China AI startup DeepSeek\'s latest reasoning model, DeepSeek-R1, recorded a much higher hallucination rate than its previous model, according to an evaluation by Vectara using the HHEM 2.1 benchmark. DeepSeek-R1 logged a 14.3 percent hallucination rate, about four times DeepSeek-V3\'s 3.9 percent. Vectara said cross-checks using Google\'s FACTS methodology showed a stronger tendency to add unsupported information, a concern as AI agent tokens automate crypto analysis and trading.

Surging electricity demand from the spread of AI is exposing the limits of transformer-based large language models and boosting calls for next-generation architectures. With many models still relying on more compute, layers and data, analysts expect data centre spending to rise sharply by 2030. Inference-focused models can consume far more power per prompt. Industry attention is turning to post-transformer designs that activate only relevant parts of a model, aiming to cut inference costs, reduce token use and remain compatible with existing infrastructure.