On-device generative AI could make smartphones more exciting—if they can deliver on the promise
With specialized chips and extensive mobile OS integration, smartphones could become smart—even intelligent. Will users embrace the new approach?
Chris Arkenberg
Duncan Stewart
Gillian Crossan
Kevin Westcott
Ariane Bucaille
Smartphones have become the most widely used piece of consumer technology in the world.1 They have absorbed many other devices, and their advanced and miniaturized components have flowed downstream into innumerable consumer and industrial devices.2 Their at-hand convenience and utility has reshaped behaviors and the competitive landscape. Yet, despite this, recent smartphone innovations seem to have failed to excite the market, appearing more incremental than revolutionary.
Now, dominant mobile ecosystem providers are starting to reorganize their devices around next-generation operating systems and advanced chips that aim to bring generative AI into the center of the smartphone experience.3 More original equipment manufacturers are now shipping gen AI–capable smartphones.4 Looking to bottle the lighting of generative AI, providers could make smartphones exciting again, but it may not be without some risk.
Deloitte predicts that in 2025, global smartphone shipments will see a modest lift to around 7%, up from about 5% annual growth in 2024.5 Some of this lift could be due to resetting the typical device upgrade cycle as consumers upgrade to the latest models. And some may be from early adopters and developers enthusiastic about next-generation phones shipping with chips designed to support generative AI locally on-device. Deloitte further predicts that the share of shipped gen AI–enabled smartphones could exceed 30% by the end of 2025.6
There is excitement about generative AI, but can the technology deliver on its promises, and will users embrace a new way of interacting with the most widely used consumer device?7
Can generative AI on smartphones boost the next upgrade cycle?
In the near term, leading smartphone designers may see generative AI integration as a way to stoke demand for their premium models: Sales of smartphones had been down for two years prior to 2024.8 In part, this was due to a degree of market saturation: It’s estimated that nearly five billion people now own smartphones—more than half of all humans.9 Upgrade cycles have been getting longer in recent years: People have been upgrading their phones every two to three years, on average, and more households have reported feeling inflationary pressures that limit their discretionary spending.10 At the same time, more are opting for higher-end devices, knowing they will be using them for a few years.11 This has likely put more pressure on the need for not just better hardware, but also more compelling value and utility in the smartphone user experience.
In 2025, smartphones are expected to put the utility of generative AI to the test.
The first quarter of 2024 showed stronger growth in smartphone sales, from increased consumer confidence and some apparent early interest in premium generative AI–enabled devices.12 Deloitte’s 2024 Connected Consumer Study found similar evidence: Fewer households now report affordability issues affecting their purchases of connected devices.13 This recovery seems evident in Europe as well, which saw continued growth in smartphone sales during the second quarter of 2024.14 So, in 2025, the upgrade cycle is likely to rebound, more people will likely upgrade their smartphones, and more of those upgrades could be for higher-priced premium devices with onboard generative AI features.
It should be noted that, while generative AI could become a driver for smartphone upgrades, the amount likely varies between markets and age groups. The same Deloitte study shows that 7% of US respondents agree that generative AI features make them likely to upgrade their smartphones sooner than they had planned, but the number jumps to 50% for those between the ages of 24 and 45 years old, who may be more dependent on smartphones and more likely to embrace new tech.15 In Deloitte UK’s 2024 Digital Consumer Trends report, however, only 4% of UK respondents report using generative AI daily, with 23% of respondents saying they don’t find it helpful, and 19% saying they’re not satisfied with the answers it gives.16
Will generative AI help create a greater boost to smartphone upgrades? It depends on how much value and utility it delivers. In 2025, smartphones are expected to put the utility of generative AI to the test.
Generative AI in personal computers
The same considerations for user experience, utility, and value, as well as the broader pressures shaping the evolution of hyperscale generative AI, apply to a new generation of PCs shipping with on-device chips dedicated to generative AI.
Results from Deloitte’s 2024 Connected Consumer Survey suggest that consumers are interested in buying gen AI–enabled PCs: Thirty-four percent of US respondents planning to upgrade their laptops or PCs agree that generative AI chips are likely to accelerate their purchasing. Deloitte believes that individual consumers will make up about 50% of annual PC sales, so this could be an important factor.17 For enterprise buyers, there’s some uncertainty about which gen AI coprocessor models on PCs make the most sense for businesses, with various PC original equipment manufacturers offering various options at various price points.18
It’s expected that over time, most high-end PCs will have gen AI functionality via special silicon. One estimate is that 80% of all PCs will have these kinds of chips by 2028.19 Another estimate suggests that nearly 9 million “AI capable” machines were shipped in the second quarter of 2024, although it’s unclear how many of these include neural processing units strong enough to run generative AI workloads.20 Indeed, potential customers might wait a year or so for the next generation of machines to deliver greater performance before they upgrade.
