Nvidia's Computex 2026: Jensen Huang Declares War on Every Layer of Computing

Jensen Huang walked into the Taipei Music Center and, in two hours, announced six data center chips, a PC superchip, a 500-billion-parameter open AI model, a humanoid robot reference design, and the death of the traditional PC. Nvidia’s Computex 2026 keynote was not a product launch. It was a declaration that the $5.23 trillion company intends to own every layer of the AI economy, from the power grid to the application.

The sheer density of announcements obscures the real story. This is not Nvidia expanding its product line. This is Nvidia building a vertically integrated operating system for the AI era, and doing it before any competitor can assemble the pieces.

Six chips, one data center.

The Vera Rubin platform is the architectural leap Nvidia needed after Blackwell. Six new chips — the Rubin GPU (336 billion transistors, dual-die, TSMC 3nm), the Vera CPU (Arm-based), the NVLink 6 switch, the ConnectX-9 SuperNIC, the BlueField-4 DPU, and the Spectrum-6 Ethernet switch — form a complete data center compute plane. Performance: 50 petaflops FP4 per NVL72 rack, with 288GB HBM4 per GPU and 260TB/s scale-up bandwidth. The Rubin Ultra, arriving in 2027, doubles that to 100 petaflops. Against Blackwell, Nvidia claims 5x inference performance and 10x lower cost per token.

The first Vera Rubin rack is already running at Microsoft Azure. AWS, Google Cloud, and Oracle are confirmed. Full production ships H2 2026. That is a staggering timeline for a chip of this complexity.

Inference gets its own chip.

Alongside the training-focused Rubin, Nvidia unveiled Rubin CPX — purpose-built for massive-context inference. The specs: 128GB GDDR7, 30 petaflops NVFP4, integrated video encoder/decoder for generative video. The NVL144 CPX platform delivers 8 exaflops AI compute and 100TB fast memory per rack.

This is the most strategically significant announcement of the keynote. Nvidia is acknowledging that inference — not training — is where the volume market is heading. Rubin CPX is designed for always-on AI agents processing million-token context windows. It is Nvidia building its own inference ASIC before customers build theirs. Every hyperscaler is working on custom inference silicon. Nvidia just told them: we saw that coming, and we are already shipping.

The PC is dead. Long live the AI PC.

“40 years of traditional PCs is now at an end.” That was Jensen’s most provocative claim, accompanying RTX Spark — Nvidia’s first PC superchip. A 20-core Grace CPU (co-developed with MediaTek) plus a Blackwell RTX GPU with 6,144 CUDA cores. Up to 128GB LPDDR5X unified memory. 1 petaFLOP AI performance. NVLink-C2C at 600GB/s.

Dell, HP, Lenovo, Microsoft, Asus, and MSI will ship devices by holiday 2026. This is a direct shot at Apple’s M5, Qualcomm’s Snapdragon, and Intel’s entire client business. Nvidia promises to “turn Windows into an agentic AI OS.” The bet is that the PC’s future is not a general-purpose compute device but a local AI inference appliance that happens to run Windows.

Software is the moat multiplier.

Nvidia released Nemotron 3 Ultra, a 500-billion-parameter open model designed for complex reasoning and agentic workflows. Unlike general-purpose models, Nemotron 3 Ultra is optimized for multi-step task execution with minimal human oversight — the kind of model that would run on Rubin CPX racks or RTX Spark laptops.

Alongside it: NemoClaw, a streamlined blueprint for building agentic workflows, and DSX (AI factory framework) with DSX MaxLPS delivering 40% more GPUs within the same power budget. DSX OS is open-source.

The software stack matters because it is the glue that locks customers into Nvidia hardware. CUDA’s 18-year head start is one moat. Adding an open foundation model, an agent framework, and a factory operating system creates three more. Nvidia is not just selling shovels in the gold rush. It is selling the maps, the claim stakes, and the assay equipment.

Physical AI: the 2030 revenue.

The least-discussed but potentially most significant announcements came in physical AI. GR00T N2, Nvidia’s next-generation vision-language-action model for humanoid robots, ranks #1 on both MolmoSpaces and RoboArena benchmarks with 2x the success rate of leading competitors. It ships end of 2026.

Nvidia also unveiled Isaac GR00T Reference Humanoid Robot (Unitree hardware plus Sharpa hands plus Jetson AGX Thor compute), Cosmos 3 (world simulation model), and Cosmos Reason (contextual understanding for robot navigation). Plus Alpamayo 1.5, a reasoning model for autonomous vehicles with multi-camera support.

This is Jensen’s long game. Data center GPUs are today’s revenue. Physical AI — where every robot, every autonomous vehicle, every industrial system runs Nvidia silicon and software — is the 2030 revenue. The reference robot is a signal: Nvidia is not waiting for customers to figure out the hardware. It will show them how.

Nvidia is not just selling shovels in the gold rush. It is selling the maps, the claim stakes, and the assay equipment.

The gaming footnote.

Almost lost in the AI announcements: DLSS 4.5 with 2nd-generation Ray Reconstruction arrives August 2026. Larger transformer model, better training data, same performance impact. 11 new supported games including Gothic 1 Remake and Phantom Blade Zero.

Gaming is now reported under “Edge Computing” in Nvidia’s financials. A reclassification that tells you exactly how Jensen views it: important, profitable, but no longer the strategic center of gravity.

What this means for AI builders.

The question is no longer whether Nvidia can sustain this. With $81.6 billion in quarterly revenue and 75% gross margins, the machine is self-funding. Every major cloud provider is already committed to Vera Rubin. The question is whether any competitor — or any combination of competitors — can build an alternative stack before Nvidia’s lead becomes permanently structural.

Google has models and cloud but no client chips or robotics hardware. Apple has client chips and an ecosystem but no data center AI or robotics. Microsoft has cloud and software but designs zero silicon. AMD has GPUs and CPUs but no software stack, no models, no robotics.

Nvidia at Computex 2026 demonstrated that the AI economy rewards vertical integration across every layer. The company is building the operating system of the AI era, from the silicon to the software to the robots that run on it. For AI builders, the strategic implication is uncomfortable: you can bet on Nvidia’s stack, or you can bet against it. There is no third option that exists yet.