Compute & Infrastructure

SK Hynix's $26.5 billion US ADR offering — the largest first-time US share sale ever by a foreign company — is strategically timed to capitalise on investor demand for direct AI infrastructure exposure. Memory, specifically High Bandwidth Memory, is the single most constrained component in current AI accelerator supply chains, and SK Hynix holds the dominant position in HBM3e supply to NVIDIA. The ADR gives the company a permanent US fundraising channel that will likely be tapped to fund the Indiana fabrication facility it announced in partnership with Purdue University, though that facility remains in early construction phases with volume output years away. Bloomberg

Simultaneously, Commerce Secretary Lutnick is applying direct bilateral pressure on both SK Hynix and Samsung to accelerate US-based memory output, framing the ask explicitly around addressing a global shortage of components critical to AI development. Bloomberg The political and financial incentives are now aligned in the same direction, but the structural reality is that memory fabs — particularly for advanced DRAM — require three to four year build cycles and depend on the same constrained equipment supply chain (ASML EUV, Tokyo Electron deposition) as logic fabs. US-produced HBM at scale is a 2029-2030 story at the earliest, meaning current AI buildout will remain dependent on Korean production for the foreseeable planning horizon.

Gartner's projection that global data centre electricity consumption will reach 565 TWh in 2026 — up 26% from 447 TWh in 2025 — and that AI servers alone will exceed all conventional data centre hardware consumption by 2027 represents a demand curve that most utility and grid operators did not model even 18 months ago. Tom's Hardware The distinction between training clusters and inference infrastructure matters here: training remains concentrated at a small number of hyperscale campuses where operators can negotiate directly with utilities for dedicated grid connections or co-located generation. Inference, by contrast, is distributing rapidly across edge and regional deployments, creating a more diffuse and harder-to-plan load profile.

The cooling dimension compounds the power challenge. The shift toward liquid cooling — including two-phase immersion systems — is accelerating as GPU thermal density makes air cooling architecturally inadequate for next-generation racks. Universal liquid cooling is transitioning from an optional premium to a baseline specification for new AI-optimised builds. Data Centre Dynamics This has supply chain implications: the liquid cooling component ecosystem — pumps, cold plates, manifolds, dielectric fluids — is not scaled for the volume ramp now being demanded, and lead times are extending.

FTC clearance for Elon Musk's acquisition of Mesh Optical is significant less for the company's current scale than for what it completes in Musk's AI infrastructure stack. With Terafab providing custom silicon design capability and Gigasat handling satellite-based bandwidth, Mesh Optical inserts the optical interconnect layer that determines how efficiently compute nodes communicate at cluster scale. Tom's Hardware The strategic logic mirrors NVIDIA's acquisition of Mellanox in 2020: controlling the interconnect layer allows the stack owner to optimise end-to-end latency and bandwidth in ways that pure chip performance benchmarks obscure.

The broader signal here is that interconnects — both within the data centre (NVLink, InfiniBand, UALink) and between facilities — are becoming a primary axis of competition. As GPU cluster sizes scale toward 100,000-accelerator configurations, the fabric connecting those accelerators increasingly determines effective training throughput. Vertically integrated players who control silicon, interconnect, and networking software can tune the full path; those dependent on third-party networking components face a combinatorial integration challenge that grows with cluster size.

Two distinct European compute sovereignty plays advanced this week. The Aether consortium announced a target of two AI gigafactories in Strasbourg, France, with an initial 42MW capacity — a modest but concrete step toward the EU's stated goal of hosting 20% of global AI compute capacity. The Strasbourg location is notable for access to Rhine hydropower and cross-border grid connectivity with Germany, partially addressing the power availability constraint. Data Centre Dynamics Separately, Dutch infrastructure operator NorthC is partnering with Volt to launch a Netherlands-based AI cloud service, with a potential future tie to Volt's planned AI gigafactory. Data Centre Dynamics Both remain announced plans rather than confirmed capacity — the Aether gigafactories in particular face the standard 24-36 month gap between announcement and operational status.

Canada is also advancing regional inference capacity, with DeepInfra deploying its first non-US data centre in Toronto — a modest but operationally confirmed step that reflects both Canadian regulatory comfort for US-adjacent workloads and the availability of relatively lower-cost grid power in Ontario. Data Centre Dynamics The pattern across all three announcements is that mid-tier operators are filling the gap between hyperscale campuses and local edge deployments, targeting inference rather than training workloads where power and latency requirements are more manageable.

Italy's Technoprobe, a manufacturer of probe cards used in semiconductor wafer testing — a component directly in the NVIDIA supply chain — has surged 330% as investors seek exposure to AI infrastructure picks-and-shovels plays beyond the primary chipmakers. Bloomberg This matters from a supply chain vulnerability perspective: probe card capacity is a non-trivial bottleneck in chip production ramp cycles, and concentration of that capacity in a single European supplier creates a fragility that is currently invisible because demand is being met — but which would become acute if Technoprobe faced operational disruption or if NVIDIA's production volumes require faster scaling than Technoprobe's capacity expansion plans allow.

China's Nexchip raised $891 million in Hong Kong and debuted up 14%, demonstrating that the Chinese domestic semiconductor capital market remains functional and active despite US export controls constraining access to leading-edge equipment. Bloomberg Nexchip focuses on mature-node display driver and power management ICs rather than leading-edge logic, meaning its ramp does not directly compete with TSMC at 3nm — but the capital formation success signals that Chinese chipmakers at all nodes are accessing growth funding through Hong Kong markets as US capital markets remain largely closed to them.