Apple hardware chief says new Macs carry no cost per token

By Splitfeed EditorialLast updated Sep 26, 2026, 20:18 UTC

On September 22, 2026, Apple introduced four new Mac models alongside a claim from its hardware chief that the machines carry "no cost per token" for AI inference. The phrase refers to the per-call pricing structure that cloud GPU services use when developers or users run AI workloads remotely. Apple's argument is that running inference locally on Apple silicon eliminates that recurring cost entirely, because the compute happens on hardware the customer already owns. Nvidia's stock moved in both directions across the following two days, initially gaining before slipping as markets worked through what the announcement meant for cloud GPU demand.

The framing matters because it targets the economics of a market Nvidia currently dominates. Cloud inference, where requests are routed to data centers running Nvidia GPUs and billed per call, has become a major revenue story for the chip industry. Apple is not arguing that its chips are faster or more capable in absolute terms; it is arguing that the cost model is different, and that for many workloads the on-device alternative removes an ongoing expense entirely. Whether that argument holds up depends on what kinds of inference tasks a Mac can handle locally compared to what developers actually need from cloud services, which is the genuine point of dispute here.

None of the three camps, Pro-AI, Anti-AI, and Middle Ground, had published specific reactions at the time of writing. Pro-AI voices would typically treat on-device inference as a democratizing development, lowering barriers for developers and individuals who cannot absorb cloud billing at scale. Anti-AI voices would likely question whether cheaper and more accessible inference simply accelerates deployment of systems they already consider underregulated or harmful. Middle Ground observers would be expected to probe the fine print, asking which model sizes actually run on Apple silicon, whether quality is comparable to cloud alternatives, and what happens when workloads outgrow local hardware.

The argument will sharpen as developers test the four new Mac models against real inference workloads and publish benchmarks comparing output quality and latency to cloud equivalents. Nvidia's next earnings commentary on data center demand will also be a signal worth watching, since executives there will likely address whether on-device compute is being treated internally as a meaningful competitive pressure or a separate market segment.