Seven RTX 5090s, 224GB of VRAM, and a photographer's stubborn refusal to rent what he could own.
In early 2024, I made a decision that most people in my life thought was reckless: I spent a significant amount of money building a GPU cluster from scratch instead of using cloud services. I'm a fashion photographer. I run a small AI company. I am not, by any traditional definition, the kind of person who builds data center infrastructure.
But here's the thing about the future of creative AI: the people who control the hardware control the experience. And I wanted to control the experience.
Let me start with what's in the box, because the hardware matters in ways that aren't immediately obvious.
The primary node runs seven NVIDIA RTX 5090 GPUs. That's 224 gigabytes of VRAM — the high-bandwidth memory that determines how large a model you can load, how fast you can run inference, and how many users you can serve simultaneously. The CPU is a 32-core, 64-thread processor that handles orchestration, scheduling, and the thousand small tasks that keep a multi-GPU system coherent.
I also have secondary nodes in my cluster — machines with RTX 4090s — that handle overflow and specific workloads. The whole system is networked on a local fabric that keeps inter-GPU communication fast and data transfer off the public internet entirely.
This is not a data center. It's closer to a very serious home lab that grew ambitions. And that's exactly the point.
This is the first question everyone asks, and it deserves a thorough answer.
Cloud GPU services — AWS, GCP, Lambda Labs, RunPod — are excellent products. They offer flexibility, scalability, and the ability to spin up and down as demand requires. For many use cases, they're the right choice.
For what I'm building, they're wrong in three specific ways.
Privacy at the hardware level. ZSky AI serves creative professionals who work with confidential material — unreleased campaigns, embargoed designs, proprietary brand assets. When I tell them their data never leaves my infrastructure, I mean it literally. There's no cloud provider in the chain. No third-party logging. No subprocessor with access to GPU memory dumps. The air gap between our cluster and the public internet is real and auditable.
Cost predictability. Cloud GPU pricing is volatile and, over time, expensive. An A100 instance on AWS costs roughly $3-4 per hour. Running seven GPU-equivalents at that rate for 16 hours a day comes to over $40,000 per month. My hardware paid for itself in months of equivalent compute time. The ongoing cost is electricity — roughly $400-600/month depending on load — and occasional thermal paste.
Performance consistency. Cloud instances share physical hardware. Your performance depends on what your neighbors are doing. Spot instances disappear without warning. Quota limits throttle you during peak demand — which, invariably, is when you need the compute most. Our cluster delivers the same performance at 3 AM and 3 PM, on weekdays and weekends, during product launches and quiet periods.
Building the cluster was an education in physics, economics, and humility. Here are the lessons that mattered most.
Thermal management is the real engineering problem. Seven RTX 5090s under sustained load generate an extraordinary amount of heat. My first configuration hit thermal throttling within twenty minutes of full-cluster inference. I redesigned the airflow three times before arriving at a solution that keeps all cards under 80°C during sustained operation. The cooling system is, in some ways, more sophisticated than the compute system.
Power delivery is non-trivial. The cluster draws significant power under full load. This required dedicated circuits, UPS systems, and a conversation with my electrician that started with "So, how much amperage can we add?" and ended with a minor electrical panel upgrade. If you're considering building something similar, start with the power budget. Everything else is downstream of whether your building can deliver the watts.
Software is harder than hardware. Racking GPUs is mechanical work. Making them cooperate is computer science. Model parallelism, load balancing across heterogeneous hardware, memory management during concurrent inference requests, graceful degradation when a card hits thermal limits — these are hard problems that took months to solve well. The open-source ecosystem (vLLM, ComfyUI, custom CUDA kernels) was essential, but nothing worked out of the box for our specific configuration.
Hardware is a solved problem. Making hardware serve creative workflows elegantly — that's the unsolved problem I'm most interested in.
Here's what nobody talks about in the cloud-vs-own debate: ownership changes how you think.
When compute is metered — when every GPU-second costs money — you optimize for efficiency. You run the smallest model that produces acceptable output. You limit experimentation. You batch requests. You make creative decisions based on cost, not quality.
When you own the hardware, the calculus inverts. The GPUs are there whether you use them or not. The marginal cost of an extra generation is essentially zero. This changes creative behavior in profound ways.
I experiment more. I run larger models. I generate fifty variations where I used to generate five. I fine-tune on whims — "What would photorealistic look like if it had only ever seen Helmut Newton photographs?" — because the cost of finding out is electricity and time, not a cloud bill that makes my accountant uncomfortable.
This freedom is, I believe, the actual competitive advantage of self-hosted AI infrastructure for creative work. Not the cost savings. Not the privacy (though that matters enormously). The freedom to explore without a meter running.
I think we're heading toward a bifurcation in the creative AI market.
On one side: massive cloud platforms serving millions of users with standardized models and standardized interfaces. These will be good, affordable, and generic. They'll serve the same role that stock photography serves today — competent visual content for people who need images but don't need specific images.
On the other side: smaller, specialized platforms running custom infrastructure tuned for specific creative disciplines. A platform for fashion that understands fabric and light. A platform for architecture that understands materials and space. A platform for product photography that understands surface texture and brand consistency.
ZSky AI is my bet on the second future. The seven RTX 5090s in my cluster aren't just compute — they're a creative instrument, tuned by a photographer, for photographers and visual artists who need more than generic output from a generic platform.
Without hesitation. But I'd do three things differently.
I'd start with better cooling from day one. The thermal redesigns cost me weeks. I'd invest more in monitoring and alerting infrastructure — knowing that GPU 4 is running hot before it throttles is worth every hour spent on Grafana dashboards. And I'd budget more time for the software layer. Making seven GPUs work together is straightforward. Making them work together well is a project that's never really finished.
The cluster hums quietly in the next room as I write this. Seven GPUs, 224 gigabytes of VRAM, generating images for people around the world who trust us with their creative work. It started as a photographer's side project. It became an infrastructure company. And it taught me something I should have known from years behind the camera: the best tools are the ones you build yourself.
Cemhan Biricik is a fashion photographer and the founder of ZSky AI, a privacy-first generative AI platform running on self-hosted GPU infrastructure. He writes about the intersection of photography, AI, and creative technology. More at cemhanbiricik.com.