A measurable shift is underway in how enterprises deploy AI inference: they are moving workloads off cloud GPUs and onto the edge silicon already sitting in their devices. The beneficiaries are Qualcomm, Apple, and MediaTek. The companies watching revenue trajectories shift are Amazon Web Services, Microsoft Azure, and Google Cloud.
The economic case is simple. Running a medium-scale language model inference task on AWS costs roughly $0.002 per query at current spot pricing. Running the same query on a Qualcomm Snapdragon X Elite laptop or an Apple M4 Pro chip costs nothing incremental — the hardware is already paid for. For organizations running tens of millions of queries per month, that difference is not a rounding error.
Qualcomm has been the most aggressive in quantifying this. The company's AI Hub, launched in late 2025, gives enterprise developers access to pre-optimized model variants of Llama, Mistral, and Phi tuned for Snapdragon Neural Processing Units. In Q1 2026, Qualcomm reported that AI Hub had more than 4,000 enterprise registrations, up from under 500 at launch. The company does not break out revenue from AI Hub directly, but the Snapdragon X series now commands a 15 to 22 percent ASP premium over its non-AI predecessors in enterprise procurement cycles.
Apple is pursuing a different angle. Apple Intelligence, the marketing name for on-device AI in iOS 18 and macOS Sequoia, expanded to 40 languages in late June 2026. The enterprise significance is less about individual productivity and more about what it does to IT policy: organizations that have blocked cloud AI tools for data residency reasons now have a compliant, on-device path.
MediaTek is the least discussed of the three but arguably the most important at scale. Its Dimensity 9400 and 9500 chips power a majority of Android flagship devices globally — in markets from Southeast Asia to Latin America where Qualcomm and Apple have lower penetration. MediaTek's AI processing benchmarks on the Dimensity 9500 are within 15 percent of Qualcomm's best, at a chipset cost roughly 30 percent lower.
The cloud vendors are not standing still. AWS announced custom Trainium 3 chips, Google continues to expand its TPU v5 fleet, and Microsoft is deploying Maia 2 silicon across Azure regions. But the inference cost gap at the edge is widening faster than cloud-side custom silicon can close it.