Software engineering faces deep transformation shaped by token economics, compute constraints, and energy limits rather than continuous exponential AI progress

AI adoption is automating parts of software development, but progress is constrained by economic and physical factors rather than following a continuous exponential trajectory. Training compute has historically grown by a factor of 4–5x year on year, driven by more complex processing and larger datasets.

Token economics will become a first-order constraint as AI subsidies unwind and pricing reflects true costs. Organizations will need to assess whether building more advanced models is justified for a given use case, likely leading to fragmentation with more specialized models. The cost of tokens is driven by compute efficiency (GPU purchasing cost, electricity, and data centre operations) and data acquisition costs, which grow as higher-quality curated data becomes necessary.

Supply constraints in semiconductors, electricity production, and data centre capacity risk derailing AI's growth. Data centres are projected to consume an ever-growing share of global electricity, and expanding nuclear energy or upgrading transmission infrastructure will take years to materialize. Improvements in compute efficiency—such as quantisation, pruning, and more efficient attention mechanisms—provide partial counterbalance, but long lead times limit expansion.

Distillation (using frontier models to generate training data for smaller, task-specific models) and per-task cost optimization offer practical offsets to rising inference costs, though frontier models remain preferred for complex coding tasks where higher reliability reduces retries.