Technology Health Check: October

by Brad Gastwirth Global Head of Research and Market Intelligence

Executive Summary

AI infrastructure demand remains robust but more strategically paced as hyperscalers shift from raw capacity expansion to utilization and energy efficiency. Power constraints continue to slow several U.S. and European data-center projects, even as capex commitments remain elevated.

Component markets remain tight. DDR4 and DDR5 pricing remains elevated, supported by constrained wafer output and sustained enterprise demand. HBM3E and substrate capacity remain structurally limited, while NAND prices appear firmer.

The enterprise HDD market continues to tighten, with 32 TB and larger drives facing extended lead times well into 2026. At the same time, CPU architecture and deployment strategies are evolving rapidly balancing GPU workloads, driving power density, and reshaping system designs.

Across compute, memory, and power electronics, the market remains defined by elevated pricing, multi-quarter lead times, and allocation-based supply, requiring increasingly proactive procurement planning.

5 Key Signals to Watch

GPU Allocation Discipline: NVIDIA’s H200 and Blackwell shipments remain tightly controlled; channel supply limited.

HBM3E Supply Constraint: TSV and packaging capacity remain bottlenecks; relief unlikely before mid-2026.

Memory Pricing Elevated: DDR5 remains firm; DDR4 continues to hold higher levels amid legacy-node supply cuts.

CPU Power Density Rising: AI-optimized CPUs (AMD Turin, Intel Granite Rapids) push rack-level design limits and sustain tight supply.

MLCC & HDD Shortages: High-capacitance MLCC lead times now exceed 40 weeks; HDD supply tight into next year.

Enterprise Compute & AI Infrastructure

Hyperscaler buildouts remain steady but increasingly phased around energy availability and cooling capacity.

Server architecture continues to migrate toward 80–120 kW racks, blending liquid and immersion cooling.

In Asia, Alibaba and Baidu are accelerating use of in-house accelerators, marking a clear pivot toward local hardware ecosystems.

Compute Trends: CPU Side of AI Infrastructure

While GPUs dominate AI acceleration, CPUs are evolving to manage orchestration, data handling, and inference workloads that complement GPU clusters.

Design Shift: New architectures from AMD (Turin / Zen 5) and Intel (Granite Rapids, Sierra Forest) focus on bandwidth, I/O, and energy efficiency. Hyperscalers are deploying mixed compute racks, pairing dense-core CPUs with GPUs for preprocessing, inference, and coordination tasks.

Supply & Pricing: CPU supply remains balanced, but high-end SKUs are tight due to shared advanced-node capacity. Lead times average 10–16 weeks; pricing remains firm on steady enterprise refresh cycles. Power envelopes exceeding 400 W are forcing new rack-level PDU and cooling requirements.

Competitive Landscape: AMD continues to gain hyperscaler share with dense-core EPYC CPUs. Intel’s Granite Rapids and Sierra Forest target efficiency and throughput leadership. Arm-based CPUs (AWS Graviton, NVIDIA Grace, AmpereOne) are gradually gaining ground in inference and cloud-native environments.

Integration Trend: Compute architecture is shifting toward CPU–GPU convergence, using high-speed interconnects like NVLink-C2C and Infinity Fabric. CPUs increasingly manage orchestration and inference layers, while GPUs focus on model training.

Implications: CPU demand now intersects with GPU-related bottlenecks: substrates, MLCCs, and power delivery. Expect modest lead-time pressure and pricing resilience into 1H26 as AI-driven refresh cycles persist.

Memory & Storage Trends

DRAM: DDR5 pricing remains elevated on sustained AI-server demand. DDR4 also remains tight, with limited legacy output keeping pricing firm.

HBM: CoWoS and interposer capacity remain fully allocated through mid-2026; output prioritization favors GPU vendors.

NAND: Prices are firmer as hyperscalers restock high-density SSDs under supplier discipline.

Enterprise HDD: High-capacity drives (≥28 TB) remain constrained; 32 TB lead times extend into late 2026.

Regional Component Dynamics

Taiwan: TSMC’s CoWoS and InFO packaging remain fully utilized; Unimicron, Kinsus, and Nan Ya PCB remain backlogged through mid-2025.

Korea: SK Hynix leads HBM3E output; Samsung balancing DDR5/LPDDR5 production while investing in substrate expansions.

Japan: Murata and TDK now quoting 40–46 week lead times for large-form MLCCs; inductors also tight.

China: Alibaba’s T-Head chips in early deployment; SMIC and Cambricon expanding capacity under state support.

Emerging Bottlenecks:

Power transformers, PDUs, interposer substrates, and high-cap MLCCs continue to gate data-center deployment timelines. Freight and packaging constraints in Taiwan and Korea remain persistent.

Industry Headlines to Watch

Alibaba Qwen3-Max (>1T parameters): Expands China’s domestic AI infrastructure.

Alibaba 7 nm inference chip testing: Early substitution progress for domestic accelerators.

TSMC CoWoS capacity expansion: Incremental relief expected by late 2026.

AMD Turin & Intel Granite Rapids: New server platforms entering OEM validation.

Power & Cooling Infrastructure: Transformer and thermal-system delays slowing hyperscaler rollouts.

Call to Action: Procurement Guidance

Given persistent tightness across compute, memory, and power systems:

Extend Forecasts: Secure visibility two quarters out for CPU, GPU, and DRAM categories.

Negotiate Allocations Early: Lock in framework agreements to manage supply volatility.

Diversify Sourcing: Incorporate certified refurbished systems for inference and mid-tier compute.

Prioritize Power Components: MLCCs and transformers now sit on the critical path.

Monitor CPU–GPU Convergence: Track AMD, Intel, and Arm design cycles for early signals on server refresh timing.

Outlook: The global compute ecosystem remains supply-constrained entering 2026. HBM and substrate bottlenecks continue to cap GPU availability, while CPUs and power components emerge as new gating factors. With pricing elevated and lead times extended, procurement agility and multi-quarter visibility remain essential through the next investment cycle.

by Brad Gastwirth Global Head of Research and Market Intelligence