Air Cloud Customer Case Study: How HEIMDEX Cut GPU Costs Through Infrastructure Diversification
- 5 days ago
- 7 min read
Updated: 3 days ago
Sustainable AI Infrastructure After the Credits Run Out
- Why HEIMDEX Started Diversifying Cloud Vendors with Air Cloud
Running a video AI service depends as much on the GPU infrastructure behind it as on model performance.
In the research phase, teams need to repeatedly experiment with different algorithms and models. In the service phase, they need to index customer data quickly and scale GPU resources with traffic. But because GPU usage isn't constant, it's difficult to secure the performance you need while keeping costs predictable.
HEIMDEX is a video intelligence company that uses AI to analyze scenes and meaning within video, helping users search, summarize, and manage large volumes of video content. Users can find the exact scene they need with simple search terms, people, places, dialogue, emotions, without having to watch an entire video themselves.
In this interview, we spoke with HEIMDEX co-founder and CEO Jangwon Lee and the development team about why they adopted Air Cloud, how they actually use it, and what cloud vendor diversification means for AI companies thinking long-term.
Q. Could you introduce HEIMDEX and your role there?
Hi, I'm Heejo Kang, co-founder and CSO of HEIMDEX. We're building a video intelligence service where AI understands what's happening in a video so users can search for the scene they need in natural language.
Video contains far more than dialogue. Faces, objects, places, actions, and emotions are all embedded in it. HEIMDEX analyzes this information with AI and indexes it by meaning, turning the video data companies already have into a searchable, reusable asset.
I oversee HEIMDEX's technology and service operations overall, including per-customer container deployment, GPU infrastructure configuration, and how we scale workloads.
Q. What infrastructure challenges have you faced running an AI service?
HEIMDEX's GPU workloads fall into two broad categories: research and experimentation, and live service operations.
In the research phase, we connect to servers over SSH to repeatedly compare algorithms and models. In the operations phase, we deploy per-customer workers and apply different settings for each customer through environment variables. Indexing a new customer's video data for the first time requires a large amount of GPU resources in a short period, and once that initial indexing is done, the resources needed drop back down. Keeping the same number of GPU servers running at all times creates unnecessary cost, while manually building servers each time they're needed slows down how quickly we can respond to customers. That's why flexible infrastructure, one that can quickly secure resources when needed and scale back down once the workload drops, mattered so much to us.
Q. Why did you look into other clouds when you still had AWS credits available?
The AWS credits available to startups help a lot with reducing early infrastructure costs. But credits aren't a permanent way to save money. As a service grows and GPU usage increases, you eventually have to think about your cost structure after the credits run out. And if all your workloads depend on a single cloud, your options can become limited later if costs or operating conditions change.
CEO Jangwon Lee explained it this way:
"AWS credits definitely help ease the initial cost burden. But what matters in the end is whether you can keep running the service sustainably after the credits are gone. As a service grows, you have to look at cost structure and scalability together, and at some point, you need to think about diversifying your cloud vendors."
For HEIMDEX, vendor diversification isn't just about adding a cheaper cloud. It's a strategy for reducing dependence on any single vendor while increasing the stability and scalability of service operations. Because Air Cloud can support everything from fast experiments and PoCs to customer service deployment and large-scale indexing, it became a key part of HEIMDEX's sustainable AI infrastructure.
Q. What was your first impression when you started using Air Cloud?
The HEIMDEX team had already used several GPU clouds, including RunPod. Once Air Cloud added SSH support, our researchers could carry over the same repeated experiments they used to run on other GPU clouds without much friction. Getting from sign-up to the console wasn't complicated either. The team noted that the structure, organization, project, endpoint, is simpler than AWS's, which meant even developers new to GPU clouds could quickly find what they needed. The HEIMDEX development team explained it this way:
"AWS has so many features that it takes time to figure out where to even start as a new user. Air Cloud's structure is relatively simple and intuitive, so projects and containers were easy to understand."
Being able to see environment variables and container settings on a single screen was also a convenience for whoever handled deployment.
Q. What advantages did you notice on the cost side?
Air Cloud's price competitiveness was most noticeable during research and experimentation. Our research team repeatedly spins up and tears down GPU environments to compare different models and algorithms. These don't need to run constantly like a production server, but they do need enough GPU performance while the experiment is running.
