A growing number of artificial intelligence startups are subscribing to GPU-as-a-Service (GPUaaS), a cloud-based subscription model that lets companies rent graphics processing units on demand. The trend spans startups seeking to reduce time and cost burdens through more efficient GPU utilization to those choosing the service for high performance and stability. GPUaaS is rapidly establishing itself as essential infrastructure. The shift is accelerating as companies that want to purchase GPUs outright continue to face not only high costs but also persistent supply shortages, according to industry analysts.
Soombit AI, a medical AI startup, began using GPUaaS from Elice Group in the second half of last year, according to industry sources on Tuesday. Soombit AI develops software that uses AI to generate draft radiology reports for chest X-rays. As medical screenings increase worldwide, the volume of imaging tests — including X-rays, computed tomography (CT) scans and magnetic resonance imaging (MRI) — is also rising. However, the number of specialists available to read these images has not kept pace. Soombit AI saw an opportunity to fill this gap with AI.
Time and cost factors drove the company to choose GPUaaS over purchasing GPUs outright. Using the service, Soombit AI trained its models on 14 million chest X-ray data points and conducted multiple experiments. The company said the service's period-customizable GPU allocation allowed it to run diverse experiments without significant cost burdens. The company's goal of developing a multimodal foundation model encompassing CT, MRI and other imaging data alongside clinical records beyond X-rays also favored GPUaaS. As data volumes grow, the need for flexible cloud infrastructure increases accordingly.
Upstage, which develops its own large language model (LLM), chose Samsung SDS's GPUaaS early on for its high performance. Upstage's LLM is based on hundreds of billions of parameters, requiring simultaneous use of multiple GPUs for development. Seamless communication between GPUs was also essential, as any interruption in training on even a single unit would cause significant losses. With training LLMs on high-performance GPUs in compressed timeframes identified as critical, the company selected Samsung SDS's service, which offered access to Nvidia's H100 — the best-performing GPU available last year. "As more companies worldwide sought H100s last year, it became increasingly difficult for firms to actually procure GPUs," an Upstage official said. "By choosing a cloud approach that provided smooth access to high-performance Nvidia GPUs even amid the global GPU supply crunch, we were able to maintain our AI development speed while reducing both the upfront investment costs and management burden of securing GPUs."
Some startups have also noted that domestic cloud providers offer more stable GPUaaS than their foreign counterparts. Theori Korea, a startup providing AI-based security audit solutions, uses NHN Cloud's GPUaaS to build the LLM needed for its security services. The company had previously experienced outages and other emergencies while using a foreign cloud provider's service but was unable to receive timely support due to time zone differences and language barriers.
Industry observers expect even more startups to adopt GPUaaS beyond these factors. As related technologies advance — from AI agents operating 24 hours a day, 365 days a year, to world models that understand real-world dynamics and simulate the future — demand for GPUs will only intensify. An official at an early-stage startup currently developing a world model said, "We received seed funding from venture capitalists and bought GPUs first. Even so, we still have a significant GPU shortage."