AI Infrastructure Daily — June 24, 2026

California SB 53 frontier model safety provisions take full effect. Google announces expanded 1GW advanced nuclear portfolio. CoreWeave ships first GB200 NVL72 racks to customer datacenters.

California SB 53 Frontier Model Safety — Full Effect

California's SB 53, signed by Governor Newsom in October 2025, reaches full enforcement today. The law requires frontier AI developers training models above defined compute thresholds to report safety incidents, conduct third-party safety evaluations, and maintain AI emergency response plans.

Unlike the vetoed SB 1047, SB 53 focuses on safety incident reporting and whistleblower protections rather than blanket liability. Key provisions include mandatory reporting within 72 hours of critical safety incidents, with civil penalties up to $100,000 per violation.

All major AI labs with California operations — including OpenAI, Anthropic, Google DeepMind, Meta, and xAI — are now subject to these requirements. The California AI Safety Task Force begins enforcement operations this week.

Google Expands Advanced Nuclear Portfolio to 1GW

Google announced an expanded advanced nuclear energy portfolio totaling 1GW of capacity, through additional PPAs with Kairos Power, Oklo, and Terrapower. This builds on the initial 500MW Kairos agreement from October 2024.

The expanded portfolio is designed to power Google's AI datacenter expansion through the 2030s, with first reactors expected online in 2030. Google aims to achieve 24/7 carbon-free energy for all datacenters by 2030.

This follows similar moves by Microsoft (Three Mile Island, Quad Cities, Crane restarts) and Amazon (X-energy, SMR investments), as hyperscalers increasingly turn to nuclear power to meet surging AI compute demand.

CoreWeave Ships First GB200 NVL72 Racks

CoreWeave announced the deployment of its first NVIDIA GB200 NVL72 racks at the Quintus-X AI Campus in West Columbia, Texas. Each NVL72 rack contains 72 Blackwell GPUs and 36 Grace CPUs, delivering 13.8TB of HBM3e memory and 648 PFLOPS of FP4 compute.

The Quintus-X campus, backed by Blackstone financing, is scaling to 350MW with GB200 clusters leased to AI labs for frontier model training. CoreWeave's $7.5B debt facility secured against GPU inventory is funding the massive infrastructure buildout.

Pricing for GB200 NVL72 cloud access remains TBD, but industry estimates suggest $30-40/hr per GPU equivalent — making full rack-hour pricing potentially $2,500+/hr.

South Korea AI Basic Act — First Comprehensive Asian AI Law

South Korea's AI Basic Act, enacted in December 2025, is now fully in effect as of January 2026. As the first comprehensive AI law in Asia, it establishes the AI Committee, requires transparency for high-impact AI systems, and mandates risk assessments.

The law affects all AI developers and deployers in South Korea, including foreign companies serving Korean users. High-impact AI obligations take effect July 1, 2026, covering sectors including healthcare, finance, and public services.

With Japan pursuing voluntary guidelines and China maintaining strict content moderation measures, South Korea's approach represents a middle ground — comprehensive regulation with pro-innovation industry support provisions.

GPU Market Update: B100 and H200 PCIe Now Available

NVIDIA's B100 GPU, the lower-power Blackwell variant (700W vs 1000W for B200), is now shipping via CoreWeave at $2.80/hr. The B100 offers 192GB HBM3e memory and 3,600 FP8 TFLOPS — making it attractive for mainstream training and inference workloads where power is constrained.

The H200 PCIe variant has also reached general availability at $2.10-2.15/hr via CoreWeave and RunPod. With 141GB HBM3e memory, it's positioned as a high-memory inference GPU for large model serving.

AMD's MI325X is gaining traction at $2.49/hr via CoreWeave, offering 288GB HBM3E — the largest memory capacity of any shipping GPU. This makes it particularly attractive for large-model inference workloads.