Up for Scraps

Executive briefing

The architecture of computing has flipped, and the consequences are now arriving at retail. The dominant volume product on every leading-edge process node is no longer a PC CPU — it is a smartphone SoC. Whichever product anchors the highest volume on a node sets the cost basis for everything else fabricated on it. The smartphone is the new desktop PC; the laptop is the new server. The structural advantage that made x86 unbeatable in the 1990s — desktop volume amortising the node — is now operating in favour of ARM, with the mobile handset as the volume anchor.

The MacBook Neo, Apple’s $599 ARM laptop launched in March 2026, is the visible end of this shift. The silicon inside it is a binned A18 Pro die — the same chip that ships in the iPhone 16 Pro, with one defective GPU core fused off. The R&D and fab capacity were paid for by the iPhone Pro programme; the MacBook Neo monetises the natural yield-tail, an estimated 12-15 million dies a year that would otherwise be scrap. That binned tail alone equals a fifth to a third of the entire global budget laptop CPU market.

The mechanism is structural, not tactical. No vendor competing with Apple in the budget Windows laptop segment has access to a primary-volume product on TSMC’s leading edge. Apple has paid for the capacity, the yield curve, and the priority allocation through fifteen years of iPhone wafer commitments. The Neo’s silicon line is structurally inaccessible to HP, Lenovo, Dell, or Asus. Hsu, Asus’s co-CEO, called the launch a “shock to the entire market” on his Q4 2025 earnings call.

The same logic operates one stack-level up. AWS now ships most new EC2 capacity on its in-house ARM-based Graviton silicon rather than on Intel or AMD x86; the chips business exceeded ten billion dollars in annualised run rate by Q4 2025. Nvidia has built ARM-based CPUs (Grace, Vera) optimised for AI training clusters and is designing custom microarchitectures (Olympus) rather than licensing Neoverse off the shelf. In July 2025, Nvidia disclosed it is also porting CUDA to RISC-V hosts, framing the move explicitly as preparation for a “compute landscape” reshaped by export controls. Four major actors — Apple at the consumer edge, AWS in the cloud, Nvidia in AI training, the EU/India/China at the sovereign tier — have independently arrived at the same architectural answer through different forcing functions: cost, workload fit, geopolitics. The convergence is not a fashion. It is what happens when the same volume mechanism is applied at every layer of the compute stack by actors with the resources to act on it.

Intel and AMD are caught in an institutional trap. Their x86 server margins fund the engineering programmes any pivot would require; their customer base wants compatibility maintained, not abandoned; their public communication strategy cannot publicly concede architectural disadvantage without accelerating the migration they are trying to slow. The trap closed in real time over an eighteen-month window: Intel published a draft x86 simplification (x86-S) in 2023, was pushed back by the ecosystem it depended on, terminated the programme in December 2024, and joined a defensive coalition with AMD. In September 2025 AMD publicly denied that ARM offered an efficiency advantage over x86. The trap is not stupidity. The trap is incentive structure.

ARM is the transitional answer, not the destination. ARM Ltd is itself a foreign-controlled licensor in the critical path of every ARM design — a dependency layer that became briefly visible during the 2020-2022 governance crisis at Arm China, when the JV CEO physically retained the company’s official seal and ran the operation independently for nearly two years. Sovereign-compute planners noticed. The European Union’s SiPearl Rhea-1 processor (eighty Neoverse cores, exascale system 2026) is an ARM design with a stated RISC-V successor roadmap. China’s T-Head has shipped over four billion XuanTie RISC-V cores; the server-grade C950 entered production in early 2026 with reported initial deliveries of 470,000 units. India’s Digital India RISC-V Programme funds an indigenous processor portfolio explicitly framed by the government as eliminating “license costs” and supporting “digital sovereignty.” Three independent blocs, three independent capital structures, the same architectural answer.

What is up for scraps is every part of the existing compute stack that depends on a foreign-controlled licensor: the cost structure of the budget Windows laptop, the margin pool of the x86 server market, the Arm Ltd royalty layer, and the geographic concentration of advanced-node fabrication. Several scenarios could slow this trajectory — Intel succeeding as a merchant foundry, ARM consolidating its licensee base, a US-China thaw, RISC-V’s software ecosystem stalling — but none resolves the structural argument; they only change the slope of the curve. The MacBook Neo is the visible part of a much larger sorting. The last instruction set will be the one nobody owns.