Nvidia GTC Berlin: Faster AI factories, lower power use, more secure agents - Network World

GTC Berlin's real signal isn't the FLOPS headline—it's Nvidia pivoting the competitive axis from peak compute to "effective inference tokens per watt, in a verifiable execution context." This fractures the upstream stack: tighter per-die power budgets make 800V HVDC and immersion cooling non-optional, rewriting BOMs for Vertiv, Schneider, and peers. Secure agents demand hardware enclaves baked into silicon—no software patch substitutes—forcing a re-partition of next-gen GPU dies and compounding CoWoS yield pressure. On compliance, the Berlin venue is a deliberate alignment play with the EU AI Act and GDPR. Auditable inference chains raise enterprise deployment costs, but simultaneously build a moat that AMD's ROCm ecosystem and custom-ASIC camps can't replicate within 18 months of certification cycles. Strategic counter: AMD will likely push MI350's "open architecture + lower TDP" narrative, but secure-enclave IP isn't built overnight. The real wildcard is the ASIC camp—when TPU v6 and Trainium 2 embed secure inference into silicon, Nvidia's full-stack premium gets compressed. 12–24 month call: the power wall, not the compute wall, becomes the binding constraint on datacenter expansion. Procurement criteria shift from GPU count to "secure inference tokens per kWh"—a fundamentally new pricing anchor. Europe emerges as a third infrastructure market on sovereign-AI narratives, yet the bottleneck remains locked in advanced packaging capacity in Taiwan, China.