Nvidia prices a 120-billion-parameter inference box into the $3,000 laptop bracket

Apple's is the only memory figure in the account: up to 192 GB of unified memory in the M5 generation ^[4], which set against 120 billion parameters ^[13] works out to 1.6 bytes per parameter ^[11], and on a unified-memory design the million-token window ^[13] has to draw on that same pool ^[3]. Nvidia's equivalent number is absent from cryptobriefing's account of the Computex Taipei launch ^[1][6], and it is the figure that decides whether running a 120B model locally describes a machine somebody works on or a slide somebody screenshots.

So price carries the argument instead. The midpoint of the $3,000 to $4,000 band ^[2] is $3,500 ^[9], and a machine bought there and held three years runs about $97 a month ^[10], which is the figure a team lead compares, informally and usually badly, with what the same developer bills in API calls. The comparison crosses a budget boundary: capex signed once against metered spend that renews without a meeting. That asymmetry, more than the silicon, is what moves a refresh plan.

Worth naming what Nvidia gives up to make the case. The revenue base is still data center ^[5], so Arm silicon and laptop-level system software are engineering hours not spent on the business that pays for them, and the report frames the whole exercise as diversification rather than a core shift ^[5]. The offsetting asset is CUDA, a standard for over a decade, which lets a developer already writing CUDA against data-center GPUs run the same code at the desk, continuity Apple cannot offer ^[7]. Nvidia itself never named Apple; analysts supplied that framing ^[14].

The narrow version is the one the piece's own analysts flag, where $3,000 to $4,000 buys a specialty workstation and the relocated slice of workload never shows up in anyone's cloud invoice ^[15]. The software version has developers staying on Macs because the integrated stack works and Nvidia's system-level optimisation is not there yet ^[16]. The version both vendors say out loud is hybrid, local for some tasks and cloud for others ^[17], which reads less like a compromise than a description of the iteration loop: prototyping and long-context grinding at the desk, serving in the rack.

This is probably wrong, but the workload that actually leaves is the one nobody meters carefully, the developer's own loop, and it leaves for accounting reasons rather than token-cost ones. Nvidia's stated bet is broader, that workloads now confined to server farms migrate to personal machines ^[18]; I would take the narrower version, or rather the more interesting narrow version, in which the cloud keeps production inference and loses the experimentation that never had a line item. What would falsify it is checkable: Spark ships with a memory capacity too small for 120B at any usable context, or the $3,000 to $4,000 band ^[2] holds long enough that no fleet buyer standardises on it, in which case Apple's on-device privacy pitch ^[8] keeps the dev laptop and Nvidia keeps the data center. One spec sheet settles most of that.