Google & Amazon Capex Surges: AI Supply Chain

With Google and Amazon reported astonishing 2026 capex of $180/$200 Bn, $50/$60 Bn above wall street consensus pre results respectively, the investment sentiment comes back. Previously investors worry about OpenAI share loss could weaken the overall infra investment, yet Anthropic Cowork launch reignites competition, or even accentuate the competition.

For Google supply chain, the most important hardware partner is Celestica on both server and switch, and holds 100% share for Google in-house TPU server and OCS switch. Celestica has 32% revenue from Google as of C4Q25. Previously there were rumors about Inventec taking share of TPU server, while Inventec is the next assembly step L11, not Celestica’s L10. Other key listed partner include Innolight for optical module, Isu Petasys for PCB, and Furukawa for cold plate.

For AWS supply chain, two important hardware pure plays are both in Taiwan – Accton and Alchip - both have >50% sales to AWS. Accton does the Trainium accelerator board and majority of switch, while Alchip does back-end design service for Trainium. Investors worried a lot on Accton's share in Trainium accelerator board but ignored the acceleration of switch upgrade - Google being the first one migrating to 1.6T from 2H26, and Amazon who is in 800G now need to speed up migration. Other key listed partners of AWS include Astera Lab for PCIe, FLEX for server assembly, Credo for AEC.

We like both Celestica and Accton, while neutral on Alchip due to design service industry issue (we are neutral on whole design service industry).

We do not like Nvidia AI GPU server plays such as SMCI, Dell, Hon Hai, given higher competition with more players, and lower margin due to standardization.