Monolithic vs Microservice
I am not talking about websites, even though the concept of microservice started from web application architecture. I am talking about methodologies in general. I am not boasting here - I am an expert in microservice architecture, in cloud or on-prem. Maybe some people would argue microservice and on-prem are oxymoron. But in my practices, software architecture has become ever more platform agnostic.
So why is above related to DeepSeek vs NIM?
From software architecture standpoint, up until DeepSeek R1 release, all hyperscalers deploy LLM models in a monolithic architecture. History of software engineering has already demonstrated that paradigm is shifting toward microservice architecture from monolithic architecture, particularly loosely-coupled micro-service architecture. There are exceptions in corner cases in various vertical problem domains. But general trend for large-scale cloud-native SaaS is going loosely-coupled micro-service architecture. DeepSeek R1 is the first top-performing tier-one model that can fit in, while other top models that hyperscalers have deployed are all monolithic that can only be deployed in mega data centers with humongous GPGPU farms. Therefore, one of the fundamental impacts DeepSeek brings upon the current AI revolution is that it has started a paradigm shift in deployment architectures. In other words, it is not that DeepSeek is cheap and free that bring so much fear, rather it is innovations in its implementation techniques as well as overall software architecture that has fundamentally change the game of generative AI.
This is just the beginning ......