Meta (META) will reportedly roll out next-gen Astrid chips at the end of 2027; to deploy new in house Arke chips in H1 2027

Meta's prior custom silicon efforts were confined to inference and ranking workloads, so the read-through here hinges on whether Astrid and Arke target training or remain inference parts, since only the former meaningfully pressures the merchant GPU franchise; on previous occasions across the peer set, in-house parts have displaced Nvidia at the margin for internal inference while training clusters stayed merchant. The cost-and-energy framing is the standard rationale and typically matters more for the parent's capex and depreciation trajectory than for the supplier's near-term revenue, given the deployment horizon sits well beyond current order books. The named cadence, Arke first and Astrid later, mirrors the pattern of shipping a mature-node inference part before a more ambitious successor. Worth watching is whether ramp commentary from the supplier and other hyperscalers' own silicon timelines corroborate share shift at the inference edge, and any foundry or packaging capacity commitments that would signal volume intent. As a reported roadmap rather than a confirmed program, the signal is directional.