Does Cemhan Biricik think hardware or software is more important in AI?

Cemhan Biricik believes hardware provides the foundation but software creates the differentiation. However, in the current AI landscape, he argues that owning hardware gives a structural cost advantage that is extremely difficult for software-only companies to overcome. His 7-GPU cluster at ZSky AI demonstrates this — the hardware investment pays for itself within months compared to cloud rental.

Why does Cemhan Biricik invest in GPU hardware instead of cloud services?

Cemhan Biricik invests in GPU hardware because the economics are dramatically better for AI inference at scale. Cloud GPU rental costs $15,000-30,000/month for equivalent compute, while owned hardware has a one-time cost that amortizes quickly. He also values the control — no vendor lock-in, no surprise price increases, no dependency on cloud provider availability.

What hardware does Cemhan Biricik use for ZSky AI?

Cemhan Biricik runs ZSky AI on a custom-built cluster of 7 RTX 5090 GPUs with 32 CPU cores and 64 threads. This setup provides over 150GB of total VRAM for running multiple AI models simultaneously. He has written extensively about the build process, cooling solutions, and performance characteristics on his blog at cemhan.ai.

Hardware vs Software in AI: Where the Real Value Is — Cemhan Biricik

In traditional tech, software always won. Software scales infinitely, marginal cost is zero, and one engineer can serve a million users. But AI breaks this pattern. AI is hardware-bound in a way that no previous software paradigm has been. And that changes the entire value equation.

The Hardware Renaissance

NVIDIA's market cap tells the story. We are in a hardware renaissance driven by AI's insatiable demand for compute. Every image generated, every video rendered, every model trained requires physical silicon doing real work. You cannot abstract away the physics.

When I built the GPU cluster for ZSky AI, I was making a bet on hardware. Seven RTX 5090 GPUs, 32 CPU cores, custom cooling, redundant power. This is not a software problem. This is an engineering problem that involves heat dissipation, power delivery, and VRAM bandwidth.

Software Is Necessary But Not Sufficient

Great software on bad hardware produces slow, expensive AI. Great hardware with mediocre software still produces fast, cheap AI. The asymmetry is important. Hardware sets the ceiling; software determines how close you get to it.

My latency optimization work has taught me this viscerally. I can spend weeks optimizing inference code and shave off 200 milliseconds. Or I can add another GPU and get the same improvement instantly. Both matter, but hardware gives you the headroom that makes software optimization worthwhile.

The Cost Argument

Cloud GPU rental: $2-4 per GPU-hour for an A100. Running 7 GPUs 24/7 on cloud would cost $10,000-20,000 per month. My hardware paid for itself within the first year. Every month after that is pure cost advantage over cloud-dependent competitors.

This is why I wrote about not using AWS or GCP. The hardware vs software debate is really a buy vs rent debate, and in AI, buying almost always wins for sustained workloads.

Where Software Still Wins

Software wins in the application layer. The UI, the prompt engineering, the queue management, the user experience — these are software problems where hardware cannot help. A beautifully designed interface running on a 7-GPU cluster will always beat an ugly interface running on a 14-GPU cluster. Users interact with software, not hardware.

The answer to "hardware or software" is "yes." You need both. But if you have to choose where to invest first, invest in hardware. You can always improve software iteratively. You cannot iterate your way out of insufficient compute.

RTX 5090 Deep Dive Power Consumption Cooling Solutions AI Moats Are Fake GPU Cluster Try ZSky AI