Upscale, a startup with NVIDIA backing, is building software to make AI chips from competing vendors work together in the same data centre. The technical rationale is straightforward: cloud providers are designing their own silicon, specialised accelerators are entering the market, and NVIDIA GPUs can no longer be assumed as the only hardware in an AI cluster. Orchestrating workloads across that mix is difficult when each vendor’s stack was designed to run alone.
For companies staffing AI infrastructure teams, the coordination layer is where hiring gets hard. A client in Frankfurt building out a hybrid GPU environment asked us for someone who could optimise workloads across NVIDIA and custom TPU-style accelerators while keeping the MLOps pipeline stable. The role sat open for four months. Candidates with deep NVIDIA experience struggled to demonstrate fluency in alternative architectures. Candidates from cloud-native backgrounds lacked the low-level GPU programming.
Interoperability software like Upscale’s may eventually abstract some of that complexity, but the abstraction still needs someone who understands what sits underneath. Companies running heterogeneous AI compute are already paying premiums for engineers who can move between vendor ecosystems without retraining from scratch. The rarer profile is someone who has actually shipped production workloads on more than one accelerator type and can articulate the trade-offs to procurement and engineering simultaneously.
The hiring constraint centres on people who can hold two or three hardware paradigms in their head at once and make scheduling decisions that affect both cost and model performance. As DACH companies build out sovereign AI infrastructure and reduce dependence on any one supplier, that profile will determine how quickly they can deploy.
Prompted by reporting from BigDATAwire.