NVIDIA enters the PC.
NVIDIA shipped its first PC processor at Computex 2026. RTX Spark packs 70 billion transistors, 128 GB of unified memory, and 1 petaFLOP of AI compute into a 45-to-80 W laptop, and added $319 billion to NVIDIA's market cap in a single session. Qualcomm fell 8 %. Intel fell 5 %. Six OEMs (ASUS, Dell, HP, Lenovo, Microsoft and MSI) ship 30+ devices by Fall.
For thirty years the PC was an x86 duopoly. NVIDIA just walked in through the AI door, and the incumbents felt it the same afternoon.
Jensen Huang, NVIDIA founder & CEO. Photo: Peter Dasilva / Wikimedia Commons (CC BY 4.0)
NVIDIA (US) designs the GPU, owns the CUDA software moat, and integrates the chip.
MediaTek (Taiwan) co-designs the 20-core ARM CPU half of the SoC.
TSMC (Taiwan) fabricates the dies on its 3nm process.
Samsung / SK Hynix / Micron supply the 128 GB of unified LPDDR5X memory.
Microsoft provides Windows-on-ARM, the OS that makes the chip a mainstream PC.
ASUS, Dell, HP, Lenovo, Microsoft, MSI assemble 30+ finished laptop models.
1. NVIDIA and MediaTek co-design the SoC (GPU + ARM CPU on one die).
2. TSMC fabricates the dies on 3nm; yields set the supply ceiling.
3. Unified LPDDR5X memory is sourced and co-packaged beside the SoC.
4. Finished packages ship to the six OEMs.
5. OEMs build laptops and price them from $2,899 to $7,000.
6. Software runs on Windows-on-ARM plus NVIDIA's CUDA stack.
Neutral process view. Where a step is single-sourced (TSMC 3nm, NVIDIA CUDA) is noted without judgment.
Three things had to line up.
This didn't come from nowhere. First, Apple proved in 2020 that an ARM laptop with unified memory could beat x86 on performance-per-watt, then took half the US premium-laptop market while Intel watched. Second, the AI workload changed what fast means: the bottleneck is no longer clock speed, it's how much model you can hold in memory at once. Third, Qualcomm's ARM-exclusivity deal with Microsoft expired, opening Windows-on-ARM to anyone with the silicon.
NVIDIA had the GPU, the CUDA moat, and, via its MediaTek partnership, finally a CPU. RTX Spark is what happens when the company that already owns datacenter AI decides a laptop is just a smaller datacenter. The 128 GB unified pool isn't a spec-sheet flex; it's the whole point. CPU and GPU address one memory space, so a 120-billion-parameter model runs locally, with no copying, no cloud round-trip, and no data leaving the machine.
TSMC (every chip is a 3 nm order), MediaTek (co-designed the CPU, gains ARM credibility), Microsoft (Surface Laptop Ultra is the halo device), and the LPDDR5X memory makers (Samsung, SK Hynix, Micron), handed a new high-margin demand curve.
Intel, whose Lunar Lake and Arrow Lake-H still split CPU and GPU memory, exactly the bottleneck Spark removes. Qualcomm, whose Snapdragon X Elite can't match Spark's system-level compute at the workstation tier. And the discrete-GPU laptop model itself.
The startup opening. Someone needs the ARM-migration tooling for enterprises with x86-locked stacks. Someone needs the on-device LLM app layer that assumes 128 GB is there. And the advisor who audits a Fortune 500's refresh roadmap against this shift won't lack for clients. The infrastructure layer of physical AI is being rewritten, and the picks-and-shovels haven't been claimed yet.
NVIDIA edge-AI silicon on a reference board at Computex. Photo: 4300streetcar / Wikimedia Commons (CC BY 4.0)
1. NVIDIA RTX Spark press kit: specs, OEMs, OpenShell runtime [5 min]
2. Morgan Stanley: N1 / N1X pricing & ARM-share scenarios [8 min]
3. Tom's Hardware: Computex 2026 roundup [6 min]
4. Apple M1 unified-memory explainer (the original playbook) [4 min]
Unified memory: one pool of RAM both the CPU and GPU read directly, with no copying of data between them.
LPDDR5X: the low-power memory standard laptops use instead of desktop DDR5.
petaFLOP: one quadrillion math operations per second; a yardstick for AI compute.
ARM vs x86: two chip instruction sets. x86 (Intel/AMD) ran PCs for 30 years; ARM (phones, Apple Silicon) is far more power-efficient.
Beachhead: a small foothold a company seizes first, then expands from. NVIDIA doesn't need the whole PC market, just a place to stand.
NPU: neural processing unit, a chip block built specifically for AI, separate from the CPU and GPU.
Corrections & coffee: [email protected]
