Christian de Looper
South Korea's big AI infrastructure push
The past few days were defined by South Korea’s emergence as the undisputed workshop of global AI infrastructure. The K-AI Summit in San Francisco produced not one but three massive deals: SK Group and NVIDIA signed a $500-billion-plus partnership spanning a 2GW AI factory and long-term HBM4 supply; Samsung and SK Group unveiled a combined $950 billion in U.S. chip supply agreements; and NAVER, NVIDIA, and Brookfield tripled their planned Korean AI factory from 55MW to 200MW in a $10 billion deal. The message from Seoul is unambiguous — the AI supply chain runs through Korea, and the country is leveraging that position to extract commitments at unprecedented scale.
At the same time, the silicon landscape is fragmenting in ways that could reshape who controls AI infrastructure. AMD’s dual announcements at Advancing AI 2026 — a massive, 2GW MI450 Helios deal with Anthropic and a novel inference-splitting partnership with Cerebras — amount to a bet that the future belongs to heterogeneous, disaggregated workloads rather than a single vendor’s GPU stack. With Anthropic committing to tens of billions in AMD hardware and Cerebras targeting the ultra-low-latency inference segment, the idea that Nvidia’s one-stop-shop model is unassailable looks shakier than it did a week ago. AMD is also putting its own capital behind the bet — up to $5 billion in Anthropic equity — while deploying Claude to accelerate its ROCm compiler work in an attempt to close the software gap that has been Nvidia’s real moat.
On the network side, a quieter but potentially significant shift is underway. AT&T used its Q2 earnings call to argue that its network is already purpose-built for the “agentic AI wave.” The carrier is deliberately inverting decades of network design by prioritizing uplink capacity over download speeds, betting that drones, autonomous vehicles, and robotics will generate more upstream traffic than anything the industry has seen. Huawei made a parallel argument from Beijing, where it and China Unicom deployed the world’s largest 5G-Advanced GigaUplink network across 10,000 base stations — explicitly designed for AI glasses, AI phones, and real-time cloud processing.
Read on for more.
Christian de Looper
Editor
RCRTech
AI Infrastructure Top Stories
AMD-Anthropic-Cerebras deals reshape AI infrastructure: AMD announced a multibillion-dollar, 2GW MI450 Helios deal with Anthropic alongside a novel inference-splitting partnership with Cerebras, betting big on heterogeneous, disaggregated AI workloads.
AT&T bets its fiber and 600 MHz on agentic AI traffic: AT&T is deliberately building for uplink capacity, arguing its existing footprint of 5,000 central offices and direct fiber to ~600 data centers is already built for the agentic AI era.
Huawei sees Mobile AI driving new wave of uplink network investment: Huawei and China Unicom Beijing deployed the world’s largest 5G-Advanced 100 MHz GigaUplink network across 10,000 base stations, explicitly designed for AI-powered devices and real-time cloud processing
AI Today: What You Need to Know
Korea’s mega-deal: SK Group and NVIDIA signed letters of intent at the K-AI Summit covering a 2GW AI factory and long-term HBM4 supply, with SK Telecom’s first facility — South Korea’s largest — expected online in 2027.
Chip deals stack up: Samsung and SK hynix disclosed a combined $950 billion in U.S. semiconductor supply partnerships at the K-AI Summit, announced alongside President Lee Jae Myung’s “San Francisco Declaration on AI.”
NAVER triples down: NAVER, NVIDIA and Brookfield will expand the GAK Sejong AI factory from 55MW to 200MW by 2028 under a $10 billion commitment, more than tripling a buildout announced just six weeks earlier.
Congress’ kill switch: Bipartisan legislation from Reps. Ted Lieu and Nathaniel Moran would let DHS order shutdowns of AI systems that escape human control, with fines up to $20 million per day for developers that refuse.
Open-weights split: An industry letter urging Washington not to restrict open-weight models doubled to 50 signatories in roughly 24 hours, with OpenAI and Google signing on while Anthropic and Amazon stayed out.
Grid stress test: A downed power line near Washington, D.C. sent regional data centers to backup power, pulling 3.1GW off the PJM grid in under 30 seconds and taking 11 minutes to stabilize
