Build Applications on NVIDIA BlueField Faster with NVIDIA DOCA Agent Skills | NVIDIA Technical Blog - NVIDIA Developer

NVIDIA is replicating its CUDA ecosystem lock-in strategy onto the DPU track. DOCA Agent Skills is not a developer-tool refresh—it is a strategic repositioning of BlueField-3 from an ops-oriented offload engine into an application development platform. By packaging network-compute-security capabilities into callable skill modules, NVIDIA compresses cloud-vendor development cycles (VPC isolation, storage offload, zero-trust policy) from months to weeks, directly severing the value chain of legacy NIC vendors. Geopolitically, DPUs inherently carry data-sovereignty functions. DOCA's closed ecosystem will accelerate China's cloud providers toward domestic DPU alternatives, yet North American hyperscalers face prohibitive switching costs—BlueField dependency is still climbing. On competition: AMD Pensando trails by a full generation in dev-tooling; Intel's IPU roadmap is effectively stalled; Marvell OCTEON lacks AI-inference integration. NVIDIA's play is not peak performance—it is minimizing full-stack friction, the same playbook that made CUDA untouchable. The 12-24 month thesis: DPUs will evolve from offload engines into edge-inference nodes, dissolving the last-mile communication bottleneck in GPU clusters via DPU-side inference pre-processing. NVIDIA's endgame is making BlueField the network extension of the CUDA ecosystem—a dual-chip GPU+DPU lock-in. Whoever ships Agent on DPU first defines the next-generation data-center architecture.