NVIDIA's GB200 NVL72 — a rack-scale system pairing 36 Grace CPUs with 72 Blackwell B200 GPUs across a high-bandwidth NVLink fabric — has entered full production volume, with supply chain sources indicating that Taiwan Semiconductor Manufacturing Company's CoWoS-L advanced packaging lines are now running at target capacity after months of yield ramp challenges.
The NVL72 is architecturally distinct from prior NVIDIA data-centre products. Each rack functions as a single logical compute unit: the 72 B200 GPUs share a unified 1.4TB GPU memory pool accessible at NVLink speeds of 1.8 TB/s, eliminating the PCIe bottlenecks that constrain GPU-to-GPU communication in conventional server designs. For large language model training at the frontier scale — models with hundreds of billions of parameters — this architectural change reduces inter-GPU communication overhead by roughly 60% compared to A100 or H100 cluster designs using InfiniBand.
Hyperscale cloud providers are the primary customers. Amazon Web Services, Microsoft Azure, Google Cloud, and Oracle Cloud Infrastructure have all publicly disclosed Blackwell deployments or partnerships, and industry analysts estimate that combined GB200 allocations for calendar year 2026 will represent over $50 billion in revenue for NVIDIA, setting a new record for a single product line in semiconductor history.
The inference economics of GB200 are drawing particular attention from AI startups and enterprise customers. Because the NVL72 can serve multiple concurrent inference requests across a shared memory pool without the expensive model re-loading that individual GPU deployments require, the effective cost per token for large models drops significantly compared to H100 clusters. Early benchmarks from cloud providers suggest GPT-4 class model inference costs 35–45% less per token on GB200 infrastructure compared to equivalent H100 configurations.
Competitive pressure is also building. AMD's MI400 accelerator is scheduled for production release in Q4 2026, and several AI chip startups including Groq, Cerebras, and d-Matrix have positioned their products as alternatives for specific inference workloads. Intel's Gaudi 3 continues to undercut NVIDIA on price for medium-tier deployments.
Despite this, NVIDIA's software ecosystem — particularly the CUDA programming model and the NCCL collective communication library — continues to give it a durable advantage in a market where rewriting training frameworks for new hardware architectures carries substantial engineering risk. Most AI labs remain committed to CUDA-native tooling for the foreseeable future.