Worldwide end-user spending on generative AI (GenAI) smartphones is projected to total $298.2 billion by the end of 2025, according to Gartner, Inc. a business and technology insights company. It will represent 20% of total AI end-user spending in 2025.
Gartner defines GenAI smartphones as devices equipped with a built-in neural engine or neural processing unit (NPU) capable of running small language models. The forecast covers both premium smartphones and basic smartphones (under $350) but excludes utility smartphones as they are no expected to have NPU capabilities at any time.
“Currently, most users still rely on text or touch for all tasks, and voice interactions remain limited in scope. However, as conversational AI becomes more natural in the future, users are expected to become more comfortable with AI as a proactive digital companion, rather than just a reactive tool,” said Ranjit Atwal, Senior Director Analyst at Gartner.
With all mobile vendors increasingly integrating on-device GenAI models and applications, end-user spending on GenAI smartphones is expected to total $393.3 billion in 2026, an increase of 32% from 2025 (see Table 1). Gartner expects that 100% of premium smartphones will feature GenAI capabilities by 2029.
Table 1: Worldwide GenAI Smartphone End-User Units and Spending, 2024-2026
| 2024 | 2025 | 2026 | |
| Total GenAI smartphone units (in thousands) | 260,433.4 | 369,347.0 | 559,000.6 |
| Total GenAI end-user spending (millions of dollars) | 244,735.1 | 298,189.5 | 393,297.7 |
“The broad use of new NPUs in smartphones will allow GenAI models to run faster and more efficiently, requiring users to upgrade to the latest smartphone hardware for optimized experiences,” said Atwal. “In 2025, Gartner forecasts almost all premium GenAI smartphones will include NPUs and 41% of basic GenAI smartphones will have NPUs.”
By 2027, Gartner expects on-device NPUs performance levels exceeding 40 tera operations per second (TOPS) will become standard in premium GenAI smartphones. They will enable real-time execution of complex multimodal AI workloads without excessive power consumption.