Hyojeong Ryu, Senior Developer and Researcher at HEIMDEX, put it this way:
"When we used other GPU clouds for experiments, the cost felt pretty high. With Air Cloud, it was clear that experimentation was noticeably cheaper."
CEO Jangwon Lee also pointed to fast experimentation and PoC-building as one of Air Cloud's strengths.
"When you're running a simple experiment or building a PoC, how fast you can get infrastructure ready matters a lot. With Air Cloud, you can spin up the GPU environment you need quickly, without configuring a lot of complex cloud features one by one. Factor in the scalability from autoscaling, and for these kinds of workloads, it's clearly more intuitive and more affordable than AWS."
Q. How are you using Air Cloud today?
HEIMDEX uses Air Cloud differently depending on each team member's role.
Researchers connect over SSH to repeatedly compare algorithms and models, while whoever handles deployment manages the workers and containers each customer's service needs. Even when the same image is used, per-customer environment variables keep each customer's environment separate.
GPU usage temporarily spikes when a new customer's data is indexed for the first time. HEIMDEX currently adjusts the number of containers via API based on workload, and is also working with the Air Cloud team to validate more advanced autoscaling capabilities.
Being able to quickly spin up multiple containers when workload increases, then scale resources back down once the job is done, matters a great deal for a service like HEIMDEX that handles a lot of per-customer initial indexing work.
Q. What matters most to you when it comes to autoscaling?
For HEIMDEX, autoscaling isn't just about response speed. Convenience and cost optimization need to work together.
"Indexing a new customer's data for the first time needs a lot of GPUs for a short window. On the flip side, if containers keep running after the job is done, that costs us money. What matters is scaling up fast when it's needed, and scaling back down automatically when it isn't."
For example, if a large indexing job runs overnight and someone has to keep checking its status and manually shut down containers once it's done, a lot of the operational efficiency you'd expect from the cloud gets lost. HEIMDEX expects that as autoscaling features, like container scale-up/down alerts and automatic scale-down after a job finishes, get more advanced, they'll be able to cut both cost and operational burden at the same time.
Q. Was there any support experience with Air Cloud that stood out to you?
The HEIMDEX team valued not just the features themselves, but how quickly improvement requests were addressed.
In the early days, Air Cloud didn't have SSH, so we assumed it served a different purpose than the GPU clouds we already used. Once SSH was added, it became a natural fit for research and experimentation too. The Air Cloud technical team also responded quickly to an issue we ran into while entering environment variables. We're still sharing operational feedback today, on build logs, documentation, cost management, and container search, and Air Cloud is folding that feedback into product improvements.
Having a provider look directly at your workload and improve the product alongside you is a different kind of advantage than simply being handed a finished piece of infrastructure.
Q. What kind of companies would you recommend Air Cloud to?
I'd recommend it to AI startups that are using AWS credits but already thinking about their GPU costs once those credits run out. You don't need to migrate all your infrastructure at once. You can start by testing a smaller, self-contained workload on Air Cloud first, an experimental environment for validating a new model, or a customer PoC.
It's especially well-suited to teams like these:
Development teams that repeatedly experiment with AI models
Services that need a lot of GPU capacity in a short window when onboarding new customers
Companies running video analysis, generative AI, or search/indexing workloads
Startups looking to diversify cloud vendors while lowering GPU costs
Teams that want to focus on fast PoCs and product validation rather than complex infrastructure setup
CEO Jangwon Lee summed it up this way:
"Choosing a cloud shouldn't come down to the credits you have available right now. You also need to look at whether the cost still works once your service has grown. Air Cloud can be a good choice because you can start with small experiments and scale up to real service operations."
Design AI Infrastructure That Outlasts Your Credits
Cloud credits can help a startup get off the ground, but they don't guarantee your service stays sustainable.
HEIMDEX is preparing for cloud vendor diversification by separating its workloads, research and experimentation versus customer data indexing, and choosing the right infrastructure for each. Through fast GPU environment setup, intuitive container management, cost competitiveness that's especially clear during experimentation, and a scalable structure, Air Cloud helps AI startups design infrastructure that lasts well beyond their credits.
If GPU infrastructure costs or complex cloud setup are slowing down how fast you can experiment and ship your AI service, consider evaluating Air Cloud based on your actual workload.
🙋♀️ Does HEIMDEX's story sound familiar?
