__26J01qhO10.png)
INFINITIXixCSP / AI-Stack
White Paper / 2026
From Infrastructure to
Token Factory
Key gaps and solutions for operators building their own AI data centers
5%
Average enterprise GPU utilization, measured across roughly 23,000 Kubernetes clusters.
Cast AI, 202686%
of enterprises running their own GPU infrastructure report utilization below 50%.
VentureBeat Research, 2026Chapter 1
The unit of output changed. The budget structure has not.
A data center used to be valued by rack count and PUE; AI workloads replaced the unit of output with the token. Self-build operators are mature at the facility and hardware layer, yet the layer that turns compute into a billable service usually never enters the budget.
The gap between cost and revenue
Chapter 2
Four key gaps, all in the software layer
Low compute utilization
Heterogeneous resources go unmanaged, the lack of virtualization lets a single job monopolize an entire accelerator, and teams hoard reserved capacity.
Insufficient multi-tenancy and governance
No tenant isolation across compute, storage and network, and no resource quotas, access control, audit trails or compliance reporting.
No usage metering or billing
Token metering is an order of magnitude harder than GPU-hours: separate input and output pricing, per-model rates, quotas and overage, reconciliation and audit.
Customer onboarding and time to launch
Without standard APIs and a gateway, a subscribable service catalog and a customer self-service portal, hardware depreciates first and revenue arrives later.
Chapter 3
A three-layer reference architecture, and how the products divide the work
The three layers must not only be in place, they must connect: accurate billing has to identify which tenant, which model and which physical device produced every token.
AI cloud operations layer
Multi-tenancy · Metering · Billing reconciliation · Service catalog · Self-service portal
AI platform layer
Model deployment · Inference serving · Version control · Model catalog
Compute infrastructure layer
Heterogeneous resource management · Virtualization · Scheduling · Monitoring and operations
With only the infrastructure layer, you are a colocation provider; with the first two, a GaaS provider; with all three, a Token Factory. AI-Stack and ixCSP can be adopted in stages as a project progresses, and neither requires a new facility.
Free download
Get the full white paper
From Infrastructure to Token Factory: Key Gaps and Solutions for AI Data Center Operators, with full source citations.
