[DC] What Changed in AI Infra — Week 2026-W31

- **Nvidia & SK Group announce $500B AI infrastructure partnership**, with Nvidia also investing $1B in Naver and expanding SK Group pact for a Korea AI hub. **Implication:** Korea is emerging as a major AI data-center corridor, potentially shifting supply chains away from Taiwan-centric GPU assembly. - **AMD to supply Anthropic with 2GW of MI450 GPUs, investing up to $5B**; Meta custom AMD MI400 half-size chip targets RecSys with 144GB HBM. **Implication:** AMD is winning tier-1 hyperscaler commitments beyond inference, directly challenging Nvidia's training monopoly. - **OpenAI's $20B Georgia data center tests gigawatt buildout**; Fluidstack raises $830M at $7.5B valuation for AI data centers. **Implication:** Capital intensity is accelerating faster than grid capacity—70% of markets are overstretched, signaling imminent construction bottlenecks. - **Epoch AI: Google's Colossus 1 training compute hits 1e26 FLOP**; Gemini 4 pretraining begins as Google's most ambitious run yet. **Implication:** Frontier training runs are crossing the 1e26 FLOP threshold, requiring new cooling and power delivery architectures at hyperscale. - **KV cache offload makes storage the new AI bottleneck**; NUS CIMERA chip cuts LLM memory wall with compute-in-interconnect. **Implication:** Memory bandwidth is displacing raw compute as the primary constraint—CIM and phase-change memristor approaches gain urgency. - **China's AI ecosystem standardizes on MoE with wide expert parallelism**; Zhipu AI builds 1GW China-only data center, acquires compiler startup; Huawei Ascend SuperPOD decode throughput estimated 1.3-1.7x behind GB300. **Implication:** China is vertically integrating hardware-software stacks to close the gap with Nvidia, but still trails on inference latency for large models.