Cloud giants to disrupt the semiconductor scene with their in-house AI accelerators?

Jay Liu, Taipei; Julie Chang, DIGITIMES Asia 0


Despite the achievements of Nvidia, companies developing their own AI models are moving towards custom chips as costs escalate, potential compatibility constraints surface, and to circumvent supply shortage from reliance on a solitary provider.

Major players in the industry have respectively already reached varying stages in their development. Google, as a long-standing contributor in the field of AI, is already on its fourth iteration of its in-house Tensor Processing Unit (TPU). Other companies are in hot pursuit. Meta's inaugural custom AI chip, Meta Training Inference Accelerator (MTIA), is slated to go live in 2025. Moreover, in the wake of its staggering success post the deployment of OpenAI, Microsoft has also decided to throw its hat into the AI hardware arena. The cloud service behemoth, Amazon, though, with its duo of chips - Inferentia and Trainium - under its belt, has determined that software development will remain its major focus.

The decision to develop in-house AI chips isn't just about diversifying supply chains; it's largely about fulfilling operational prerequisites. Despite the superior performance of Nvidia's AI accelerators, the exorbitant computational burden and the associated spiraling costs could deter cloud providers that boast unique AI models and bespoke computing prerequisites. In-house chips will curtail the costs for acquiring AI hardware and also tackle compatibility issues that could likely arise from employing uniform AI chips.

The decision of cloud providers to assert control will also give rise to alliances and revenue prospects for teams specializing in the design of application-specific integrated circuits (ASICs), as most cloud companies grapple with challenges such as talent shortage, inadequate IP block reserves, and a steep entry barrier in their chip development journey.

Rumor has it that Microsoft is set to bolster its strategic partnership with Advanced Micro Devices (AMD), while Meta is extending its chip technology collaboration with Broadcom on AI and metaverse. Taiwan-based IP and ASIC powerhouses like MediaTek, Global Unichip, and Alchip Technologies are likely contenders for a stake in the AI and high-performance computing (HPC) realm. It could well be possible that these alliances will pose a significant challenge to the current semiconductor titans.