NVIDIA's RTX Spark just landed: 20 ARM cores, 6,144 CUDA cores, 128GB unified memory, 100% performance on battery. Also ~273 GB/s of memory bandwidth, a 2-year-old Blackwell GPU tile, and OEM power budgets ranging 45W-95W because apparently we're doing TDP roulette again. Let me save you 45 minutes of review reading. IT WINS: • beats a portable RTX 4070 • crushes M5 Pro/Max at ray tracing • Tensor cores demolish Apple in ComfyUI • runs 70B models an 8GB laptop GPU can't even load • ~99-100% perf unplugged (x86 rivals lose half) IT LOSES: • Prism emulates x86 CPU while GPU runs native → CPU bottleneck caps your frame rate, and DLSS can't save you • Apple wins multi-thread and Premiere/DaVinci loops • 300 GB/s vs Apple's 400-800 GB/s. Loading a 70B model ≠ running one fast • £2,100-3,700 in UK. $2,400 in the US. Entry configs missing or stuck with 512GB. Now: didn't Qualcomm already fail at this? Yes. But Snapdragon X Elite didn't flop because ARM Windows is impossible. It flopped because you can't launch a platform on emulation and hope devs show up. Anti-cheat wouldn't port. Adobe was patchy. Hardware was fine, software was the moat. And to be fair to Qualcomm — it led multi-threaded, beat Intel/AMD on battery, and held clocks on battery when x86 dropped 50%. Not a flop. A foundation. Difference this time: NVIDIA brought CUDA. The biggest dev moat in computing. Fortnite and The Finals already work on ARM. CS2 and PUBG still don't. Verdict: not a gaming laptop. A local AI workstation that happens to run games. If you're the developer who's been on a MacBook for the unified memory and quietly missing CUDA — this is the first machine that stops making you choose. If you just want to play games, buy something else and keep £1,000. Rumour: NPU might handle DLSS frame gen entirely, which would kill the CPU bottleneck. Unconfirmed. Genuinely want to know if it's real. Sources below. 🔧