Global – November 10, 2026 -- Roughly 9,000 nodes of current-generation NVIDIA hardware cost about $15.6 million in GPU hours for every 30 days they sit idle, according to a calculation CIQ presented with a customer whose hardware arrived before networking, cooling and software were ready. The figure is the presenter's own estimate, offered at AI Infrastructure Field Day to argue against buying a complete, single-vendor "AI factory."
CIQ says every AI stack layer will change within a year
The presenter said layers are changing at once, some on one-month to three-month cycles, while hardware is revised within a year. Teams are simultaneously asked to move from research to production in about three months.
That pressure pushes buyers toward one stack from one vendor. CIQ's objection is that an AI factory reflects one vendor's view of AI infrastructure at a single point in time, so a change in one layer can leave the rest built around outdated parts within three to 12 months.
CIQ proposes independent layers instead of a fixed stack
Under the "kit" approach, buyers add only the layers they lack and run them next to existing infrastructure, in-house components and third-party products. Each product must be useful on its own, integrate with what the customer runs and be replaceable. The presenter gave this answer when a Field Day delegate asked whether one stack could cover everything.
Mixed hardware and hybrid deployment drive the argument
AI workloads now run on CPUs, GPUs, DPUs on network cards, NPUs, TPUs, FPGAs, ASICs and accelerators from multiple vendors. They also span on-premises, cloud and virtual private cloud environments.
The presenter described an "AI hangover": teams moved to the cloud for lack of GPUs, received the bill, bought their own GPUs, and must now run the same workloads on-premises as well. Workloads also differ, with some prioritizing stability and others maximum node performance.
CIQ's answer is a common Enterprise Linux foundation, with versions tuned per workload. A delegate noted that operators want to manage everything from one place; the presenter agreed with the goal but rejected a single system containing every package, citing extra software on every node.
CIQ's kit starts at the operating system and adds three layers
The presenter said most AI stack diagrams omit the operating system, though the kernel still configures devices and controls how data streams in from storage. Unified-memory architectures reduce that work without removing the kernel from the path.
- Operating system: Rocky Linux from CIQ Pro (RLC Pro) is the supported base. RLC Pro AI is tuned for GPU workloads, and RLC Pro Hardened targets regulated environments.
- Kubernetes: CIQ Stratum Kubernetes Engine (SKE) is an open source engine, and CIQ Stratum Pro adds fleet management across SKE clusters.
- Workflow orchestration: Fuzzball runs HPC and AI workflows across on-premises clusters and cloud.
Buyers can take one product and skip the rest. The presenter's advice is to start with the layer causing the largest loss of time or GPU hours rather than replacing the whole stack at once.