Compute & Infrastructure

Iran's Islamic Revolutionary Guard Corps published video on April 3rd threatening the complete destruction of OpenAI's planned $30 billion Stargate data center in Abu Dhabi, including satellite imagery of the 1GW facility, if the US attacks Iranian power infrastructure. The threat, reported by both Tom's Hardware and The Verge, represents the first publicly documented military threat against a named AI data center facility and demonstrates adversaries have mapped the geographic concentration of critical AI infrastructure.

The Stargate facility, announced as one of the largest AI data center projects globally, is being built in partnership with the UAE government. Its location in the Gulf region places it within range of Iranian missile systems, and the IRGC's willingness to publicly identify the facility suggests systematic intelligence collection on AI infrastructure targets. The threat comes during escalating tensions in the region following renewed conflict.

Samsung Electronics reported an eight-fold increase in quarterly profit driven by AI memory chip sales, far exceeding analyst estimates, according to Bloomberg. The results demonstrate sustained demand for high-bandwidth memory (HBM) chips used in AI training despite market volatility from Middle East conflict. Separately, Hon Hai Precision (Foxconn), a key Nvidia manufacturing partner, reported 29.7% quarterly sales growth, also meeting estimates during the conflict period, as reported by Bloomberg.

The earnings reports suggest that AI infrastructure buildout has continued largely uninterrupted through the initial phase of regional conflict, though this may reflect orders placed before hostilities began. IEEE Spectrum notes that HBM remains a critical bottleneck for AI inference speed, with hyperscaler demand continuing to outpace supply despite production increases. The memory shortage is structural rather than cyclical, driven by the fundamental architecture of large language models.

Nvidia revealed plans at GTC to use photonic interconnects to integrate over 1,000 GPUs into single systems by 2028, moving beyond the already-massive GB200 rack configurations, as reported by The Register. Industry analysis published by Semiconductor Engineering confirms that all AI data center interconnects will transition to optical within five years as electrical signaling reaches fundamental physical limits. The shift requires indium phosphide (InP) and silicon photonics (SiPho) to join CMOS as critical manufacturing technologies, with co-packaged optics (CPO) and optical circuit switching (OCS) becoming standard architectures.

The transition addresses a hard physical constraint: electrical interconnects cannot deliver sufficient bandwidth at the densities required for next-generation AI clusters without prohibitive power consumption and latency. This represents a major supply chain expansion beyond traditional semiconductor manufacturing, requiring different materials science, packaging techniques, and testing methodologies. MSI showcased desktop and rack systems at GTC incorporating Nvidia's GB300 architecture, indicating commercial systems are already incorporating optical interconnect technology.

Anthropic disclosed its revenue run rate has grown from $9 billion at year-end 2025 to over $30 billion currently, and confirmed plans to deploy 3.5GW of Google TPU chips manufactured by Broadcom, according to Bloomberg and The Register. The 3.5GW commitment represents one of the largest single compute deployments by an AI developer and marks Anthropic's strategic bet on Google's custom silicon rather than Nvidia's standard offerings. Broadcom is building both the AI accelerators and data center networking chips for Google under the arrangement.

The deal structure is notable because Anthropic, despite being partially owned by Google, is making an explicit commitment to non-Nvidia infrastructure at enormous scale. The 3.5GW figure refers to chip power consumption and would require substantially more total facility power when accounting for cooling and overhead. Separately, Nvidia-backed data center builder Firmus Technologies raised $505 million led by Coatue Management, as reported by Bloomberg, indicating continued capital availability for AI infrastructure despite market volatility.