NVIDIA's Blackwell GPUs are shipping at scale. We break down what this means for hyperscaler capex cycles, competitive dynamics with AMD and custom silicon, and whether the infrastructure moat is durable.
NVIDIA's Blackwell GPUs are shipping at scale. We break down what this means for hyperscaler capex cycles, competitive dynamics with AMD and custom silicon, and whether the infrastructure moat is durable. The GPU market is undergoing a structural shift. NVIDIA's dominance in AI training and inference hardware has been the defining feature of the AI infrastructure buildout since 2023. But as Blackwell ramps, the competitive landscape is evolving in ways that matter for long-term capital allocation. ## The Blackwell Advantage Blackwell represents a generational leap in several key metrics. The B200 delivers roughly 2.5x the training performance of the H100 while improving energy efficiency by 25%. For inference workloads — increasingly the dominant use case — the improvement is even more pronounced. This matters because the transition from training-dominated to inference-dominated workloads fundamentally changes the economics. Training is a capex event. Inference is an opex event. Companies that invested heavily in H100 clusters for training are now evaluating whether Blackwell's inference efficiency justifies a hardware refresh cycle. ## Hyperscaler Capex Implications Microsoft, Google, Meta, and Amazon collectively spent over $200B on AI infrastructure in 2025. The Blackwell transition creates a complex capital planning decision. Early data suggests Blackwell inference costs are 40-60% lower per token than H100-based setups, which means the ROI on migration is compelling — but the timing depends on workload mix. Our analysis of hyperscaler earnings calls and capex guidance suggests the Blackwell upgrade cycle will play out over 18-24 months, with Microsoft leading (already deploying at scale) and Amazon trailing (still ramping Trainium as a hedge). ## The Competitive Response AMD's MI350 is a credible competitor for the first time. Benchmarks show it within 15-20% of Blackwell on standard training workloads, and AMD's software ecosystem (ROCm) has matured significantly. However, AMD's real opportunity is in inference,