Blink: CPU-Free LLM Inference by Delegating the Serving Stack to GPU and SmartNIC

Blink represents a high-end systems research approach to solving the 'noisy neighbor' and CPU bottleneck problems in LLM serving. While the project currently has 0 stars, the 5 forks within 9 days of a paper release (likely Arxiv/SOSP/OSDI track) indicate immediate interest from the systems research community. The defensibility is high (7) because building a CPU-free stack requires deep co-design of GPU kernels and SmartNIC networking, a skillset far beyond typical application developers. This is not just a wrapper; it's a fundamental re-architecture of the serving stack. However, the platform domination risk is 'high' because the primary beneficiaries are hyper-scalers (AWS, Google, Meta) and hardware providers (NVIDIA), who are incentivized to build similar proprietary offloading capabilities into their own stacks (e.g., NVIDIA's BlueField/DOCA ecosystem). Blink competes conceptually with vLLM and HuggingFace TGI, but specifically targets the infrastructure inefficiency those projects currently ignore by relying on host OS scheduling. Its moat is the complexity of implementation, but its weakness is the requirement for specific hardware (SmartNICs) and the niche nature of low-level systems optimization.