Why current LLM costs are not sustainable

4 hours ago

Many companies face high AI costs, with examples like Uber exceeding budgets and others like Microsoft cutting spending.
AI offers significant benefits in programming, design, and data tasks, but frontier models like GPT 5.5 are expensive due to high input/output token costs.
Model performance plateaus, reducing justification for price hikes as improvements diminish and training data becomes scarce.
Open-weight models, such as GLM-5.2, offer competitive performance at lower costs by avoiding the overhead of frontier AI labs.
Specialized chips (e.g., TPUs) and model architecture advancements are lowering inference costs over time.
Zero switching costs, enabled by AI gateways, allow quick model changes, increasing competition and driving prices down.
Local models will become feasible with chip and RAM improvements, reducing reliance on cloud models for simple tasks.
Predictions indicate overall price pressure will benefit consumers through reduced AI costs in the future.

Hasty Briefsbeta