Amazon’s (AMZN) $200 billion AI spending plan is rewriting the rules of the chip game.
Last week, the tech giant unveiled its 2026 capital expenditure budget: $200 billion earmarked for AI data centers, a nearly $70 billion increase from 2025. The market immediately thought of Nvidia. But Amazon’s custom chip push has brought another company into the spotlight: Marvell Technology (MRVL).
Speaking at an investor conference, Amazon CEO Andy Jassy noted that Nvidia GPUs will remain the workhorses in upcoming data centers. Yet a faster growth curve is taking shape: AWS’s custom chip business has surpassed a $10 billion annual run rate, with triple-digit year-over-year growth.
Jassy is particularly bullish on the Trainium series. Trainium2 recorded the fastest demand ramp-up in company history. Trainium3, announced last December, is expected to be fully committed by mid-2026. And Trainium4, though not yet launched, is already generating strong customer interest.
Marvell is the design partner behind the Trainium chips. In late 2024, the two companies signed a five-year agreement for Marvell to supply a full range of chip designs for AWS data centers—from AI accelerators to networking chips covering interconnect, switching, and storage.
The market has fretted that Amazon might switch design partners, with rumors surrounding Trainium3 and Trainium4 weighing on Marvell’s stock. But Marvell CEO Matt Murphy pushed back in December: “Nothing has changed in our business outlook since the Trainium3 announcement.”
Marvell’s ties to Amazon go beyond AI accelerators. Networking chips currently represent a larger chunk of its business, and these designs can seamlessly migrate from Trainium2 to next-generation platforms. The recent acquisition of Celestial AI further strengthened Marvell’s moat in AI-focused interconnect chips.
Marvell also designs Maia-series AI chips for Microsoft. Despite concerns that Microsoft might seek alternative suppliers, management expects custom AI accelerator revenue to ramp sharply in fiscal 2028, aligning with the production timeline for Microsoft’s Maia 300 chip.
Compared to AI chip peers, Marvell’s stock hasn’t overheated. The shares trade at just 22.6 times forward earnings estimates. As hyperscalers collectively pivot toward custom silicon, companies like Marvell—deeply embedded in these customized designs—stand to enjoy both revenue and margin upside.
Nvidia’s golden age continues, but Marvell’s story is just beginning.
