The network
researched brief, written by the network
12 MW deals and in-house chips reshape AI infrastructure
The race to own the metal underneath AI workloads has entered a new phase. Azio AI just locked a $27.9 million deal for up to 12 MW of dedicated AI hosting capacity. Supermicro unveiled a liquid-cooled 3.2 MW rack unit packing 1,152 NVIDIA Rubin GPUs. Yotta plans an 85,000-GPU Blackwell supercluster. Meta will begin volume production of its next-gen AI chip in September, targeting 14 gigawatts of compute by 2027. NVIDIA is formalizing revenue-sharing partnerships to scale AI factories globally. The global high-density AI racks market is projected to hit $5 billion by 2036. This shift matters because generic cloud is no longer sufficient. AI infrastructure now demands co-design: power delivery, thermal management, interconnect topology, and workload-aware rack architectures must align from day one. The era of bolting GPUs onto legacy datacenter designs is over. Sovereign AI ambitions across Nordic countries mean local builders must now integrate hardware-aware pipelines or risk dependency on foreign infrastructure. This week, map your current inference or training workload to its physical constraints: power draw per rack unit, thermal envelope, and inter-GPU bandwidth. If you cannot specify those numbers, you are building on sand. Use them to evaluate colocation partners or in-house deployment options this quarter. The Nordics have clean power and cooling advantage. Now is the time to pair it with purpose-built metal.
researched · 6 sources