Info-Tech Research Group has published a blueprint, Define Your Target AI Infrastructure, arguing that enterprise AI bottlenecks come from workload misalignment rather than a shortage of compute. The advisory firm’s message is that organisations should stop buying hardware and start matching architecture to the work. That is a sensible framing, and the independent data behind it is far harsher than the release lets on.
What Info-Tech published
The blueprint is a five-phase methodology. It asks infrastructure leaders to assess workload characteristics, align processor strategy to those workloads, identify constraints across compute, memory, storage and networking, design balanced architectures, and set operational strategies for cost and risk. An accompanying workbook offers seven reference architecture patterns.
The core argument is that AI workloads are not interchangeable. Training, inference, retrieval-augmented generation, agentic AI and edge AI place different demands on the stack. Standard capacity planning fails, the firm says, because AI demand is variable and nonlinear.
John Donovan, the firm’s principal research director, puts it plainly. Successful strategies begin with the workload, not the technology.
The numbers the blueprint is arguing against
Info-Tech’s framework is a response to a measured problem. Cast AI’s 2026 Kubernetes report, drawn from telemetry across roughly 23,000 production clusters, found average GPU utilisation of about 5%. Enterprises provision roughly twenty times more GPU capacity than they consume. CPU utilisation averaged 8% and memory 20%.
Survey research points the same way. In a June 2026 poll of 573 technical leaders, VentureBeat Research found that 83% of enterprises running their own GPUs reported utilisation of 50% or less. Fewer than half could rigorously track what their compute costs. Only about one in five ran AI in production at scale.
Spending keeps climbing regardless. Gartner estimates AI infrastructure is adding about $401bn in new spending in 2026. At 5% utilisation, most of that capacity is a fixed cost that produces very little.
That is the gap Info-Tech’s blueprint names, and more gently than the data warrants.
Why the network is the hidden constraint
The most concrete point in the blueprint is about traffic. Traditional enterprise traffic flows north-south, from users to servers, and tolerates moderate bandwidth and higher latency. AI traffic is east-west, compute-to-compute, and demands both high bandwidth and low latency.
That shift is well documented. The scale-out network has become a first-order bottleneck as training clusters scale past 100,000 GPUs. Meta has described building one of the world’s largest fabrics of its kind to connect more than 100,000 GPUs across multiple buildings.
The consequences are measurable. In NVIDIA’s own comparison, standard Ethernet training step time inflates from 735 milliseconds to 1.18 seconds once noisy-neighbour traffic appears, while a congestion-isolated fabric holds steady. Under a worst-case test, some GPU pairs on standard Ethernet collapse to 25 Gbps.
This is where part of the utilisation gap lives. A cluster that looks idle may be waiting on data movement, not compute.
The AI infrastructure procurement trap
The blueprint frames the fix as architecture. The telemetry suggests the root cause is behaviour as much as design.
Cast AI’s analysis describes a self-reinforcing loop. Enterprises secure GPUs through waitlists, sign long commitments out of fear of losing allocation, then hold idle capacity because reacquiring it would take months. The cost of over-provisioning is invisible; the cost of under-provisioning pages an engineer. Teams rationally over-request.
Orchestration compounds it. Kubernetes assigns whole GPUs to a single workload by default, so a card can show low activity yet stay locked to a job that other work cannot use. The same dataset includes a fleet that reached 49% utilisation on 136 H200s through better scheduling alone.
That is the uncomfortable part of the argument. The gains on offer are mostly operational, not capital.
What to watch
Two signals. Whether enterprises adopt utilisation and unit cost as first-class metrics, since fewer than half currently measure them rigorously. And whether networking and orchestration get treated as part of the AI budget, rather than as infrastructure plumbing.
Info-Tech’s blueprint is a reasonable checklist. The harder question is whether buyers will audit the hardware they already own before signing for more.
Editor’s Note
This analysis draws on the October 2026 Info-Tech Research Group release and blueprint, Cast AI’s 2026 Kubernetes optimisation report, VentureBeat Research surveys, Gartner’s 2026 AI infrastructure spending estimate, published networking research from Meta and NVIDIA, and independent reporting on enterprise AI infrastructure. Where a figure or claim is a company’s own, we say so.