Deloitte predicts that roughly 30% of all PCs sold in 2024 will have had some local generative AI processing capabilities,21 and we further predict that close to half of all PCs sold in 2025 will have this capacity.
Although the computer market is not as large as the smartphone market—about 261 million units expected to be sold in 202422 compared to 1.23 billion smartphones23—the higher average selling price of computers means they often punch above their weight in dollar terms. Computer sales are estimated by Deloitte to be about US$220 billion in 2024, compared to smartphone sales of about US$520 billion for the year.24
What is unclear is what effect gen AI–enabled machines could have on the PC sector. We believe that there will be an average selling price increase of PCs caused by these devices, adding a premium of about 15% to each PC.25 However, PC sales are expected to rise only in the single-digit percentage range in 2025.26
For consumers, advancements in components for both smartphones and PCs are likely to shape supply chains and push costs down, enabling such components to move into many more devices. Generative AI capabilities are expected to become more common across connected device categories.
Generative AI could make smartphones intelligent
The term “smart” in smartphones has often meant they’re connected and can run apps. Generative AI may offer a way for smartphones to become more personalized and aware of user interactions and intentions, and more intimate through conversational interfaces. Although prior attempts at voice assistants may not have lived up to expectations, some people are already forming relationships with the latest conversational large language models.27 This could become a new interaction paradigm with conversational AI as a way to interface with digital systems, and a new model for trusted intelligent agents that could learn to act on an individual’s behalf.
On-device gen AI models could answer questions like, “How early should I leave for my 2 p.m. appointment?” by inferring the user’s intention and understanding the full context of the user’s calendar, their location, and the best route to the destination within the timeframe. They are expected to focus on doing narrow tasks well, leveraging neural processing units that can deliver enough performance—at least 30 tera operations per second, by some estimates28—to support on-device inference. The model could further recognize if a question is beyond the local scope, and then assign the task to larger, cloud-based models better able to answer. This hybrid approach to high-performance mobile computing can allow for more immediate and secure interactions on the device, with direct access to models in the cloud.29
With smaller models running on the device, user interactions and data can be contained and secured locally as needed, and more low-latency operations that might require very fast responses, like real-time translation, could be enabled.30 These features may help secure the trust of users and offer more obvious utility. Providers may also see a new flywheel of data from user interactions that could help inform local and cloud models to deliver better results to users—and greater insights to their business.
A more distant goal is that smartphones—arguably the center of consumer interactions— could become much more personalized and intelligent, tuned to individual behaviors and predictive of our needs. (See our 2025 TMT Prediction on agentic AI.) This kind of “agentic” functionality could push smartphones—and the device ecosystem they often interact with and gradually transform—to evolve from merely “smart” to “intelligent.” (See our 2025 TMT Prediction on agentic AI.)
Just as there is strong market pressure to justify the costs of frontier models by establishing their product fit, there is pressure to make them more cost-effective to build and operate.
The coming year will show how quickly users onboard onto the new experience, testing the value—and comprehensibility—of early gen AI features. Providers are expected to roll out new features over the coming months, and are likely assuming that broader adoption will take time.31 The year ahead will likely also test the capabilities and limits of small models running on-device rather than going to the cloud. In time, this could change the economics of generative AI. If more generative AI tasks shift from expensive data centers to consumer devices, the capital intensity of the generative AI build-out could be softened
Can the industry spend its way past generative AI hype?
Just as there is strong market pressure to justify the costs of frontier models by establishing their product fit, there is pressure to make them more cost-effective to build and operate.32 Leading model providers have spent billions of dollars to develop current frontier models and are investing billions more to build out the data centers they believe will be necessary to meet demand at scale.33 By some estimates, US$600 billion is being spent each year to support generative AI.34 Such capital intensity, however, can only go on for so long before it demands economic value which, in turn, may require better product fit.
