Generative artificial intelligence (GenAI) tools are increasingly used in physics classrooms, yet little is known about why high school students accept them for subject-specific learning. Prior technology-acceptance research, especially the Technology Acceptance Model (TAM), shows that perceived usefulness and perceived ease of use shape attitudes and behavioral intention toward a tool. However, this work has rarely examined whether students’ beliefs about the subject they are learning also shape those technology perceptions. Drawing on TAM and Expectancy-Value Theory (EVT), we surveyed 340 Grade 10 students who had used DeepSeek in physics lessons and tested an extended structural equation model. Results showed that students were more willing to use DeepSeek when they perceived it as easy to use and useful and when they held positive attitudes toward using it for physics learning. Physics-specific beliefs helped explain these technology perceptions: students with stronger expectancy for success in physics perceived DeepSeek as easier to use; students who saw physics as useful for their future perceived DeepSeek as more useful; students who enjoyed physics held more positive attitudes; and perceived cost weakened both ease-of-use and attitude judgments. Students who regarded doing well in physics as personally important also tended to evaluate DeepSeek more favorably, partly because they saw it as useful for achieving physics-learning goals. Nested model comparisons clarified what the EVT extension added beyond the basic TAM framework. Adding physics task values produced the largest incremental gain, especially for perceived usefulness (ΔR²=0.151). Supplementary follow-up evidence indicated that the same level of intention could reflect learning-oriented or shortcut-oriented use, and that some students linked shortcut-oriented use to low confidence and topic difficulty rather than to disinterest alone. Overall, the findings suggest that GenAI acceptance in physics is shaped by both technology perceptions and students’ motivational relationship with physics, supporting a domain-sensitive extension of TAM for subject-specific GenAI learning.