From Cloud to Core: The Rise of On-Device AI and the Pressure on SaaS

From cloud AI to local processing

For most of the past decade, artificial intelligence has lived in the cloud. Data is sent to remote servers, processed there, and returned to the user. That model is now being challenged. More AI workloads are moving directly onto phones, laptops, and connected devices. These systems run models locally, without a constant connection to data centers.

The shift is driven by clear limits in cloud-based AI. Sending data back and forth adds delay. It raises privacy risks. It also ties software companies to large and unpredictable infrastructure costs. As AI features spread across everyday products, these drawbacks become harder to ignore.

On-device AI changes the economics. Large language models and other systems can now run directly on smartphones, laptops, and IoT devices. They work offline, respond faster, and give users more control over their data. That capability undercuts a core assumption of traditional Software as a Service: that intelligence must live on centralized servers. As local inference improves, the subscription-heavy SaaS model faces real pressure.

Why cloud-based SaaS is struggling with AI costs

Traditional SaaS pricing is built on stability. Companies charge per user, per seat, or per unit of storage. AI breaks that model. Most AI services are priced by tokens, queries, or compute usage. Costs fluctuate with user behavior and model complexity.

This creates what some analysts call an “AI cost illusion.” The price per million tokens for older models keeps falling. At the same time, total token usage is rising sharply as applications chain together prompts, embeddings, vector searches, and long-running workflows. The result is higher overall spending, even as unit prices drop. This mismatch makes financial planning difficult for SaaS providers that depend on cloud inference.

As AI features become standard, SaaS companies are forced to absorb or pass on these costs. Either option weakens the predictability that made SaaS attractive in the first place.

What on-device AI changes: efficiency, privacy, sustainability

Running AI locally avoids many of these problems. Processing data on the device removes the need for constant data transmission to cloud servers. Studies estimate that local inference can use 100 to 1,000 times less energy per task than cloud-based processing, largely because it avoids network and data center overhead.

Privacy is another major factor. When data stays on the device, it is less exposed to breaches or misuse. This matters most in personal devices. Features such as facial recognition, biometric authentication, and personal assistants already rely heavily on local processing for this reason.

The environmental impact is also measurable. Running AI workloads on devices such as the Samsung Galaxy S24 can reduce energy use by up to 95 percent and carbon emissions by 88 percent compared with cloud platforms like Google Colab. These gains come from chips designed for efficiency rather than raw compute power. Software optimizations in data centers help, but they do not close this gap.

Together, lower energy use, stronger privacy, and predictable costs make on-device AI a serious alternative to cloud-heavy SaaS.

The technology making on-device AI viable

This shift is enabled by advances in both hardware and software. Apple’s M5 chip is a clear example. It includes a 16-core Neural Engine and unified memory bandwidth of 153 GB/s. On devices such as the 14-inch MacBook Pro and iPad Pro, this allows local execution of diffusion models and large language models through tools like webAI.

Apple’s own local models are not the most advanced in raw capability. Their advantage lies elsewhere. Apple controls consumer hardware and operating systems and can build large semantic indexes from on-device data while keeping much of that processing local. Its “private cloud compute” approach blends limited cloud use with device-first design.

Other companies are pushing similar ideas. Liquid AI’s Liquid Foundation Models are designed to run across devices ranging from wearables to cars. They focus on efficiency, multimodal input, and privacy-sensitive deployment. The company’s LEAP platform and Apollo app lower the barrier to building and testing these models locally.

At the software level, techniques such as quantization, pruning, knowledge distillation, LoRA, and QLoRA reduce model size while preserving performance. On iOS, Swift and Core ML provide direct access to Apple’s Neural Engine, making it easier to ship personalized, device-level AI features.

These tools reduce reliance on centralized infrastructure and make local AI practical at scale.

Market growth and consumer demand

The market data reflects this momentum. One forecast projects the on-device AI market growing from USD 26.61 billion in 2025 to USD 124.07 billion by 2032, a compound annual growth rate of 24.6 percent. Another estimate puts the market at USD 160.24 billion by 2029, growing at 34.5 percent annually. Smartphones dominate, with a projected 46.2 percent share in 2025.

Consumer behavior supports these numbers. Around 70 percent of Chinese smartphone users in the Asia-Pacific region prefer devices with on-device AI features. Samsung’s AI-enabled phones account for roughly 40 percent of global smartphone sales. In 2023, on-device AI already represented about 15 percent of the overall AI market.

Underlying costs are also falling. According to the Stanford AI Index, inference costs for systems comparable to GPT-3.5 dropped by a factor of 280 between 2022 and 2024. Hardware costs declined by about 30 percent per year, while energy efficiency improved by roughly 40 percent annually. Open-weight models have narrowed the performance gap with closed systems. These trends favor self-contained AI over cloud dependence.

How user behavior and SaaS models are changing

On-device AI is reshaping everyday interactions. Google now uses local AI for scam call detection. Apple Intelligence handles private messaging features, text summarization, and other tasks directly on the device. Image correction, summarization, and real-time filtering increasingly happen without a round trip to the cloud.

This reduces the need for standalone SaaS products that once provided these functions. In response, many companies are adopting hybrid architectures, splitting workloads between device and cloud depending on complexity. This improves latency and privacy but weakens the logic of pure subscription models.

Apple is less exposed than most SaaS companies because its core business is hardware. For others, the shift is more disruptive. Deloitte’s 2026 TMT predictions show AI agents reshaping budgets, customer experience, and work patterns, with more than half of the firm’s forecasts tied to AI. To adapt, SaaS companies must rethink data management, governance, pricing, and compliance in a world where users increasingly expect fast, local intelligence.

What comes next

The spread of on-device AI has policy and strategic implications. Governments could accelerate adoption to capture energy savings and lower emissions. Developers in Apple’s ecosystem are being pushed toward Core ML and Metal Performance Shaders to improve performance without cloud costs.

For SaaS providers, the path forward is narrower. Hybrid architectures are becoming the default, as studies comparing Nvidia cloud GPUs with Snapdragon-based devices show clear efficiency gains at the edge. Careful model selection and optimization are essential. Pricing models must reflect the more predictable costs of local inference rather than volatile token usage.

Companies that fail to adapt risk being bypassed by software that lives closer to the user.

Conclusion

On-device AI is no longer a niche optimization. It is reshaping how intelligence is delivered, paid for, and trusted. Faster responses, lower energy use, and stronger privacy all favor local processing. As the market grows and consumer expectations shift, traditional cloud-centric SaaS models are under strain.

Some companies will adapt by blending cloud scale with device-level efficiency. Others will not. The balance between cloud and device will define the next phase of software. The outcome will matter not just for SaaS economics, but for how AI fits into everyday digital life.