The New RTX Spark Has Changed the AI World Forever

At Computex 2026, NVIDIA CEO Jensen Huang announced the RTX Spark — a new superchip designed to run AI agents directly on your laptop or desktop, no cloud required. Built in partnership with Microsoft and MediaTek, the RTX Spark combines an NVIDIA Grace CPU (20 ARM cores) with a full Blackwell RTX GPU (6,144 CUDA cores) on a single chip, connected by NVIDIA's NVLink-C2C interconnect. It's manufactured on TSMC's 3nm EUV process — the same node as Apple's latest silicon.

The pitch from NVIDIA is simple: your PC is about to stop being a tool and start being a teammate. Every laptop and compact desktop running RTX Spark will be purpose-built for running AI agents locally, with full access to NVIDIA's AI stack — CUDA, TensorRT, and the full suite of developer tools that have powered AI in data centres for the past decade.

The headline number is 1 petaflop of AI performance. To put that in context: that's supercomputer-grade compute sitting inside a 14mm thin, 3-pound laptop. It's achieved through fifth-generation Tensor Cores running FP4 precision — a lower-precision format optimised specifically for inference, which is what you're doing when you run a model locally.

The memory story is just as significant. RTX Spark supports up to 128GB of LPDDR5X unified memory at 300 GB/s bandwidth. That's the same architecture Apple uses in its M-series chips — CPU and GPU sharing one memory pool — which eliminates the bottleneck of moving data between separate chips. 128GB means you can load a 120 billion parameter model entirely in memory and run it locally, with up to 1 million tokens of context. For reference, GPT-4 is estimated to be around 1.8 trillion parameters — so a 120B model is a serious, production-quality model, not a toy.

For AI developers and builders, the practical unlock is running large language models fully offline. A 120B parameter model with 1 million token context running locally means no API costs, no latency, no rate limits, and complete privacy. If you're building agentic workflows, automations, or AI-powered products, this changes the economics significantly — especially for use cases where you're making thousands of model calls per day.

Beyond LLMs, the RTX Spark handles creative workloads that were previously cloud-only: rendering 3D scenes larger than 90GB, editing 12K video, generating 4K AI video. And it can play AAA games at 1440p at over 100 frames per second. This isn't a trade-off machine — it's built to be the best at everything simultaneously. The first laptops will come from ASUS, Dell, HP, Lenovo, Microsoft Surface, and MSI, launching in fall 2026.

For the past few years, running serious AI locally meant buying a desktop with a discrete GPU, or accepting the limitations of cloud APIs. Apple's M-series chips changed the story for Mac users — but NVIDIA's CUDA ecosystem, which powers virtually all serious AI development tooling, stayed locked to data centre GPUs or desktop RTX cards. RTX Spark breaks that lock. For the first time, you'll be able to run the full NVIDIA AI stack — the same environment researchers use, the same tools developers build on — on a thin laptop.

The comparison to Apple's M5 is unavoidable, and it's clearly what NVIDIA is gunning for. On raw AI performance and memory bandwidth, RTX Spark is competitive. On the developer ecosystem, NVIDIA wins by a significant margin — CUDA has a 15-year head start. What Apple wins on is the integrated software experience and energy efficiency. We won't know the full picture until fall 2026 benchmarks are out. But the era of being forced to choose between the NVIDIA ecosystem and portable, efficient hardware is ending.