AI Rack Supply Chain Revenue Climbs as Thermal Power and Rail-Kit Makers Post Record August

The three Taiwanese supply chain categories closest to the physical AI rack - thermal management, power supply and rail kits - booked combined August revenue of NT$121.48 billion, equal to roughly US$3.86 billion, according to monthly filings with the Market Observation Post System. DIGITIMES uses the aggregate as a tracker for how quickly hyperscale AI rack deployments are converting into physical hardware orders. Rack-level AI systems concentrate far more value in thermal management, power delivery and mechanical infrastructure than traditional servers do, which is why this cohort is a cleaner read on buildout pace than finished-system shipments.

Individual suppliers confirm the ramp. HEC Compucase's AI server-related rollout continues to gain traction as AI racks, cooling modules and high-end power supplies move into mass production or volume shipments, with fourth-quarter shipments and operating momentum expected to strengthen further. The company is targeting double-digit full-year 2026 revenue growth. Cooling is the clearest beneficiary, since higher rack power densities make liquid cooling a design requirement rather than an optional upgrade, and cold plates, manifolds and quick disconnects are shipping in higher volumes alongside the racks themselves.

Bottlenecks are shifting down the stack. Capacity for advanced semiconductor processes and packaging remains tight, and mature-node chips required for AI-related peripheral applications are also facing mounting supply pressure, a scramble DIGITIMES says may persist through 2028. That makes supply chain management itself a source of bargaining leverage across the AI hardware stack. Because a rack integrates compute, networking, storage, power and thermal subsystems, a delay in any single component class - including comparatively low-value interconnect and power parts - can hold back an entire deployment.

Memory connectivity has become one of the most active design fronts. Astera Labs expanded its Leo Smart Memory Controller family with the Leo X-Series and the Leo 2 E- and P-Series, targeting fabric-attached memory for agentic AI and rack-scale pooling. Paired with Scorpio fabric switches, Leo X-Series provides a dedicated near-GPU memory tier for KV-cache offload, delivering up to 62% faster time to first token and up to 22% more tokens per second. Leo 2 doubles memory bandwidth and capacity over the prior generation with four DDR4 or DDR5 controllers and CXL 3.2 support, and enables reliable reuse of already-deployed DDR4, a capability that matters directly in a memory supply constrained environment.

The commercial read-through is that value is migrating from the accelerator to the surrounding rack infrastructure. Thermal, power and mechanical suppliers are capturing a rising share of AI capital expenditure as densities climb, and their revenue is now reported monthly and trackable. With the expanded Leo family sampling at hyperscaler customers and OSAT and substrate vendors adding capacity, the near-term constraint is less about headline GPU availability and more about whether thermal, power, memory and interconnect supply can stay on the same delivery cadence as the accelerators themselves.