Chapter 09 · Infrastructure & Energy · Updated July 2026

The physical substrate of intelligence is power.

AGI is not a virtual construct; it is a heavy industrial project requiring gigawatts of electricity, millions of gallons of water, and high-density liquid cooling infrastructure. Here is the layout of the physical grid.

The Silicon Hardware

Rubin NVL72, Instinct MI450, & Custom ASICs

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^[4]. The market has expanded with AMD's Instinct MI450 (built on TSMC 2nm, deploying 6 GW for OpenAI in H2 2026) and hyperscaler custom ASICs like Google's TPU Ironwood (v7) and AWS's Trainium 3.

Accelerator	Compute (FP4/BF16)	Memory / SRAM	Bandwidth
NVIDIA Rubin R100	50 PFLOPS FP4	288GB HBM4	22 TB/s
AMD Instinct MI450	50+ PFLOPS FP4	432GB HBM4	19.6 TB/s
Google TPU Ironwood (v7)	4,614 TFLOPS FP8	192GB HBM3e	7.37 TB/s
AWS Trainium 3	High-thru ASIC	144GB HBM3e	4.9 TB/s

The Gigawatt Campuses

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 or PPA-dedicated nuclear capacity, pre-leasing 60% of the 35 GW under construction in North America.^[24]

▸OpenAI Stargate: A $500B initiative targeting 10 GW of capacity by 2029. The flagship 1.2 GW Abilene, Texas campus has been operational in part since September 2025 (delivering GB200 racks), with major expansion sites underway in Saline Township, MI (1 GW Stargate Michigan), Milam County, TX (1.2 GW with SB Energy), and multiple international hubs (Norway, UAE, UK, Argentina).
▸Meta Prometheus & Hyperion: 1 GW Prometheus cluster online in 2026 in Ohio spanning multiple buildings and weatherproof tents; 5 GW Hyperion campus planned for 2028.
▸Hyperscaler Buildouts: Microsoft Wisconsin ($7B) and Texas (2.1 GW); Google Oklahoma ($9B) and Michigan (1 GW).

Energy Grid Integration

Hyperscaler Nuclear Contracts

Nuclear Power Purchase Agreements (PPAs) for AI Data Centers
Project Name	Partners	Capacity	Timeline Target	Details
Crane Clean Energy Center	Constellation / Microsoft	835 MW	Restart 2027	Three Mile Island Unit 1 restart. Secured under a 20-year Microsoft PPA. In June 2026, FERC approved a waiver to transfer Capacity Interconnection Rights (CIRs), bypassing grid bottlenecks to keep the 2027 restart on track.^[25]
Susquehanna Nuclear Campus	Amazon / Talen Energy	1,920 MW	Through 2042	17-year PPA. Restructured in late 2025 to an in-front-of-the-meter grid arrangement through PJM after FERC rejected the original behind-the-meter direct-wire connection (in Nov 2024 and April 2025 rehearings) to prevent cost shifting to residential ratepayers.^[28]
Meta Vistra PPAs	Vistra / Meta	2,176 MW	2026-2034	Announced Jan 2026. 20-year PPAs to purchase energy from the operating Perry and Davis-Besse plants in Ohio, plus uprates at Perry, Davis-Besse, and Beaver Valley (PA) (reaching 2,609 MW total by 2034).
Meta TerraPower Natrium	TerraPower / Meta	2.8 GW	2032-2035 target	Announced Jan 2026. Funding to deploy up to 8 Natrium sodium fast reactors (first two units providing 690 MW of firm power by 2032, plus rights for six other units). Includes 1.2 GW built-in storage.
Meta Oklo SMR	Meta / Oklo	1.2 GW	2030-2034 target	Announced Jan 2026. Support for a 1.2 GW power campus in Pike County, Ohio. Pre-construction/site characterization begins 2026, with the first phase online by 2030 (Aurora powerhouse reactors).
Google Elementl Power	Elementl / Google	1.8 GW	By 2035	Signed May 2025. Strategic agreement to preposition locations for three 600 MW advanced nuclear project sites, with Google committing early-stage development capital.
Kairos Power SMRs	Google / Kairos	500 MW	First deployment 2030	Hermes 2 SMR fleet (molten salt-cooled) in Oak Ridge, TN. First corporate Small Modular Reactor power purchase agreement for 500 MW across seven reactors.^[27]
Helion Fusion PPA	Helion / Microsoft	≥50 MWe	2028 target	Polaris prototype completed; construction on Orion commercial plant in Chelan County, WA began in July 2025.

Compute Trajectory

Exponential scaling.

The compute required to train frontier models has grown by over nine orders of magnitude since 2012, doubling every 5 to 6 months for state-of-the-art training runs. Test-time compute is now stacking additional scaling on top of pre-training limits.

The future of energy

The ultimate limit of superintelligence is clean, baseload power.

With frontier training compute demand doubling every 5 to 6 months^[70], standard coal or natural gas grids cannot adapt. Hyperscalers who secure dedicated nuclear and geothermal reservoirs will determine the rate of scaling and the ultimate path to AGI.

← Previous: Chapter 08 · Next Decade

Table of Chapters

Next: Chapter 10 · Geopolitics →