Compute & Infrastructure

Wood Mackenzie projects US data centre power-generation equipment spending will reach $65 billion by 2030, up from $2.6 billion last year — a tripling of the total US power equipment market driven almost entirely by AI workloads, according to Bloomberg. This is an analyst projection, not confirmed capital commitments, but the directional signal is consistent with what operators are already procuring.

Oracle's plan to deploy a 2.45GW fuel cell farm to power its New Mexico datacenter complex illustrates the operational reality behind those projections, per The Register. At that scale, Oracle would be building what amounts to a small utility. Fuel cells offer grid-independence and lower permitting friction than gas peakers, but at significant capital cost and with fuel supply chain dependencies of their own. The fact that a hyperscale-tier operator is pursuing this path confirms that grid interconnection queues — not land or hardware — are now the primary pacing item for large campus buildouts. Community-level resistance is compounding the problem: in Archbald, Pennsylvania, six proposed AI data centres covering 14% of a town of 7,000 people have prompted four of seven town council members to resign, per Tom's Hardware, signalling that local siting opposition is becoming a structural drag on buildout timelines.

Reports that OpenAI has missed internal targets for both active users and revenue sent NVIDIA, AMD, Oracle, CoreWeave, and SoftBank shares lower — SoftBank losing 9.9% on the Tokyo exchange alone, per Tom's Hardware. The market reaction reveals the structural concentration risk: a single company's revenue shortfall is sufficient to reprice the investment case for the entire hardware stack from semiconductors to cloud platforms.

The simultaneous restructuring of OpenAI's relationship with Microsoft — with Azure's exclusive cloud mandate now ended — adds a second-order complexity. OpenAI can now distribute workloads across multiple cloud service providers, per Tom's Hardware. This opens competitive procurement dynamics that could benefit Oracle, Google Cloud, and CoreWeave, but also means that capacity reservations made by Microsoft in anticipation of OpenAI workloads may now be partially stranded. For data centre operators, the question is whether demand absorption from other AI customers is sufficient to fill that gap.

China has announced the Lingshen supercomputer — reportedly the first exascale-class system to reach that performance tier using only CPU accelerators, packing 47,000 Huawei Kunpeng processors across 92 compute cabinets with no foreign-made components, per Tom's Hardware. If the claimed 2 exaflop figure is validated and refers to comparable double-precision performance rather than a mixed-precision metric, this represents a meaningful architectural divergence from GPU-centric Western HPC design.

The strategic significance is layered. First, it demonstrates that Huawei's domestic CPU program has matured to the point where it can serve as the foundation for a world-class system — a direct rebuke to the thesis that US export controls have decisively constrained Chinese compute capability. Second, a CPU-only architecture sidesteps the NVIDIA H100/H800 dependency entirely and suggests China may be pursuing a parallel compute paradigm optimised for workloads where CPUs remain competitive, including certain inference tasks. Third, the claim of zero foreign components, if accurate, represents genuine supply chain autarky at scale — a benchmark no Western nation has matched for comparable systems.

Victory Giant Technology reported a 28% year-on-year increase in Q1 PCB sales driven by AI server demand, per Bloomberg, while Bloomberg's broader supply chain coverage identifies multiple Asian component makers recording triple-digit growth as AI server assembly scales, per Bloomberg. TDK's CEO separately flagged AI-driven components demand as a material earnings driver in Bloomberg's Asia Trade coverage. These data points collectively confirm that AI infrastructure demand is now penetrating deep into second- and third-tier supply chain layers — passive components, PCBs, power management — not just leading-edge semiconductors.

This breadth of demand creates a different category of supply risk. Unlike TSMC advanced node capacity, which can be tracked through well-established reporting channels, PCB laminate, high-end connectors, and power conversion components are sourced from a fragmented supplier base concentrated in Taiwan, Japan, and mainland China. A disruption at this layer — whether from geopolitical escalation, natural disaster, or demand surges outpacing capacity expansion — would be harder to anticipate and slower to remedy than a chip shortage. The Wood Mackenzie power equipment projection reinforces this: transformers, switchgear, and UPS systems face their own multi-year lead times.

Tenstorrent announced general availability of its Galaxy Blackhole AI compute platform — a RISC-V-based system packing 32 Blackhole accelerators into a 6U chassis priced at $110,000, per The Register. At roughly $3,400 per accelerator slot in a fully integrated system, the price point is positioned to compete in the inference and edge-training segment where NVIDIA's GB200 NVL configurations remain out of reach for mid-tier operators. The RISC-V architecture is strategically significant: it removes ARM and x86 instruction set licensing dependencies and aligns with China's domestic compute roadmap, potentially opening export markets that NVIDIA cannot serve under current controls.

Concurrently, a research collaboration across Edinburgh, Peking University, Cambridge, and others published microarchitecture work on 3D-stacked near-memory processing optimised for LLM decoding, per Semiconductor Engineering. Near-memory processing directly addresses the memory bandwidth bottleneck that dominates inference latency costs. Combined with IEEE Spectrum's coverage of sparsity-aware hardware approaches to reduce the energy cost of large model inference, the technical trajectory points toward a generation of inference-optimised silicon that could materially shift cost-per-token economics within two to three years.