical proxy only if you want to express 'AI infrastructure still wins but composition changes'; prefer call spreads rather than outright long given the rumor-driven nature of the signal.
• Watchlist, not a trade yet: AMAT and ASML. If they begin to cite advanced packaging or foundry bottlenecks as upside drivers, that would validate the second-order supply-chain thesis and extend the trade window.