Amazon's strategy of developing Trainium in-house has compounded problems, with Anthropic finding Trainium ineffective, prompting Amazon to invest heavily in OpenAI to fill idle Trainium capacity.
Marvell's custom ASIC partnerships for Microsoft's Maia and Amazon's Trainium have resulted in postponed deployment schedules, lackluster system performance, and lower effective performance per dollar, indicating they are not effective bottleneck solvers.
Broadcom's design partnership model for custom ASICs has proven to be the winning template, with an enormous pipeline underpinning its thesis as a secular winner, addressing the compute bottleneck caused by NVDA's Blackwell delays.
AXT Inc. is a supplier of small-diameter InP wafers, which are identified as a bottleneck causer due to low yields and long lead times, making it unable to scale for 1.6T transceiver demand and vulnerable to SiPho solutions.
Tower Semiconductor is identified as the biggest winner from the SiPho transition, operating the industry's most mature and highest-performance SiPho process node, uniquely qualified at 1.6T, and planning significant capacity expansion.
Innolight, as the world's largest optical transceiver producer, partnered early with TSEM on SiPho 1.6T, granting it a meaningful time-to-market advantage as SiPho resolves the optical interconnect bottleneck.
Lumentum is a major producer of EML lasers, which are fabricated on Indium Phosphide (InP) wafers, a supply chain bottleneck severely constrained by physics and economics, making it vulnerable to the SiPho transition.
The industry is undergoing a fundamental architectural shift towards optical interconnects, implicitly conceding that copper, even with AECs supplied by Credo, cannot scale to meet next-generation bandwidth demands.