The 3-Stock Custom Silicon Basket That Could Outperform Nvidia by 2030 - The Globe and Mail

As artificial intelligence continues to evolve, traditional general-purpose GPUs are increasingly insufficient to meet the growing demand for customized AI computing. This article highlights that cust... Read original →

The surge in custom AI silicon is triggering a deep-stack reconfiguration: EUV and 3nm capacity faces upward pressure upstream, while downstream inference architectures rapidly decouple from general-purpose GPU paradigms. TSMC in Taiwan, China—commanding 90% of advanced node output—emerges as the ultimate beneficiary, yet geopolitical friction inflates supply chain redundancy costs. U.S. CHIPS Act mandates and export controls compel Broadcom and Marvell to shift select design blocks to domestic foundries, despite yield and cost disadvantages. NVIDIA, facing ASICs growing at 45% annually versus GPUs’ 15%, may be forced to open CUDA to retain cloud alliances. Microsoft and Alphabet are leveraging in-house TPUs/ASICs to assert compute sovereignty. Over the next 18 months, the industry enters an 'architectural bifurcation' phase—general and specialized chips co-evolve, but capital allocation has decisively tilted toward customization, eroding GPU hegemony.