Compute & Infrastructure

Half of planned US data centre construction projects have been delayed or cancelled, according to Tom's Hardware, as power infrastructure components become the binding constraint on AI capacity expansion. Cloud giants plan to deploy $650 billion in AI infrastructure in 2026, but availability of transformers, switchgear, and other electrical distribution equipment sourced primarily from China has created a bottleneck that software optimisation cannot solve. The delays reveal a fundamental mismatch between the speed of capital allocation and the multi-year lead times required to manufacture and deploy grid-scale electrical infrastructure.

This constraint is structural rather than temporary. Electrical equipment manufacturers operate on different timelines than cloud providers, and domestic alternatives to Chinese suppliers remain limited. The delays will force hyperscalers to prioritise existing facilities for expansion and potentially shift more workload to regions with available grid capacity, creating geographic imbalances in AI compute distribution.

Memory will account for 30% of hyperscaler AI data centre spending in calendar year 2026, up from roughly 7.5% in 2023, according to SemiAnalysis estimates reported by Tom's Hardware. The quadrupling of memory's share reflects both the explosive growth in HBM demand for AI accelerators and Nvidia's ability to negotiate preferential pricing well below standard market rates. This pricing advantage compounds Nvidia's architectural lead: customers not only get superior performance but also better total cost of ownership through supply chain leverage competitors cannot match.

The memory concentration creates a two-tier market. Nvidia's preferred access to HBM capacity at below-market rates means alternative accelerator designs face higher effective costs even when claiming comparable silicon performance. Memory suppliers prioritise Nvidia volume, reinforcing a cycle where competing architectures struggle to achieve cost parity at deployment scale.

Arm-based processors will power 90% of AI servers using custom silicon by 2029, according to industry analysis cited by Tom's Hardware. Hyperscalers building in-house CPU designs are converging on Arm architectures for efficiency and control, marginalising both x86 and RISC-V in the AI data centre segment. Amazon's Graviton, Google's Axion, and Microsoft's Cobalt all use Arm instruction sets, establishing a de facto standard for cloud-native AI workloads. The trend reflects a strategic shift where hyperscalers prioritise workload-specific optimisation over general-purpose compatibility.

Intel's data centre chief has publicly disputed claims that agentic AI requires new CPU architectures, as noted by The Register, arguing existing cores meet agent requirements. This disagreement highlights the architectural debate, but deployment decisions by AWS, Google, and Microsoft carry more weight than vendor positioning. The Arm shift in custom silicon is already underway, driven by total cost of ownership rather than theoretical performance debates.

Iran's Islamic Revolutionary Guard Corps claimed responsibility for attacks on an Oracle data centre in Dubai and an Amazon facility in Bahrain, according to Tom's Hardware. The country has explicitly threatened attacks against Nvidia, Intel, and other technology companies operating in the region. Whether the claims are accurate or inflated for propaganda purposes, the targeting of data centre infrastructure represents an evolution in state-level conflict beyond traditional cyber operations.

Physical attacks on data centres create a new risk category for infrastructure planning. The Middle East has emerged as a significant AI compute hub due to available power and proximity to markets, but this geographic diversification now carries kinetic risk. Separately, Super Micro Computer co-founder Yih-Shyan Liaw pleaded not guilty to charges of smuggling billions of dollars in Nvidia servers to China, as reported by Tom's Hardware, illustrating continued pressure on AI hardware export controls.

Nvidia's $2 billion investment in Marvell, reported by The Next Platform, extends beyond NVLink Fusion switches to encompass broader infrastructure control. Marvell supplies custom silicon for hyperscaler networking and storage, making this partnership strategic for maintaining Nvidia's end-to-end influence over AI data centre architecture. The deal ensures NVLink compatibility across multiple infrastructure layers while potentially limiting Marvell's ability to support competing interconnect standards with equal priority.

This follows Nvidia's pattern of using financial relationships and technology licensing to shape the ecosystem around its architectures. By investing in infrastructure component suppliers, Nvidia reduces the likelihood of those companies prioritising alternative interconnect standards or competitive accelerator designs.