It was the kind of keynote that makes you stop scrolling. On the morning of June 1, 2026, Jensen Huang walked into the Taipei Music Center and delivered what may be the most sweeping technology keynote of the decade — two hours that functioned as a formal declaration: the company he built has become the infrastructure the AI economy runs on.
"Useful AI has arrived," Huang said early in the presentation — not soon, not eventually, arrived. The centerpiece of the keynote was the RTX Spark Superchip, a Windows-on-ARM platform co-developed with Microsoft and MediaTek, manufactured on TSMC's 3nm process. Its stated purpose: run frontier AI models entirely on-device, without sending a single byte to a server the user has never seen.
The market's reaction was immediate. NVIDIA climbed nearly 4% on announcement day. Qualcomm — whose Snapdragon X chips had dominated the Windows-on-ARM market — collapsed 8.78% in a single session, its steepest single-day drop in months. Intel and AMD both retreated as well. In financial terms, that's a real redistribution of confidence in who owns the next decade of personal computing.
What RTX Spark Actually Is
RTX Spark is a system-on-chip combining a 20-core NVIDIA Grace ARM CPU, a Blackwell GPU with 6,144 CUDA cores, and up to 128GB of unified LPDDR5X memory on a single package. CPU and GPU connect through an NVLink chip-to-chip interconnect at 600 GB/s, eliminating the memory bottleneck that has constrained every PC built over the last four decades.
- CPU: 20-core NVIDIA Grace (ARM)
- GPU: Blackwell RTX — 6,144 CUDA cores
- Tensor cores: 5th generation, FP4 precision
- Memory: up to 128GB unified LPDDR5X
- Memory bandwidth: up to 300 GB/s
- AI performance: 1 petaflop (FP4)
- Chip interconnect: NVLink-C2C at 600 GB/s
- Process node: TSMC 3nm, 70 billion transistors
- Max local model size: roughly 120 billion parameters
- Launch: second half of 2026
The 128GB of unified memory — shared between CPU and GPU in a single high-speed pool — means a developer can load a 120-billion-parameter language model entirely on-device, a capability that normally requires tens of thousands of dollars in cloud compute to run at production latency. Conventional PCs have separate CPU and GPU memory banks and must copy data between them; the NVLink interconnect at 600 GB/s removes that obstacle, so CPU, GPU, and memory behave as a single system.
Vera Rubin and the Age of Agents
RTX Spark was, in some ways, the consumer hook for a much larger strategic arc. The real weight of the keynote was in what Huang said about data centers and about the nature of AI itself. The Vera Rubin platform — the next-generation multi-rack system for agentic AI factories — is already in full production, built from six co-designed chips including a new standalone Vera CPU that runs 1.8 times faster than x86 for agentic workloads.
Early customers include Anthropic, OpenAI, and SpaceX, with Microsoft Azure already running an engineering sample and mass shipments beginning in the second half of 2026. The supply chain involves more than 150 Taiwanese partner companies, and accumulated demand visibility reportedly exceeds $1 trillion.
Alongside the hardware, Huang presented Nemotron 3 Ultra, an open-weights AI model with 550 billion parameters that leads the American open-weights rankings, plus Cosmos 3 — the first open physical-AI model for robotics — and a partnership with Cadence to speed up chip-design verification. The message: NVIDIA is no longer a chip company. It now controls data center infrastructure, consumer PC chips, agent runtimes, foundation models, robotics simulation, and the supply chain that manufactures all of it.
Where This Technology Will Arrive
Hardware announcements can be seductive in the abstract. The more useful question is what changes, concretely, and for whom:
- Healthcare: a notebook running a 100-billion-parameter model locally can analyze imaging and cross-reference drug interactions without sending a single patient record to an external server — for hospitals under HIPAA, GDPR, or CCPA, on-device inference may be the only legally viable path to scaling AI.
- Law and finance: firms dealing with information too sensitive for cloud-based AI can now run contract review, document analysis, and financial modeling locally.
- Creative industries: NVIDIA claims RTX Spark renders 3D projects exceeding 90GB, edits 12K video, and generates 4K AI clips — in a notebook, ending the old trade-off between mobility and professional-grade power.
- Education and research: a local machine that runs 120-billion-parameter models at zero marginal cost per token can meaningfully redistribute who gets to do frontier AI research, beyond whoever can afford cloud compute.
- Small business: one RTX Spark machine can replicate capabilities previously reserved for large teams — drafting documents, managing calendars, summarizing email, writing code — with no variable cost and full data control.
