Two announcements landed within hours of each other in early September, and together they frame the next fight in AI: who defines consumer computing once models run on the desk instead of in the cloud. On 1 September Tim Cook stepped down as Apple chief executive to become executive chairman, handing the company to John Ternus, a mechanical engineer who had run hardware engineering since 2021. Four days later, at the IFA fair in Berlin, Nvidia pushed its RTX Spark desktop AI chips into the product lines of six PC makers.
Apple bets the product man can catch up on AI
Cook’s exit letter framed Ternus as someone who understands what building world-changing products costs. The choice is deliberate. Under Cook, iPhone revenue reached USD 209.6 billion in a financial year and services hit USD 109 billion, while Apple Watch, AirPods and Vision Pro all landed on his watch. Active devices passed 2.5 billion by early 2026. That record is built on packaging technology into experience through hardware design and supply chains.
AI is the one field where that genetic code misfires. Apple has never led in foundation models, its on-device model sits at roughly 3 billion parameters, and the contextual Siri shown at WWDC 2024 was delayed twice before arriving with iOS 27 in 2026. Where Siri cannot answer, Apple hands the request to Google’s models, and heavy users are steered toward paid iCloud+ tiers for more capacity. A company built on end-to-end control now depends on someone else’s models and someone else’s cloud.
Nvidia wants to sell the whole mine, not just the shovels
Nvidia’s RTX Spark chip, announced at Computex in June and now inside Lenovo, Asus, Dell, HP, Microsoft Surface and MSI machines shipping in October, is a declaration. It pairs a Blackwell GPU with 6,144 CUDA cores, fifth-generation Tensor cores, a 20-core Grace CPU and 128GB of unified memory, delivering around 1 petaflop of AI compute on an ARM design that leaves x86 out. Jensen Huang’s framing is that the machine stops being a tool and becomes a teammate, running a 120-billion-parameter model locally, always on, with no per-token meter running.
The deeper threat is the stack behind the chip: CUDA, TensorRT and thirty years of software, a Windows base hardened for on-device agents, and PAIR, an open-source framework released in September that spreads inference across every PC in a home. Nvidia’s move squeezes Qualcomm’s Snapdragon X AI PCs, Intel and AMD’s NPU pushes, and Apple’s on-device ambitions at once. On the day Nvidia’s entry became concrete, Qualcomm fell 8.78 per cent, Intel 4.67 per cent and AMD 1.16 per cent, while Nvidia rose about 6 per cent.
Who actually wins the desktop
The clearest investment signals sit in unremarkable layers. Unified memory and advanced packaging are the first, with 128GB of unified memory close to a requirement for local large models and HBM capacity already locked up by data-centre customers for 2026. The software stack is the second: CUDA’s accumulated ecosystem is harder to replicate than any chip. Device entry points are the third, and Apple, Lenovo, Asus and Dell all collect a toll there.
The biggest risk is usage, not shipment. Gartner expects AI PCs to be 54.7 per cent of the PC market in 2026, but shipped is not active. If the on-device agent ends up another never-invoked assistant, the hardware race becomes one more replacement cycle. For Apple the clock is tighter: Siri is already two years late, and if Ternus cannot make the on-device experience work within two years, ARM-based Windows machines can pinch Apple’s moat from both sides.
Editor’s note: This is an adapted translation of the original Sohu Tech report. It has been trimmed and restructured for readability for an international business audience. The full original (in Chinese) is at https://it.sohu.com/a/1072106856_116132.