This article investigates the morphological and orthographic compatibility of standard tokenization algorithms with the agglutinative structure of the Uzbek language. Through empirical evaluation of 15 transformer-based language models—including global large language models, multilingual encoders, and Uzbek-specialized models—significant linguistic and computational inefficiencies are identified. Experimental results based on a dataset of 357 unique affixes and a 206-sentence orthographic stress-test corpus demonstrate that algorithms such as Byte-Pair Encoding (BPE) and WordPiece frequently cause “semantic decontextualization”. These tokenizers illogically split word stems and incorrectly segment special apostrophe-based graphemes (o‘, g‘), treating them as punctuation boundaries. Such excessive fragmentation artificially increases token sequences by up to 54.5%, leading to a 138% exponential increase in Floating Point Operations Per Second (FLOPS) during model inference. The study concludes that standard tokenizers are insufficient for Uzbek NLP tasks. Therefore, it proposes the development of a tokenization algorithm that accounts for the linguistic characteristics of the Uzbek language, along with the construction of a token vocabulary, in order to improve the computational efficiency of Transformer models.