"The Arm PC copied the Mac twice: the instruction set, and the pool. Its marketing leads with a third number the Mac's decode never divides by."
The Second Convergence: the Instruction Set
The CPU track told the story of Apple moving the Mac to its own Arm cores in 2020. The Windows PC followed on a longer fuse: Qualcomm's Snapdragon X Elite in 2024 was the first Arm chip sold at scale into Windows laptops with a translation layer for the x86 software the industry had accumulated — Rosetta's problem, four years later, at the other vendor. Its product brief lists LPDDR5x at 8,448 MT/s, "Up to 64 GB", "135 GB/s", and an NPU of "45 TOPS"; that is a 128-bit bus at the M4's class of bandwidth, with the M4's class of ceiling — a 9B at 30 tokens per second, a 27B not worth running. The 2026 X2 Elite raises the memory to 128 GB and the top part, the X2 Elite Extreme, widens the bus to 192 bits for "228 GB/s" and an 80-TOPS NPU; the lower X2 SKUs keep 128 bits at 152. Both are pools; both are the Mac's diagram; both sit at a fraction of the Max's bus width, a step narrower than AMD's 256 bits.
The Number the Marketing Leads With
Every Arm PC launch leads with TOPS — trillions of NPU operations per second — and the number is real, and it is the wrong number for a language model's decode. The Core ML lesson measured why on Apple's own unit: a neural engine is a compute engine with a narrow door to the pool, and decode is a bandwidth stage that reads every weight per token. The X2 Elite Extreme's 80 TOPS decodes a 9B at the same 51 tokens-per-second ceiling that a GPU on the same 228 GB/s would, because the ceiling is bytes over bandwidth and TOPS never appear in it. Sort the code block's table by the bandwidth column and the Arm PCs land between the M4's row and the M4 Pro's (the Extreme's 228 against 120 and 273); sort it by TOPS and they lead the M4's Neural Engine, and the ordering means nothing for the token. Apple's own M4 newsroom line — the Neural Engine "capable of up to 38 trillion operations per second" — is the same kind of claim from the other side, and this quest treats both as vendor metrics for the compute stage, which is where they belong.
The Datum That Changes the Track
The press reports a third Arm PC vendor for 2026: NVIDIA, with a chip the press calls RTX Spark, "the consumer-oriented sibling of the GB10" — an Arm CPU and a Blackwell GPU behind NVLink-C2C with "up to 128GB of unified memory", built for Windows laptops and compact desktops from the major OEMs, reported to be arriving as Qualcomm's exclusive arrangement for Windows on Arm lapses. This quest has measured none of it and labels it press. If it ships as described, it is the strongest datum in the track: the Mac's shape — Arm cores, a large GPU, a 128 GB pool on a 256-bit bus — with the CUDA platform from the rival track attached, sold into the PC market at the same bandwidth point as the DGX Spark. The copy that matters was never a wider LPDDR bus; it is the pool with twenty years of software on it, and it is the reason the next lesson asks what copying proves about the mouse.