Marvell Technology announced that it has lost Amazon Web Services’ Tranium 3 and Tranium 4 custom‑chip programs to Taiwanese competitor Alchip. The loss removes a key source of custom silicon work that had been a cornerstone of Marvell’s strategy to win hyperscaler contracts. Amazon’s Tranium line, which began with Tranium 2, is a core part of AWS’s plan to reduce reliance on third‑party vendors and to deliver higher performance for AI training and inference workloads.
Amazon’s decision to award the Tranium 3 and 4 designs to Alchip was driven by a combination of technical and supply‑chain considerations. Amazon licensed SERDES technology from Synopsys and partnered with Alchip for design, foundry, and backend support, a package that Marvell was unable to match after delays and execution challenges with Tranium 2. The shift underscores the importance of end‑to‑end integration and rapid delivery in the hyperscaler silicon market.
Marvell’s Q3 fiscal 2026 results showed record revenue of $2.075 billion, up 37% year‑over‑year, with data‑center revenue rising 37.8% YoY. Custom‑chip revenue, while lower in margin, is projected to reach $2.5 billion by fiscal 2026, reflecting the company’s focus on scaling its AI‑chip business. The loss of Amazon’s programs is a significant headwind, but the broader custom‑chip pipeline remains robust, and Marvell’s overall financial health is supported by strong demand in its data‑center and optical‑chip segments.
Amazon’s AI infrastructure continues to expand, with AWS reporting a 20% year‑over‑year revenue increase to $33 billion in Q3 2025, the fastest growth since 2022. The company is actively developing Tranium 3 and 4, with Tranium 4 slated to integrate Nvidia’s NVLink Fusion technology, signaling a hybrid approach that blends in‑house silicon with external GPU capabilities. The loss of Marvell’s involvement does not materially alter Amazon’s silicon roadmap, which remains focused on performance and cost efficiency.
Investors reacted with caution after the announcement, citing concerns about Marvell’s future growth in the custom‑chip market and the potential shift of Microsoft’s custom‑chip work to Broadcom. Benchmark analysts highlighted that Marvell’s XPU growth could slow to about 20% in 2026, a downgrade from earlier expectations. The event illustrates the competitive pressure hyperscalers exert on chip designers and the importance of securing long‑term contracts.
The episode highlights the volatility of the custom AI‑chip ecosystem, where hyperscalers can pivot quickly between partners. Marvell will need to accelerate its win‑rate on new hyperscaler contracts to offset the loss of Amazon’s programs, while Amazon continues to invest heavily in its own silicon to maintain a competitive edge in AI compute. The market will watch how both companies navigate the evolving demand for high‑performance, low‑latency AI workloads.
The content on BeyondSPX is for informational purposes only and should not be construed as financial or investment advice. We are not financial advisors. Consult with a qualified professional before making any investment decisions. Any actions you take based on information from this site are solely at your own risk.