Anthropic, Pentagon Pressure, and Energy Constraints

Iran's drone strikes against AWS data centers in UAE and Bahrain mark the first documented kinetic attacks specifically targeting commercial AI cloud infrastructure based on its dual-use military value. The designation of AWS, Google, Microsoft, Nvidia, Oracle, and Palantir as legitimate military targets represents a doctrinal shift where concentrated cloud infrastructure—essential for AI training and inference—is treated as valid military assets during conflict. This targeting rationale appears based on these companies' Israeli business relationships and provision of cloud services to defense and intelligence customers, creating new strategic vulnerabilities that traditional cybersecurity frameworks do not address.

The economic logic of AI—which rewards massive concentrated facilities for training efficiency—directly conflicts with military logic favoring distributed resilient systems. Countries and companies making the largest AI capability investments are simultaneously creating the most attractive strategic targets. Neither commercial nor military institutions have adapted doctrine or architecture to resolve this trade-off, suggesting current AI infrastructure buildout could create brittle rather than robust strategic capabilities. For countries hosting major cloud facilities, this creates considerations about infrastructure vulnerability that extend far beyond the data sovereignty debates that have dominated policy discussions to date.

Asian bankers financing AI infrastructure are now explicitly incorporating energy cost volatility into deal terms following the Iran conflict, marking the first time power prices have become a material variable in data center financing rather than a predictable pass-through cost. This coincides with Bitcoin miners pivoting mining infrastructure to AI workloads as the network experienced its first quarterly hashrate drop since 2020, demonstrating how existing compute infrastructure is being reallocated toward higher-value AI inference when power costs make previous uses uneconomical. Rising energy prices driven by geopolitical conflict erode data center IRRs because power is a cost that cannot easily be repriced in long-term customer contracts already signed.

CoreWeave's $8.5 billion loan and Nebius's $10 billion Finnish data center commitment show infrastructure providers building capacity ahead of confirmed enterprise demand, betting that AI workload migration will justify the capital expenditure. However, if electricity costs remain elevated, leveraged infrastructure plays face margin compression or covenant breaches. This risk is amplified in Europe and Asia, where energy markets are less liquid and more exposed to geopolitical supply shocks than in the US. Power availability is replacing network latency and proximity to users as the primary constraint on data center location, fundamentally altering where new AI compute capacity can be deployed regardless of GPU availability or capital.

Shenzhen's activation of China's first 10,000-card intelligent computing cluster using Huawei Ascend 910C chips represents a milestone demonstrating that China can deploy large-scale AI infrastructure with indigenous chips, reducing the effectiveness of US export controls as a tool to limit China's AI development timeline. The facility's 11,000 petaflops capacity and deployment at scale signals Huawei has achieved sufficient yield and reliability to move beyond prototype demonstrations. This comes alongside Chinese GPU makers Biren and Iluvatar posting triple-digit revenue growth—207.2% and similar increases respectively—though both remain unprofitable, indicating China's domestic AI chip ecosystem is scaling commercially but remains pre-profitability and vulnerable to subsidy withdrawal.

South Korea's increased China manufacturing exposure complicates this picture: Samsung and SK Hynix increased investments in Chinese fabs by 67.5% to address global AI memory shortages, revealing that US export control architecture relies on allies whose commercial interests diverge from Washington's strategic objectives. Japan's Fujitsu announcement of plans for a domestically designed 1.4nm AI inference chip manufactured by Rapidus represents another sovereign compute initiative aimed at bypassing TSMC dependence. These parallel indigenous chip efforts—in China, Japan, and elsewhere—suggest that rather than containing AI capability development, export controls are accelerating fragmentation into regional supply chains with different performance profiles but growing independence from US-controlled manufacturing nodes.