AI Infrastructure Daily — June 23, 2026

Oracle-OpenAI Stargate Phase 1 declared operational in Abilene. xAI Memphis expands to 300K GPU target. EU AI Act Annex III deadline approaches in 70 days.

Stargate Abilene Campus Reaches Phase 1 Operational Status

The Oracle-OpenAI Stargate campus in Abilene, Texas has officially reached Phase 1 operational status. The facility currently hosts approximately 100,000 NVIDIA H100 GPUs with 300MW of allocated power. Phase 2 construction in Mount Pleasant, Wisconsin is already underway, with Microsoft securing 837MW of nuclear power from Constellation Energy's Three Mile Island Unit 1 restart.

This marks a critical milestone in the Stargate project, which aims to build the largest AI training infrastructure in the world. The Abilene campus is expected to scale to over 1GW total capacity across multiple phases.

xAI Memphis Supercluster Targets 300,000 GPU Expansion

xAI's Memphis facility — already the world's largest operational AI training cluster with 100,000 H100s — is planning a massive expansion to 300,000 GPUs. The expansion includes NVIDIA H200s and GB200 NVL72 systems. Power remains the primary constraint, with xAI relying on mobile gas turbine generators to supplement grid power from TVA.

The expansion would make the Memphis cluster roughly 3x the size of any other operational AI training facility globally.

EU AI Act: 70 Days Until Annex III High-Risk Obligations Take Effect

With the August 2, 2026 deadline approaching, organizations deploying high-risk AI systems (as defined by Annex III of the EU AI Act) must prepare for compliance. Key obligations include risk assessments, technical documentation, conformity assessments, and post-market monitoring.

The Annex III categories include: biometric identification, critical infrastructure, education and vocational training, employment and worker management, essential public and private services, law enforcement, migration and border control, and administration of justice.

GPU Market Update: B200 Volume Increasing, H200 Widely Available

NVIDIA B200 shipments are ramping through Q2 2026, with CoreWeave offering cloud access at $3.50/hr. H200 SXM5 units are now widely available across neoclouds at approximately $2.49/hr. H100 allocation remains constrained for new customers, though existing hyperscaler contracts continue to be fulfilled.

AMD MI300X adoption is accelerating, with Meta, Microsoft, and Oracle expanding deployments. The MI300X offers 192GB of HBM3 memory — 2.4x the H100's 80GB — making it increasingly attractive for large-model inference workloads.

Saudi Arabia's Project Transcendence: $100B Sovereign AI Bet

Saudi Arabia's PIF, in partnership with Google, announced Project Transcendence — a planned 1GW AI datacenter in the Eastern Province. At $100 billion, it represents the largest single AI infrastructure investment announced to date. The facility is positioned as sovereign AI infrastructure, aiming to make Saudi Arabia a regional hub for AI training and inference.

The project highlights the global race for AI infrastructure sovereignty, with multiple nations (UAE, Saudi Arabia, India, Japan) investing in domestic AI capacity.