Blog • February 2026
By Cemhan Biricik — Founder of ZSky AI
Every company is the sum of its decisions. For ZSky AI, five technical decisions made in the first few months shaped everything that followed — the cost structure, the user experience, the competitive position, and the kind of company it became. Here are those decisions and why I would make every one of them again.
This was the foundational decision. Buying seven RTX 5090 GPUs and building my own GPU cluster instead of renting cloud compute. The upfront cost was significant. But the math was clear: cloud GPU rental for the workloads I was planning would cost more per month than the hardware cost in total.
This single decision created a 25-50x cost advantage over cloud-dependent competitors. It made a genuine free tier economically viable. It eliminated dependency on any single cloud provider. And it forced me to develop deep infrastructure expertise that has paid dividends in every other technical decision since.
I chose open source AI models over proprietary APIs. This meant more engineering work — deploying, optimizing, and maintaining models myself. But it meant no per-generation API costs, no rate limits, no content restrictions imposed by a third party, and no risk of a provider changing terms or raising prices.
Open source also enabled the optimization work that made ZSky AI fast. I could quantize models, tune schedulers, and customize pipelines in ways that API-based platforms simply cannot. The speed advantage that users experience comes directly from this decision.
I launched with a free tier that was actually useful, not a crippled demo designed to frustrate users into paying. This was controversial advice — everyone said to gate features, limit quality, add watermarks. I disagreed. A free tier that people genuinely love creates the best possible marketing: users who voluntarily tell others about your product.
This only works because of Decision 1. When your marginal cost per generation is $0.003, free users are essentially free to serve. The economics of free are only possible on owned infrastructure.
Beyond the GPU infrastructure, I chose to self-host the entire stack — web server, database, CDN, monitoring, everything. No AWS, no Vercel, no managed databases. This is the most debatable decision on the list. Managed services save enormous time. But they also create costs and dependencies that compound.
Self-hosting means I understand every layer of the stack. When something breaks, I know where to look. When I need to optimize, I have access to every configuration. And the cost savings are real — my total non-GPU infrastructure cost is under $200 per month for everything that managed services would charge thousands for.
I built custom systems for GPU job scheduling and cluster monitoring instead of using off-the-shelf solutions. This was the highest-effort decision — weeks of engineering for capabilities that existing tools partially provide. But the GPU-aware intelligence in the custom systems produces measurably better results.
The queue system routes jobs based on real-time GPU state, not round-robin distribution. The monitoring system tracks exactly the metrics that matter for AI inference, not generic server metrics. These custom systems are now competitive advantages that would be difficult and expensive for competitors to replicate.
The lesson I take from these decisions is that early technical choices compound. A cost advantage gets bigger over time, not smaller. A speed advantage becomes a user experience advantage becomes a growth advantage. The decisions you make in month one define the company you have in year five. Choose the harder path if it creates structural advantages. The compound returns are worth the upfront effort.