Ievgen Vakulenko leads generative-AI go-to-market and business development at Amazon Web Services, applying a background in networking and cloud infrastructure to the economics of enterprise AI. Earlier in his career, he worked at Cisco; by 2024, he was a Crusoe product manager responsible for infrastructure and GPU networking.
At Crusoe, Vakulenko focused on rail-optimized InfiniBand networking, addressing a costly problem in distributed training: accelerators sit idle while exchanging data across large clusters. His architecture separated conventional customer-facing networking from a dedicated, high-bandwidth GPU communication fabric, while integrating GPU virtual machines, CPU instances, local NVMe storage, and developer-facing APIs and tools.
His most concrete example combined NVIDIA NCCL PXN with NVSwitch connections inside GPU servers. Routing cross-rail traffic through those internal links reduced dependence on higher-level switches, limiting extra network hops and congestion. In a Mixtral fine-tuning workload spanning 240 H100 GPUs, enabling PXN produced a reported 14 percent improvement. Vakulenko considers application-level results more useful than synthetic networking benchmarks because they translate into faster training and lower infrastructure costs.
- Time-to-train as a product metric: Evaluate networking decisions by GPU utilization, training duration, and customer expense.
- Energy-aware data-center placement: Situate compute infrastructure near renewable, stranded, or otherwise underused energy sources while accounting for customer geography and latency.
- Developer-centered infrastructure: Combine compute, storage, fault isolation, networking, and usable interfaces into an environment designed for machine-learning teams.
His subsequent move into generative-AI business development extends that infrastructure grounding into enterprise adoption and the commercial realities of deploying AI.
