Test report DSG-5706 · Rev C · tested November 17, 2026

AI Datacenter InfrastructureDevice under test

Clush lifts GPU utilization 24 points in a month via Acceleon

Clush lifted one government project's GPU utilization from 17% to 41% in a month, expects revenue above 50 billion won and plans a KOSDAQ listing review this month.

Read
5 min
Words
932
Node
7nm
Operator
Grace Kim

Spec summary

  1. GPU utilization on one government-funded project rose from 17% to 41% in a single month, while another client's task success rate climbed from 71% to 94% over a year
  2. Clush operated more than 7,000 of South Korea's estimated 50,000 to 60,000 GPUs last year and manages roughly 1,500 Nvidia H200 GPUs at once
  3. Revenue reached 37.5 billion won last year, up from 15.6 billion won in 2024, with more than 50 billion won expected this year
  4. The national AI computing center project carries a 2.4065 trillion won ($1.65 billion) budget for close to 50,000 GPUs
  5. Clush received A and BBB grades in its technology evaluation last month and plans to file for a preliminary KOSDAQ listing review this month
Korean Startup Clush Lifts GPU Utilization 24 Points in a Month
Fig. AKorean Startup Clush Lifts GPU Utilization 24 Points in a Month — blickpixel / pixabay

Grace Kim

Clush, a GPU operations startup founded in 2015, raised GPU utilization on one government-funded project from 17% to 41% in a single month, a 24-percentage-point gain that the company presents as proof of a broader problem: enterprises keep buying accelerators faster than they can keep them working.

Revenue reached 37.5 billion won last year, up from 15.6 billion won in 2024, and Clush expects revenue above 50 billion won this year with operating profit rising sharply. The company posted its first annual operating profit since its founding last year.

Clush operates within South Korea's government-led national AI computing center program, whose construction is under way at the Sanigi Peninsula in Haenam County, Jeonnam-Gwangju Special Metropolitan City, on the Yellow Sea. The project carries a 2.4065 trillion won ($1.65 billion) budget and will house close to 50,000 graphics processing units (GPUs).

Samsung SDS (018260), Naver Cloud and Samsung Electronics (005930) sit among the participants; Clush is the only small or mid-sized firm in the group. In an interview on the 11th, CEO Lim Jung-soon and CTO Lee Yong-hyuk estimated that South Korea had 50,000 to 60,000 GPUs installed last year, with more than 7,000 operated through Clush.

Clush typically runs several hundred to more than 1,000 GPUs at a time. A memory chipmaker, Korea's largest financial group and a homegrown cloud service provider have kept contracts with the company running for years.

What does Acceleon manage?

Clush's flagship product, Acceleon, is a management program for heterogeneous computing resources. It monitors and allocates work across GPUs, central processing units (CPUs) and neural processing units (NPUs) from a single platform, spanning data center operations and the development and management of AI agents for corporate use.

The platform runs four operating systems:

  • GPU-as-a-service: assigns and reclaims tasks across several types of AI computing resources.
  • Data: collects and uses a company's internal data and controls access permissions.
  • Machine learning: links development, training and deployment of AI models, including turning a trained model into a live service.
  • AI agents: manages deployment and usage records of agents used inside corporate work systems.

On one government project, Clush managed roughly 1,500 Nvidia H200 GPUs at once through Acceleon. "What sets Acceleon apart is that it is built specifically for running large-scale AI infrastructure," Lim said.

The deployment translates into two operating metrics. A government-funded project moved from 17% GPU utilization to 41% within a month, and a corporate client lifted its GPU task success rate from 71% to 94% over a year. GPU utilization measures the share of GPUs with assigned tasks; task success rate measures the share of those tasks completed successfully.

Clush credits the software only in part. Lee said GPUs are sensitive devices prone to frequent errors, so raising utilization and task success together needs skills beyond ordinary server management. Clush diagnoses and troubleshoots alongside clients and takes responsibility for operational services, so customers can use their GPUs as efficiently as possible.

Lim tied both metrics to capital efficiency. "GPU utilization and task success rates are directly tied to a company's return on investment," she said, adding that companies across the market keep buying high-performance GPUs without using their performance fully.

What does this change for GPU buyers?

It changes the question buyers should ask after installation. Hardware spend does not tell you how much compute a fleet delivers, and Clush's own figures show assigned capacity moving from less than one-fifth to more than two-fifths in a month, with task completion rising from 71% to 94% in a year on a separate account.

The software layer, not the silicon, drives that shift, which is why Clush is extending the platform rather than reselling servers. A significant Acceleon feature update arrives next year, with broader use of predictive scheduling.

That scheduling technology forecasts when running tasks will finish and reserves GPUs needed for large-scale jobs in advance. Clush is also building AI agents for industry-specific work, such as manufacturing processes and support for financial underwriting reviews.

How does Clush sell without a large sales force?

With depth rather than headcount. Of roughly 220 employees, only five work in sales. Lim said the company's motto is "land and expand" — winning one Clush technology with a corporate client, then adding other services. Satisfied clients create a technology lock-in effect that expands the business, so Clush serves key accounts rather than pushing sales hard.

The sequence matters: Clush targets leading companies in each industry first, then widens the solutions it supplies them. Many small and mid-sized IT firms chase a broader footprint through short-term projects; Clush instead builds around industry-specific knowledge accumulated with large clients.

The listing process has started. Clush received an A grade and a BBB grade in the technology evaluation for a technology-based special listing last month, meeting the requirements, and plans to file for a preliminary listing review with the Korea Exchange this month.

Overseas, the company is in talks to supply Acceleon to neocloud operators abroad, including in Southeast Asia. Lim called domestic and overseas AI data center construction a clear tailwind and said Clush aims to become the AI infrastructure software company that represents Korea.

The near-term test is whether a five-person sales function can carry revenue past 50 billion won this year, with operating profit rising sharply, while predictive scheduling and Southeast Asian neocloud deals move from roadmap to contract.

via en.sedaily.com (Source)

Filed under

  • gpu-utilization
  • ai-infrastructure-software
  • heterogeneous-computing
  • predictive-scheduling
  • acceleon
Share this article:

More from Grace Kim

Grace Kim

Show full bio

Market editor covering regulation, policy and their impact on hardware layer of artificial intelligence at Die Signal.

665 articles

Same lot · LOT-362D

« Previous articleNext article »