Microsoft (MSFT) announces its third-gen Maia 300 AI chip, with production starting in 2027, The Verge reports

Hyperscaler in-house silicon programs follow a familiar arc: announcement, tape-out, deployment in internal data centres first, with merchant GPU volumes displaced only gradually and mostly for inference rather than training workloads. Past episodes of this kind have tended to matter less for the hyperscaler's own multiple than for the read-through to the merchant accelerator suppliers, where the sensitivity runs through capex allocation and the share of incremental capacity awarded to custom parts versus off-the-shelf. The 2027 production start is the operative detail: chips at that horizon are typically committed against multi-year capacity agreements, and the telling comparison in prior cycles has been announced custom silicon versus actual deployed volume, which has historically skewed heavily toward the merchant incumbents. Coexistence rather than substitution has been the pattern, with in-house parts absorbing a minority share of internal workloads while external procurement continues to grow. What is worth watching is any quantification of wafer or packaging capacity booked, commentary from the merchant suppliers' own capex customers, and whether the chip targets inference economics specifically, since that is where custom parts have made their deepest inroads.