NVDA Inference Share Seen Falling to 50% by 2028 as AMD Gains Pre-Market 4.5%

A subtle but significant narrative shift is emerging in semiconductor sentiment: the focus is moving from AI capex growth acceleration to concerns about competitive fragmentation and inference-market share erosion. An analyst comment circulating this week claimed that NVIDIA's inference revenue could decline to 50% of the market by 2028 as AMD, Google TPU, AWS Trainium, Meta Maia, and specialized SRAM chip vendors capture share. This claim, unverified but gaining traction, has prompted pre-market strength in AMD and Super Micro (SMCI), both up around 4.5% this morning.

The bull case for AMD rests on three pillars: (1) inference workloads are becoming more commoditized and price-sensitive; (2) hyperscalers are incentivized to build custom silicon to reduce vendor lock-in; (3) AMD's EPYC and data-center roadmaps offer viable alternatives to NVIDIA's continued GPU dominance. The counter to the AI capex boom narrative is not that capex is slowing, but that the capex spend is diversifying across multiple chip architectures and vendors.

Funding flows support this view. Google, Microsoft, and Amazon are indeed deploying $250B+ into AI infrastructure, but that capital is split across training (GPUs) and inference (CPUs, TPUs, custom silicon). The implication is that absolute demand for compute remains robust, but NVIDIA's margin and share-of-wallet shrink if customers mix in alternatives. This is why NVDA's earnings call commentary on competitive dynamics and pricing power matters so acutely today.

Skeptics counter that NVIDIA's software ecosystem (CUDA) and architectural advantages in training remain unmatched, and that hyperscalers' custom silicon often complements rather than replaces NVIDIA products. The risk for AMD bulls is that competitive threats prove overblown. However, the narrative has clearly shifted from 'AI capex runaway' to 'AI capex diversification,' which is a material de-rating scenario if sustained.