Technology Radar

We’re entering a new phase of AI model use, one where automatically reaching for a proprietary frontier model is no longer the obvious choice. This edition highlights significant innovation in this space, encouraging organizations to think more creatively about which models to use, how and when.

Open-weight models are particularly notable here. Twelve months ago, the gap with proprietary frontier models was stark, but today it’s much smaller for many tasks. Cost is a major driver, but not the only one: open-weight models can run in your public cloud environments, on neoclouds or on your own hardware. That flexibility can help meet strict privacy requirements and reduce dependence on a single vendor. This is especially visible in software engineering, where the GLM family has made open-weight models credible alternatives to proprietary frontier models for some coding-agent workloads. Self-hosting can also help organizations keep source code and prompts within their own infrastructure. But the shift isn’t limited to open-weight models. Local coding models are rising in popularity, while developments such as Foundation models for structured data and decision models like Jev reflect growing recognition that frontier LLMs aren’t the only approach to AI.

With this growing range of options, choosing the right model for each task is becoming an important responsibility for engineering teams. Model routing is emerging as a key consideration. We caution against blanket frontier model usage and encourage teams to make deliberate model choices while remaining open to alternatives.

While the demands of AI infrastructure have appeared in several recent Radar volumes, this edition brings them into much sharper focus. As generative AI and agent use intensify, enterprises are discovering that token economics ultimately comes down to infrastructure economics. The cost of AI depends not only on model pricing but also on model selection, routing, batching, data movement and the efficiency of the underlying platform. Much of the innovation needed to make AI economically sustainable therefore sits deeper in the stack and often falls below many people’s radar.

In putting this edition together, we encountered a vibrant ecosystem emerging around AI infrastructure. Kubernetes, in particular, is becoming a key foundation for what are sometimes described as AI factories: industrialized platforms for running AI workloads at scale. Several blips illustrate this shift. KServe and KAITO support model serving, deployment and accelerator-aware workload management. Higher in the stack, llm-d tackles distributed inference and scheduling, while LiteLLM provides routing and abstraction across models and inference providers. The AI factory stack also extends beyond compute to storage and networking. Technologies such as VAST DataStore address the high-throughput storage demands of accelerator-heavy workloads, while Netris brings software-defined network automation to the underlying infrastructure, helping teams provision and operate the high-bandwidth connectivity AI clusters depend on.

The bigger shift is that models, GPUs, inference engines, storage and networks are increasingly managed as programmable resources coordinated through a common platform layer, making AI infrastructure a central factor in enterprise AI economics.