1T-Parameter Large Models Now Run on Desktops: Apple & NVIDIA Reshape the AI PC Industry Landscape - eu.36kr.com

This is not a product refresh—it is a structural migration of compute sovereignty from the cloud to the endpoint. Technical cascade: Running a 1T-parameter model locally demands 200GB+ of effective bandwidth per device. HBM4 is spilling out of data centers into consumer silicon, and CoWoS-L packaging capacity will be the scarcest bottleneck industry-wide through 2026. Apple's M4 Ultra and NVIDIA's RTX 5090 are compressing data-center-class memory-wall problems into a 300W TDP envelope—thermal and power architectures must be rewritten from scratch. Compliance & risk: Under BIS export controls, advanced packaging equipment access for China remains tightly restricted. With over 85% of global CoWoS capacity concentrated in Taiwan, China, any geopolitical disruption triggers a 6-9 month delivery gap. Compliance overhead is now baked into BOM at 15-20%. Market dynamics: Intel's Arc and AMD's Strix Halo are essentially follower pricing plays. The real wildcard is Apple—once the M-series simultaneously captures AI inference workloads across MacBook and iPhone, NVIDIA loses its "sole option" premium in consumer desktop. Forward look: Within 18 months, CXL memory expansion and LPDDR6 will shatter the linear "VRAM equals performance" assumption. The TCO crossover between local and cloud inference is projected for Q2 2026—by then, "AI PC" stops being a marketing tagline and becomes an economically rational allocation of compute.