Making models smaller, reducing the amount of data they need, and breaking them apart based on the scope of workloads may be ways to potentially reduce their costs, especially for inference tasks that can scale with use. Many tasks for consumers and workers may be exposed to generative AI, and they may be addressable or augmented by cheaper and more energy-efficient small models.
However, it’s unclear how much inference will remain on-device. Current generative AI interactions and expectations have mostly been defined by public cloud-based models. It may take time for users to understand which kinds of tasks and prompts run locally, securely, and for free, and which will traverse networks to models in the cloud. Interacting with a conversational, on-device, and cloud-enabled agent is a new paradigm with unclear implications for adoption and behaviors.
Broad adoption of generative AI still faces challenges
Deloitte’s 2024 Connected Consumer Study shows that 38% of US respondents have used generative AI, and 63% of those users say the technologies exceed their expectations.35 The magic may already be there for many who have used generative AI, but providers may need to show broad utility to wider demographics to help justify the cost of a new smartphone for consumers.
Usage of generative AI on smartphones could prove confusing, as users attempt to navigate novel interactions. They may hesitate to cede their own agency to intelligent assistants that seek, for example, to manage their calendars.36 Adoption could reveal trade-offs in battery consumption, costs levied by integrated public models, and unrecognized falsehoods that could undermine high-value use cases. Building trust between users, their personal AI agent, and public models will likely take time; losing that trust could happen very quickly.
Providers likely hope that the next generation of frontier models can unlock greater value, but it remains unclear if frontier models will continue to see such growth in capabilities, or if the curve will flatten or decline. And is there enough data to feed increasingly avaricious training sets?37 Solutions like synthetic data created by models to train themselves may cause the quality of inference to degrade over time.38 Can functionality advance without higher costs in data, training, and inference? Is there a window where functionality could improve while capital and data intensity decrease? Nervous investors could potentially demand greater revenues before the technology is able to deliver them.
Regulators could also impact development of gen AI with a broader approach to safeguarding against emerging ills, like deepfakes, misinformation, and persuasive human-like bots. Conversational bots may be establishing greater rapport and intimacy with users, that are better able to influence their ideas and ideologies.39 Personalized conversational agents could tap into the deeper realms of human interactions, potentially helping more people, but also risking addictions.40 Combining on-device generative AI with third-party models could create a larger surface area of security vulnerabilities and exploits.41 This could further provoke providers to secure their ecosystems, and regulators to install more guardrails.
Bottom line
Despite recurring talk of the “next smartphone”—a consumer device platform with the potential to transform and uplift entire markets—it hasn’t come. With billions of users, smartphones still dominate and offer a large test bed for new services and user interactions. In 2025, the number of people interacting with generative AI will likely get a boost through premium smartphones—and through personal computers. They can try it, learn its value, and test for its edges. If it succeeds, smartphones could become more exciting and could help expand their platform to enable entirely new categories of use and opportunity—and drive a new boom in personal devices. But this will likely take time, and the coming year is expected to be an onboarding effort to help introduce users to a new paradigm for personal computing.
In the coming years, the smartphone operating system could capture more interactions, such as the next generation of conversational search that can return more local summaries than remote links, disintermediating service providers and information sources. If users adopt more personalized agentic AI, the nature of digital interactions could change, potentially off-loading more tasks to a user’s device rather than demanding direct user interface. In this manner, computing could become more ambient, operating in the background on our behalf, and potentially more spatial—increasingly aware of our surroundings and network interactions.
As providers work to stoke demand, they may find themselves racing against economic pressures to offset the capital intensity and energy costs of training and operating models at scale. The industry could pursue small models, hybrid architectures, and a deeper understanding of which generative AI workloads require which kinds of computational overhead. At a time when climate uncertainty and anxiety is difficult to escape, the gen AI data center buildout is already driving up energy and water usage, as well as the energy costs borne by households and municipalities.42 If generative AI surmounts its economic debt, it may yet find itself impaired by energy debt.
As of late 2024, the bet for hyperscalers, smartphone ecosystem owners, and young public models is that the benefit they provide will turn into broad economic value. But how much of that value will they capture? Will generative AI hyperscalers follow the path trod by telecoms and the early internet, spending down their reserves on massive capex just to build the infrastructure that could ultimately power the next generation of innovators?43
Driving the emergence, deployment, and broad adoption of generative AI could constitute one of humanity’s grandest experiments since unleashing the internet to the masses—a moonshot that, destination aside, could deliver a new flood of technologies, behaviors, and business models through its development.
