Connecting Idle GPUs to Reduce AI Infrastructure Costs, The Challenge of Distributed Cloud Startup AIEEV

-NAT Traversal connects idle PCs… Auto-scaling automatically allocates resources and lowers cost

-Provides automated operational infrastructure for stable model operation, performance optimization, and cost efficiency

-Completed pilot testing and will officially launch the service next year AIEEV plans to raise Series A funding in the first half of 2026.

The core of implementing AI services is training and inference. Training is the process in which an AI model learns and identifies patterns and correlations from large-scale data, while inference is the process where the trained model generates answers to user questions and is actually operated in service. To run AI services, massive datasets and high-performance computing are essential. Companies must deploy thousands to tens of thousands of GPUs, and operate data centers equipped with high-bandwidth networks optimized for GPU-to-GPU communication. A single GPU costs at least 15 million KRW and can exceed 20 million KRW. Such enormous infrastructure costs—especially purchasing and operating expensive GPUs—have become the main reason most AI service companies fail to secure profitability.

There is a company that lowers infrastructure costs for AI service providers. While high-spec GPUs are required to develop models, inference repeatedly performs preset tasks on a completed model, so lower-spec GPUs are sufficient. Noticing this, AIEEV built a distributed virtual cloud infrastructure by connecting idle PCs such as home computers and PC cafés, dramatically reducing GPU costs. AIEEV CEO Saejin Park explained, “Training a large language model requires thousands or tens of thousands of GPUs, but inference can be done with just a few. If we lower the cost of building AI infrastructure, more people can build and use more AI services.”

Park realized during the cryptocurrency mining boom that many GPUs were being used for crypto mining, while companies building AI services were suffering from a shortage of GPUs, and founded AIEEV in early 2024. Park holds a Ph.D. in system software from POSTECH and is currently an adjunct professor in the Department of Computer Science at Keimyung University. The AIEEV team consists of system software specialists from POSTECH, KAIST, and Korea University, most with more than 15 years of industry experience. Since its establishment, AIEEV has filed 10 patents, secured two commercial customers, and is conducting PoCs with 10 companies. In August, it was selected for SK Telecom’s “AI Startup Accelerator Batch 3.” AIEEV plans to pursue Series A fundraising in the first half of 2026 for commercialization. We met CEO Park at AIEEV’s Gangnam office to hear about how idle GPUs are connected, how distributed GPUs operate reliably, and AIEEV’s vision for reducing AI infrastructure costs.

So how does it connect idle GPUs? In general, personal PCs or PC café computers do not have public IP addresses, making external access impossible. So how can these PCs be connected? AIEEV connects them using NAT Traversal technology.

When a personal PC attempts to connect to the AIEEV server, the router automatically opens a tunnel, and AIEEV is able to control the GPU through it. If this method does not work, the system connects through an intermediate relay server. As a result of technical testing, AIEEV has completed a test connecting more than 100,000 nodes. The fact that a data center is unnecessary directly becomes a price advantage. In traditional data centers, 50% of electricity costs go into cooling. Because many devices are densely packed in a narrow space, extreme heat is generated. “Constructing, operating, and staffing a data center requires significant personnel. AIEEV does not need security, operational, or maintenance staff.”

Just because it is connected doesn't mean it's finished. Managing a large-scale network requires efficient communication and fault isolation between nodes. “To manage a large number of nodes, simple network connections are not enough. You must efficiently monitor the status of each node and quickly migrate tasks to another node if a failure occurs.”

Service stability is important in distributed systems. Operating AI services requires numerous open-source libraries, and if even one is missing or versions do not match, the service does not operate properly. To solve this, AIEEV developed container technology. By packaging all open-source libraries, system settings, and dependencies required to run AI models into a single unit, models can run in a stable, identical environment anywhere. Large-scale cloud or data centers host all services in a single location. If there is a problem with the data center network or power supply, the entire service may be interrupted. Therefore, redundancy must be configured so that the service does not stop even if a server fails, and automatic recovery systems must be built. AIEEV automated the entire process. Once a model is uploaded to the AIEEV platform, all AI service processes—from performance monitoring to redundancy configuration to automated recovery—operate automatically. Even if a node fails, the internal scheduler immediately switches the workload to another node, so the user logically does not detect the failure.

