Compute & Infrastructure

Nvidia RTX 5090 cards are now trading at a minimum of $5,000 on retail and secondary markets — a dramatic premium above MSRP — and the pricing pressure extends across the entire Blackwell consumer stack according to Tom's Hardware. The driver is a combination of constrained TSMC CoWoS-L packaging capacity, sustained demand from both gaming and prosumer AI inference users, and tariff-driven cost inflation across the PC hardware supply chain.

The strategic read for infrastructure professionals is straightforward: if consumer Blackwell is this constrained, professional-grade H200 and B200 allocations remain the primary bottleneck for enterprise AI build-outs. The consumer pricing is a lagging indicator of the same CoWoS packaging chokepoint that is limiting data centre GPU supply. Until TSMC meaningfully expands advanced packaging capacity — a process measured in years, not quarters — allocation scarcity will continue to price-discover upward across both market segments.

AWS and Saudi Arabia's Humain have moved from announcement to commitment, confirming 50MW of capacity at the Saudi AI Zone by 2028 alongside a December 2026 cloud region launch that appears on track according to Data Center Dynamics. This is a confirmed investment with a near-term deliverable, distinguishing it from the broader wave of Gulf AI announcements that remain speculative. The 50MW figure is modest by hyperscaler standards but meaningful as anchor infrastructure for a sovereign AI zone designed to attract regional workloads away from European or US-jurisdiction cloud.

Separately, Microsoft and ChronoScale have announced a 50MW North American deployment based on Nvidia GB300 NVL72 systems — a next-generation Blackwell Ultra configuration — according to Data Center Dynamics. This is an announced plan, not confirmed capacity, and its execution is contingent on GB300 NVL72 availability, which remains subject to the same packaging constraints discussed above. The GB300 NVL72 rack-scale form factor also imposes significant power density requirements that will test facility cooling and power infrastructure.

AMD announced the Threadripper Halo Station at IFA 2026, pairing a 96-core Zen 5 Threadripper Pro with dual MI350P accelerators — expandable to four — and 2TB of DDR5 memory, with AMD claiming the configuration can run trillion-parameter models locally according to Tom's Hardware and ServeTheHome. The MI350P uses a liquid-cooled form factor, signalling that AMD is engineering around the thermal constraints that have historically limited workstation-class AI accelerator density.

This product targets a specific and growing segment: enterprises and research labs that need local, air-gapped, or latency-sensitive inference without committing to a full rack-scale deployment. The trillion-parameter claim deserves scrutiny — at current model precision levels, fitting a 1T parameter model in a dual-MI350P system would require aggressive quantisation — but the system's memory capacity and interconnect bandwidth represent a genuine step change for workstation-class hardware. AMD's MI300 series has already demonstrated competitive positioning against Nvidia's H100 in certain HPC workloads, and the Halo Station is a direct bid to extend that competition into the developer and enterprise on-premises market where Nvidia's DGX Station has dominated.

Infineon and Skeleton Technologies have announced a partnership combining Infineon's CoolSiC power semiconductors with Skeleton's power conversion systems, specifically targeting AI data centre power supply chains according to Data Center Dynamics. Silicon carbide power devices offer significantly higher switching efficiency and thermal tolerance than conventional silicon MOSFETs, which matters acutely in high-density GPU clusters where power conversion losses compound at scale. This is an announced partnership without confirmed deployment timelines, but it reflects a broader industry recognition that the power delivery stack — not just raw grid connection — is now a differentiated engineering problem.

Separately, a sponsored analysis from DDC published in Data Center Dynamics argues that rack containment architecture is becoming essential for AI-dense deployments. While the piece is commercially sponsored and should be weighted accordingly, the underlying engineering argument is sound: GPU rack power densities of 60-100kW per rack, as seen in NVL72 configurations, overwhelm traditional hot-aisle/cold-aisle cooling unless containment is implemented with precision. The combination of SiC power conversion efficiency gains and improved rack thermal management represents the near-term engineering envelope that data centre operators must solve before raw capacity expansion is possible.

Two technical developments from Semiconductor Engineering deserve attention from an infrastructure capacity perspective. First, imec has published advances in its CFET-based device roadmap — Complementary FET architecture represents the likely post-GAA transistor structure for sub-2nm nodes — detailing novel integration modules and standard cell configurations according to Semiconductor Engineering. CFETs stack N and P transistors vertically, enabling continued density scaling when lateral scaling runs out, but they introduce formidable process complexity. This is research-stage work, not production-ready technology, relevant to the 2028-2030 node generation.

Second, a technical piece on extending Intel's 18A process node with dual contact architecture demonstrates significant drive current and frequency improvements at matched capacitance according to Semiconductor Engineering. Intel 18A is scheduled for production in 2025-2026 and is a critical test of Intel Foundry Services' ability to compete with TSMC N2 for advanced AI accelerator designs. Performance extensions to 18A reduce the urgency of migration to the next node, potentially improving Intel's competitive window if the base process yields acceptably. Given that TSMC's packaging capacity — not process node — is the current primary constraint, any credible Intel Foundry capacity coming online relieves a meaningful supply chain chokepoint.