The physical substrate of intelligence is power.

Vera Rubin NVL72, Groq 3, & Gaudi 3

AI accelerators are scaling transistor counts and interconnect bandwidth. The NVIDIA Blackwell B300 (192GB HBM3e) is being succeeded by the Vera Rubin NVL72 system (combining 72 Rubin R100 GPUs and 36 Vera CPUs) delivering 3.6 exaflops of FP4 compute. Alternative architectures like the Groq 3 LPX Rack leverage 128GB of aggregate ultra-low latency SRAM to bypass HBM memory bottlenecks entirely.

• ▸NVIDIA Rubin R100: 336B transistors, 288GB HBM4, 22 TB/s bandwidth, 50 PFLOPS FP4 compute.
• ▸Intel Gaudi 3: 1,835 TFLOPS BF16, 128GB HBM2e, 3.7 TB/s bandwidth, priced disruptively at $15,625.
• ▸Groq 3 LPX Rack: 256 LPUs, 500MB SRAM each (128GB aggregate), 640 TB/s scale-up bandwidth.

From Megawatts to Dedicated Power Grids

Datacenter design has scaled past the megawatt envelope into multi-gigawatt campuses. Building queues for regional utility connections stretch up to five years, prompting AI developers to build their own generation infrastructure on-site, pre-leasing 60% of the 35 GW under construction in North America.

• ▸Meta Prometheus & Hyperion: 1 GW Prometheus online by 2027 in Ohio; 5 GW Hyperion campus in Louisiana with $3B on-site gas generation.
• ▸xAI Colossus (Memphis): Running 100K H100s (150MW), expanding to 200K, with a long-term roadmap targeting 1M GPUs.
• ▸Microsoft & Google Buildouts: Microsoft Wisconsin ($7B) and Texas (2.1 GW); Google Oklahoma ($9B) and Michigan (1 GW).