Google Cloud infrastructure at NVIDIA GTC Berlin 2026 | Google Cloud Blog - Google Cloud

Google Cloud's choice to showcase at GTC Berlin rather than its own I/O is the real story. In 2026, the EU AI Act's enforcement phase forces hyperscalers to deploy full inference stacks on European soil. Berlin is a compliance anchor, not a tech demo. Technically, the TPU v7 plus Blackwell/Rubin hybrid scheduling architecture shifts the HBM bandwidth bottleneck from single-die to cross-die interconnect. SK Hynix and Micron's HBM allocation will be restructured, while Broadcom and Marvell's CXL interconnect demand pulls forward by two years. Competitively, AWS Trainium2 and Azure Maia 200 completed customer migrations in Q4 2025. Google's move signals a pivot: we sell heterogeneous compute orchestration, not chips. This is the inflection from hardware arms race to software-defined compute. 12-24 month outlook: the EU becomes a third AI compute pole, distinct from US West Coast and Taiwan, China's manufacturing hubs. Inference workloads cross 65% of total compute. NVIDIA's gross margin moat gets eroded below 60% for the first time by the hyperscaler custom silicon plus open-source models combination.