Key Takeaways

  • Apple and Meta are pushing more AI inference onto phones, glasses and other edge devices.
  • Local processing can reduce latency, bandwidth use and energy per task, but cloud infrastructure remains central.
  • Enterprises will need to assess workloads individually rather than treating edge AI as a universal substitute for data centres.

Apple and Meta are converging around a consequential idea for the AI industry: more computational work can happen where users interact with AI, rather than inside a distant data centre.

Apple Intelligence’s on-device capabilities and Meta’s AI glasses represent different product strategies, but both test the same underlying model. Phones, wearables and other edge devices can run selected inference tasks locally, while cloud systems handle training, complex reasoning and workloads that exceed device limits.

That division of labour could reshape AI infrastructure spending. It also complicates the popular assumption that each new AI feature automatically requires another large cluster of data-centre GPUs.

The energy backdrop is hard to ignore. According to the International Energy Agency, global data-centre electricity demand grew 17% in 2025, while demand from AI-focused facilities rose 50%. Overall data-centre electricity consumption is projected to increase from 485 TWh in 2025 to 950 TWh in 2030. AI-focused demand is expected to triple over that period.

Capital spending is rising just as quickly. Five major technology companies spent more than $400 billion in 2025, driven largely by data-centre investment, and that total is forecast to rise another 75% in 2026.

Edge inference will not make those data centres disappear.

Large models still need extensive infrastructure for training, fine-tuning and demanding inference. Agentic systems, video generation and requests involving large amounts of context can quickly exceed the processing capacity, memory or battery budget of a consumer device. NVIDIA data-centre GPUs therefore remain complementary to device processors, not obsolete because of them.

Qualcomm Snapdragon platforms illustrate the other side of that stack. Dedicated neural-processing components can execute compact models on phones and wearables, giving developers a way to deliver AI functions without sending every request across a network. That can lower response times, preserve functionality during weak connectivity and reduce the amount of potentially sensitive information leaving a device.

Privacy may become as important as energy efficiency. Processing voice, images or business documents locally can narrow the volume of raw data transmitted to cloud services. It does not remove governance concerns, however. Models can still produce inaccurate results, local data can still be exposed, and software updates can change system behaviour.

The NIST AI Risk Management Framework offers organizations a structure for identifying and managing those risks across the AI lifecycle. The ISO/IEC 42001 standard provides a related management-system approach for governing how AI is developed, deployed and monitored. Both are relevant when inference moves across devices, private infrastructure and public cloud services.

For enterprise technology leaders, architecture decisions will become more granular. A transcription feature might run locally. A customer-service agent could use a device model for basic classification, then call a cloud model for detailed reasoning. Video generation would probably remain cloud-heavy. There is no single split that works for every workload.

Cost calculations also need care. Local inference can reduce network traffic and cloud charges, but it may require newer devices, model optimization and additional software testing. Battery consumption matters too. Shifting electricity use from a data centre to millions of endpoints changes where energy is consumed; it does not automatically prove that total consumption has fallen.

Could device-level efficiency be overwhelmed by higher usage? Quite possibly. The IEA says electricity use per AI task is declining rapidly, yet increasingly intensive applications are expanding. Lower costs can encourage developers to place AI into more interactions, creating a rebound effect in which aggregate demand rises even as each task becomes more efficient.

Apple and Meta are therefore testing more than a product design. Their shared direction points toward a hybrid AI market in which edge processors and cloud accelerators grow together. The winners may be the organizations that match each workload to the smallest practical computing environment, while measuring performance, privacy, cost and energy rather than assuming local or cloud processing is inherently superior.