How do we build Nebius Token Factory to address the challenges of building production-scale inference systems? How do the customers' SLAs look, and what delivers optimized performance around the tradeoffs of cost, latency, throughput, and quality?

As organizations race to deploy AI into real products, the gap between public models and providers' benchmarks and real-world performance has never been more apparent. In this keynote, we’ll explore why metrics fail to capture the true behavior of production-scale inference systems, and what it actually takes to build them.

We’ll share how we designed Nebius Token Factory to meet demanding customer SLAs and deliver predictable performance under real workloads. The talk will walk through the practical trade-offs between cost, latency, throughput, and quality, and the engineering techniques that make it possible to optimize across all four.