“Even if a node fails, the workload is immediately switched by the internal scheduler and executed on another node. From the user's perspective, there is a physical failure, but logically, the failure is not perceived.”

Another important function is auto-scaling. In typical cloud environments, servers must be prepared in advance according to expected traffic. The problem is that it is impossible to predict how many customers will come. So companies reserve resources conservatively. Even if the cost is high, a service outage is worse. If resources can be allocated efficiently based on usage, costs can be significantly reduced. AIEEV solves this with auto-scaling. When traffic increases, the necessary nodes and containers expand automatically, and when traffic decreases, they shrink automatically. Like adding more staff during peak hours and fewer during slow times, resources can automatically scale with traffic, reducing costs. The user can adjust sensitivity to control scaling speed.

“For one customer, this auto-scaling technology reduced costs by about 80% compared to a conventional cloud.” AIEEV provides automated operational infrastructure that ensures stable AI model operation, performance optimization, and cost efficiency. This means even a startup can immediately secure enterprise-level infrastructure stability.

AIEEV is conducting a pilot test in cooperation with a nationwide PC café management company. Once the pilot is complete, distributed idle GPUs from approximately 4,000 PC cafés nationwide will be available to supply computing resources to AI service companies. Collaboration discussions with enterprises are also active. In particular, demand is gradually increasing among companies that need small-scale AI resources, and they are expected to become AIEEV’s primary target customers.

AIEEV’s “Air Cloud” is a low-cost AI inference platform that can replace AWS. Depending on customer needs, it is provided in two options. Air Cloud Standard operates by crowdsourcing idle GPUs from personal PCs and PC cafés. Since it prioritizes affordability, startups and developers can test and deploy AI models at minimal cost. Air Cloud Plus is designed for enterprises requiring stability and high performance. It consists of a distributed cluster based on verified custom-build nodes and guarantees 99.99% availability, enabling stable operation of mission-critical AI services. Customers can dynamically allocate the necessary GPU capacity based on traffic and are billed only for the time used. This allows flexible response to unpredictable traffic fluctuations. Customers can check at any time whether their AI service operates properly through AIEEV's dashboard, which provides real-time monitoring of performance and detailed logs of API endpoints, and respond quickly if necessary.

“Air API” is a generative AI interface that developers can immediately integrate into their services. Without building a complex AI model, AI functions similar to ChatGPT can be implemented only by calling the API. Air API is charged per call, so companies only pay for what they use. Air API is a fully serverless AI solution. Customers do not need to worry about infrastructure or server management and can focus solely on business logic. When traffic increases it scales automatically, and when it decreases it automatically shrinks. Air API aims for official service launch in 2026. After launch, it will support various models based on Korean LLMs.

Private Air Cloud is a solution that centrally manages GPU resources scattered across large companies. As departments and teams secure GPUs as needed, company-wide efficient operation becomes difficult. Private Air Cloud fundamentally solves this problem. Just as Air Cloud aggregates distributed GPUs worldwide, Private Air Cloud aggregates distributed GPUs inside the enterprise and allocates them dynamically to the departments that need them.

AIEEV is named after the acronym of AI Equality, Everyone’s Value. It contains the meaning of AI equality and value for all. “I think the digital divide will continue to grow. The gap in access to information is invisible, but it will deepen. Even now, some people use ChatGPT for free, while others pay 300,000 KRW per month for ChatGPT Plus. The differences in their access to information are large.” Park’s concerns are becoming reality. As AI services expand across industries, the gap between companies with access to AI and those without is widening. “I don’t think this gap should exist. AI benefits should be available to anyone. AIEEV wants to create a world where everyone can use AI without burden — and for that to happen, infrastructure costs must come down.”

AIEEV is attempting not only to create a low-cost service, but to reduce societal inequality through technology. Infrastructure innovation is essential for this. If AIEEV’s challenge succeeds, the structure of AI usage will be completely different in a few years.

Source

VentureSquare | NOV 17, 2025 https://www.venturesquare.net/1014451