Timeline: What to Expect and When
- RTX Spark 128GB (top configuration): full local inference on models up to roughly 120 billion parameters
- RTX Spark 64GB (base configuration): quantized models up to roughly 60 billion parameters
- RTX 5090 (32GB VRAM), for comparison: limited to roughly 30 billion parameters
- A typical current laptop GPU (8–16GB), for comparison: small models only, roughly 7 billion parameters
Near term (2026–2027): the first RTX Spark notebooks and desktops reach retail in the second half of 2026 through Dell, HP, Lenovo, Microsoft Surface, ASUS, and MSI, at premium starting prices above roughly $2,500. Windows-on-ARM compatibility with legacy software will create some friction at launch, and regulatory scrutiny of NVIDIA's market position is intensifying in the US and Europe.
Medium term (2027–2030): a next-generation Vera Rubin Spark with faster memory launches in 2028, pushing pricing toward mainstream as Apple, AMD, and Qualcomm all accelerate proprietary silicon specifically to reduce dependency on NVIDIA. AI agents start becoming a standard interface for everyday computing tasks, and autonomous robotics begin entering logistics, manufacturing, and retail at scale.
Long term (2030 and beyond): the next hardware generation handles models that today require entire data centers, "AI agent as primary computer user" becomes normal, and access to powerful local AI becomes one of the defining divides of the era — which is exactly why the two arguments below matter.
The Case For: Privacy and Local Power
For years, AI assistance came with a hidden tax: your data traveled through servers you cannot audit, belonging to companies whose interests do not always align with yours. RTX Spark changes that at the hardware level — with 128GB of unified memory and 1 petaflop of on-device compute, a 120-billion-parameter model can run entirely locally, and sensitive information never has to leave the machine.
NVIDIA's new OpenShell runtime adds a policy layer that lets users define exactly what their AI agents can access and audit agent actions in real time. Running AI locally also eliminates latency, removes monthly cloud-inference subscription costs, and brings capable AI to places with poor connectivity — rural hospitals, field researchers, small businesses outside broadband reach. For professionals operating under HIPAA, GDPR, or CCPA, that combination of open model weights, local hardware, and a policy-enforced runtime is compliance built into the architecture rather than promised in a contract.
The Case Against: One Company, the Entire Stack
Step back from the petaflops and look at the shape of the announcement. A single company now designs the data center chips that power cloud AI, a CPU competing with Intel and AMD, the consumer PC chip, the leading open-weights foundation models, the robotics simulation platform, the agent runtime, and the supply chain — with over 150 partner companies in Taiwan dependent on NVIDIA's continued success. That isn't diversification; it's a moat wide enough to look like a wall, and the CUDA ecosystem alone carries 18 years of advantage no competitor can easily replicate.
The legal signals are already flashing: China's market regulator has accused NVIDIA of antitrust violations, and Intel has publicly warned about compatibility and DRM issues with Windows on ARM. Amazon, Meta, Alphabet, and Microsoft are all accelerating their own proprietary silicon for exactly this reason. There's also the question of who gets left behind — RTX Spark devices arrive first from premium manufacturers at premium prices, and 2026 may be the first year in three decades without a new consumer GeForce GPU generation, as manufacturing capacity shifts to AI chips.
The Bottom Line
June 1, 2026 marked a genuine inflection point in personal computing history. RTX Spark has a confirmed supply chain, confirmed OEM partners, and a confirmed launch timeline — this is not hype without substance, and the Vera Rubin platform is already running in production, not sitting on a roadmap slide. Whether this technology ends up serving people broadly or concentrating power narrowly is the more important question, and the answer won't come from a keynote.
Frequently Asked Questions
- What is NVIDIA RTX Spark? A superchip developed with Microsoft and MediaTek on TSMC's 3nm process, combining a 20-core ARM CPU with a Blackwell GPU and up to 128GB of unified memory to deliver 1 petaflop of AI performance, built to run frontier-capable AI agents entirely on-device.
- When does RTX Spark launch? Second half of 2026, through Dell, HP, Lenovo, ASUS, MSI, and Microsoft Surface, with base configurations starting above $2,500 and top-spec builds exceeding $4,000.
- Does RTX Spark run games? Yes — its Blackwell GPU is roughly equivalent to a desktop RTX 5070 and supports ray tracing, DLSS, and G-SYNC, though NVIDIA positions the chip primarily for AI workloads.
- What AI models run on RTX Spark? Any model in the CUDA ecosystem, via Ollama, llama.cpp, vLLM, or SGLang — it's the reference hardware for the 120-billion-parameter Nemotron 3 Super, with the larger Nemotron 3 Ultra available via cloud access or server-class configurations.