ComputeCosts Observatory Report 003

This report examines the role of local versus cloud computing in scientific computing as practiced
in chemical engineering, environmental engineering, biotechnology and process technology, with
specific attention to scientific machine learning, modelling, simulation and inference workloads
typical of PhD level research. The analysis is based on a meta analysis of publications from 2025
and 2026 extracted from arXiv and associated GitHub repositories, complemented with peer
reviewed literature in computational chemistry, engineering simulation and related fields.

Across this evidence base, a substantial body of papers explicitly reports experiments executed on
single workstation systems equipped with consumer RTX GPUs and roughly 128 GB of host
memory. The frequency of these disclosures is of the same order of magnitude as references to
cloud based computation. This is notable because the broader technology narrative strongly
promotes cloud infrastructure, largely through large scale commercial messaging, while the
research literature shows that modern local compute remains an efficient, routine and accepted
experimental platform.

Taken together, the evidence indicates that for a large majority of day to day scientific computing
tasks in these domains, a modern local workstation, for example a system with an RTX 5090 class
GPU, about 128 GB RAM, multi terabyte NVMe storage and a contemporary multi core desktop
processor, is already sufficient to execute modelling, training, inference and simulation workflows.

Cloud and cluster resources remain important for genuinely large scale computations. However,
the literature suggests a common development pattern: models are often built and stabilised on
local machines first, with larger infrastructure used later when additional scale is needed. This
reflects the iterative nature of research, where working locally helps ensure that methods behave
correctly before larger resources are applied.