Underutilized GPUs at 5%: Elice, Nota, Lablup offer the Korean answer to a global crunch

In plain terms

A CAST AI study published on July 31 and reported by ETNews estimates the average GPU usage to be 5% across approximately 23,000 Kubernetes clusters operated by the three major clouds (AWS, Azure, GCP) when Kubernetes is not optimized. Three Korean publishers - Elice, Nota, Lablup - respond on the same day with their product solutions. In other words: the GPU capex is already being paid for, usage is lagging worldwide; Korea is the first to publish publishers positioned on the fix.

The fact, in context

ETNews, July 31, 2026. The article aggregates two levels: a GLOBAL figure (the CAST AI study, 5% average usage measured on ~23,000 K8s clusters operated by AWS/Azure/GCP when Kubernetes is not optimized), and a LOCAL RESPONSE (three Korean publishers each offering an angle on the fix). Immediate context for Korea: the government completed a $950 billion package with the United States at the end of July, including a GPU capex component (Vera Rubin, a plan for 10,000 cards in the following fiscal year according to ETNews on July 19). The country is massively purchasing while a global figure reminds us that unorchestrated compute is wasted capex.

Under the hood

Three technical angles covered by Elice, Nota, and Lablup:

• Batching and scheduling - individual inference systematically underutilizes an A100/H200; dynamic batching (like vLLM) and multi-tenancy are needed.
• Model routing - much of the load does not need the large model; a router that switches to a smaller model or a quantized variant recovers 20-40%.
• KV cache management - long contexts explode memory; intelligent compression and persistence avoid underutilization of slots.

Elice (training/enterprise platforms), Nota (model compression), and Lablup (multi-tenant orchestration Backend.AI) are the three faces of the same orchestration equation.

So what

Two threads converge. Token-budget-caps - the CFO does not see the price of the token, they see the GPU bill. A GPU at 5% utilization is the same as a token at 20x the market price. APAC-AI-poc - Korea is not particular about the utilization rate, but it is the first to publish publishers positioned on the fix. The transition from POC to production is hindered by orchestration, not the model. To watch: adoption of usage benchmarks beyond academic MFU, publication of equivalent studies outside of US-cloud (Alibaba Cloud, Tencent, on-prem enterprise), and whether Korean hyperscalers (NHN Cloud, KT Cloud) integrate these native stacks.