The latest in data center infrastructure – from power and minerals to networking, CPUs, and rack-scale AI systems – shows the stack shifting toward energy, interconnects, and end-to-end efficiency.
August 18, 2026
Data Center Knowledge’s latest hardware coverage points to a system-level rethink of AI infrastructure. Beyond GPUs, the biggest stories centered on energy, mineral supply, grid interconnects, rack-scale integration, and an oncoming network supercycle, signaling an infrastructure-first phase of AI buildouts.
Here are the most-read data center hardware stories from July through mid-August:
‘GPUs Suck’: Former Intel CEO Slams Data Center Hardware Limitations At the Ai4 conference, former Intel CEO Pat Gelsinger argues that the current AI hardware stack, especially GPUs paired with HBM, is power-hungry and computationally inefficient. Commentary from OpenAI leadership also highlighted constraints across manufacturing, energy, and deployment that demand a system-wide redesign.
The Critical Minerals Crisis: AI Data Centers Face Supply Chain Strain Critical minerals, especially copper for power distribution, are emerging as a chokepoint as AI demand surges, ore grades decline, refining capacity lags, and geopolitical risks grow. While hyperscalers’ willingness to pay can secure short-term needs, it tightens the market overall. Long-term offtakes, recycling, and diversified or domestic refining are becoming essential to future-proof builds.
Nvidia’s $500B Infrastructure Bet Raises Power Stakes Nvidia’s financing initiative lowers the cost of deploying GPU hardware and can push more GPU spend off operators’ balance sheets. However, it does not resolve core bottlenecks: grid interconnects, transformers, turbines, and permitting. Scarcity is shifting toward energized capacity, increasing the value of power-ready sites and existing electrical infrastructure.
Pat Gelsinger criticized current GPU efficiency during his keynote at the Ai4 2026 conference in Las Vegas. (Photo: Shane Snider)
TSMC Expands Arizona Campus to $265B as AI Demand Surges TSMC is scaling up advanced chip supply in Arizona, adding multiple fabs and packaging capacity, and ramping up 2 nm (alongside 3 nm and 5 nm). The expansion targets GPUs, CPUs, networking silicon, and custom AI accelerators.
AMD Fires Back at Nvidia with Helios AI System, Epyc CPUs AMD introduced Helios, an integrated rack-scale AI system combining sixth-generation Epyc 9006 CPUs with new Instinct MI455X GPUs, Pensando networking, and ROCm into a single-vendor platform designed to rival Nvidia’s Vera Rubin/NVL72. AMD claims Helios delivers higher AI compute density, more memory and scale-out bandwidth, and improved tokens-per-dollar.
AMD’s Helios is a fully integrated rack-scale AI system featuring Epyc processors. (Image: AMD)
Enter the Network Supercycle: Preparing Data Center Networks for AI’s Next Wave AI clusters increasingly require multi-layer fabrics (scale-up, scale-out, and scale-across), providing 10-100 times more bandwidth, ultra-low latency, and robust congestion control to keep thousands of GPUs fully utilized. As agentic AI and always-on inference expand east-west traffic, the network becomes part of the computer. Expect major upgrades in switches, optics, topology, telemetry, and resilience.
IBM Warns AI Infrastructure Shift Impacts Software Deals, Shares Plunge IBM’s stock declined after it warned that Q2 revenue and EPS would miss expectations, citing customers shifting spending to hardware (servers, storage, memory) to lock in constrained supply and pricing, delaying software and consulting deals. Analysts characterized the shortfall as a market-sequencing issue and flagged execution risks in converting IBM’s AI story into software-and-services wins.
IBM Brings Z to 19-Inch Racks as AI Reshapes Data Centers IBM is introducing rack-mount and compact single-frame versions of its z17 and LinuxONE systems, enabling enterprises to slot mainframe components into standard 19-inch racks or deploy turnkey appliances. The form-factor shift tightens the footprint and eases cooling integration alongside existing infrastructure.
Quantum Meets the Data Center: Hybrid Systems Take Off Quantum hardware is pivoting from qubit-count races to tightly integrated, hybrid deployments in which QPUs are co-located and low-latency-linked with GPU/CPU systems for workload-level speedups. In practice, data centers must support cryogenics, power and heat density, and vibration isolation, while maintaining modular layouts to accommodate rapidly evolving modalities.
A Guide to Network Redundancy in Data Centers Resilient networking remains foundational. True redundancy requires diverse ISPs and entry paths, duplicate switches and routers, and automated failover. Bonded networking can enable near-instant continuity when components fail.
The Big Picture #
The hardware stories above depict an AI ecosystem shifting from a GPU-centric lens to an infrastructure reality defined by energy, minerals, and interconnects. Financing can accelerate purchases, but energized capacity and copper availability increasingly set the pace of deployment. Meanwhile, rack-scale systems, advanced packaging, and the network supercycle are becoming decisive levers for utilization. The competitive race now spans accelerators, power, supply chains, and end-to-end system efficiency.