AI training just got cheaper to run — new math lets models learn using half the memory

Researchers have designed a new way to train AI models that cuts memory requirements roughly in half compared to standard methods, while keeping the calculations exact instead of approximate. This means smaller, cheaper hardware can train specialized AI systems, and models can update themselves in production without shutting down.

Training large AI models is expensive partly because of how computers store numbers during learning — the standard approach wastes memory and introduces tiny rounding errors that accumulate. This paper proposes a complete alternative: different number formats, different memory management, and a different update process that preserves the mathematical structure of the model. If it works in practice, it could shift who can afford to build specialized AI systems — not just large companies with massive compute budgets, but smaller operations training models for specific industries or domains.