NVIDIA at Computex 2026: Rubin GPUs, AI PCs & the New Robotics Stack

NVIDIA just wrapped its Computex 2026 keynote in Taipei, and it was the densest hour of hardware news we have had all year. Jensen Huang walked on stage and, in under sixty minutes, reset four product categories: data-center GPUs, AI PCs, desktop developer workstations, and humanoid robotics. Here is the plain-English breakdown of what NVIDIA announced, what is actually new, and what each piece means if you build with AI — without the keynote hype.

1. Rubin and Rubin Ultra: the Blackwell successor is here

The headline is the new Rubin GPU architecture, named after astronomer Vera Rubin and first teased a year ago. NVIDIA confirmed two SKUs: Rubin for general training and inference, and Rubin Ultra with twice the high-bandwidth memory for frontier-model training runs. Both use TSMC's N3P node and ship with HBM4.

Numbers Jensen quoted on stage:

• Up to 3.3× training throughput per rack versus Blackwell B200, mostly from improved interconnect and a redesigned tensor core.
• 1.6 TB of HBM4 per Rubin Ultra board.
• New NVLink 6 fabric at 3.6 TB/s per GPU.
• Vera CPU — NVIDIA's first in-house Arm-based server CPU, paired 1:1 with Rubin in the new GB300 superchip.

For most developers this matters indirectly: every Rubin shipped means cheaper inference on the API providers you already use. If you are tracking the broader race — see our recap of the biggest week in AI history for why hyperscaler GPU supply is the bottleneck holding back agentic apps.

2. The AI PC reset: RTX 60-series and Windows 12 Copilot+

The "AI PC" category has been a marketing label looking for a product for two years. Computex 2026 is the first time it has real teeth. NVIDIA's new GeForce RTX 60-series Mobile GPUs deliver 1,000+ TOPS of on-device AI compute — enough to run a 13B-parameter language model locally at usable speeds, alongside Stable Diffusion XL Turbo and real-time video generation.

Three things changed at once to make this work:

• Windows 12 ships with Copilot+ APIs that any app can call — local LLM, vision, transcription — without bundling its own runtime.
• OEMs (ASUS, Dell, Lenovo, Razer, MSI) are launching "AI Studio" laptops starting at $1,299 with the RTX 6050.
• NVIDIA's RTX AI Toolkit 2 auto-quantizes and packages models so they "just run" across the RTX 60 lineup.

If you build creator tools or developer apps, this is the moment to ship local-AI features as a default, not a power-user toggle. For the current landscape of local-first AI, our Ollama setup guide still applies — and gets dramatically faster on RTX 60.

3. DGX Spark: a real developer desktop

The most surprising product of the keynote was DGX Spark mini: a $3,999 desktop the size of a Mac mini with 128 GB of unified memory, a Rubin-class GPU, and the same CUDA stack as the data-center DGX. NVIDIA's pitch is simple — run, fine-tune, and prototype 70B-class models on your desk without paying cloud rent.

Why this matters:

• For solo developers and small teams, $4K up-front beats $0.50/hour H100 instances within a few weeks of serious use.
• Fine-tuning Llama 4 70B locally with LoRA goes from "rented multi-GPU cluster" to "overnight job at home."
• It is a credible answer to Apple's M-series Mac Studio for AI workloads — and ships with the entire CUDA ecosystem on day one.

Developers comparing coding stacks should also read our 2026 ranking of AI coding assistants — most of them get noticeably better with local model fall-back enabled.

4. GR00T N2: open humanoid robotics, finally usable

NVIDIA first announced Project GR00T at GTC 2024 as a vision for general humanoid robotics. GR00T N2, unveiled at Computex 2026, is the version teams can actually use. It is an open foundation model for bipedal robots, trained on a mixture of real teleoperation data, simulation rollouts, and internet video.

The stack ships as three pieces:

• GR00T N2 model weights — open under a permissive license for research and commercial use up to a revenue cap.
• Isaac Sim 5 — physics simulation with photoreal rendering for training and evaluation.
• Jetson Thor T2 — the on-robot compute module with 1,500 TOPS in a 60 W envelope.

Boston Dynamics, Figure, Agility, Apptronik, Sanctuary AI, and Unitree all showed robots running GR00T N2 on stage doing factory-floor and home tasks. This is the same direction we covered in our AI agents tipping point analysis — agents that act in the physical world, not just the browser.

5. NIM 2 microservices and the enterprise inference platform

Less flashy but arguably the biggest enterprise news: NIM 2 microservices. These are NVIDIA-packaged, OpenAI-compatible API containers for every major open model — Llama 4, Qwen 3, Mistral Large 3, DeepSeek R2, Gemma 3 — that drop into any Kubernetes cluster, on-prem datacenter, or VPC.

For enterprises, this is the cleanest answer yet to "how do we run open models in production without building MLOps from scratch." It also locks the open-model ecosystem more tightly to CUDA — a strategic moat NVIDIA is happy to widen.

6. What was missing

A few notable absences:

• No new consumer desktop RTX 60-series GPUs — those are expected at CES 2027, not Computex.
• No major automotive update — DRIVE Thor 2 was mentioned only briefly.
• No comment on the OpenAI / Broadcom custom-silicon deal that has been the elephant in the room since March.

7. What this means for the next 12 months

Three predictions worth writing down:

1. Local-first AI features ship by default. Once RTX 60 laptops are mainstream by Q4, "send to cloud" becomes opt-in, not assumed.
2. Frontier training costs drop. Rubin Ultra plus NVLink 6 should compress GPT-5-class training from $500M to closer to $150M per run — opening the door for one or two new frontier labs in 2027.
3. Humanoid robotics gets its "ChatGPT moment." GR00T N2 plus Jetson Thor T2 plus Isaac Sim 5 is the first time the full stack is available to a startup with under twenty engineers.

If you are evaluating which models to build on top of right now, start with our ChatGPT vs Claude vs Gemini 2026 comparison and complete AI tools guide.

Frequently asked questions

When will Rubin GPUs be available on AWS, Azure and Google Cloud?

NVIDIA confirmed shipments to hyperscalers begin in Q3 2026, with general availability on the major clouds in late 2026 or early 2027. Expect early access programs first.

Do I need an AI PC to run local LLMs?

No — any modern laptop with an RTX 40-series or M-series Apple chip can run 7B–13B models today. RTX 60-series simply makes 30B+ models comfortable on a laptop. See our Ollama guide for the entry-level path.

Is DGX Spark mini worth $3,999 for a solo developer?

If you fine-tune models or run inference for more than ~80 hours a month, the math works in under a year versus cloud GPU rentals. For occasional experimentation, a cloud instance is still cheaper.

Is GR00T N2 really open source?

The model weights are released under a permissive license that allows commercial use up to a revenue threshold. Above that threshold, a commercial agreement with NVIDIA is required — similar to Llama's community license.

Will Rubin make ChatGPT and Gemini cheaper?

Indirectly, yes. Cheaper, faster GPUs lower the cost of inference for every API provider. Expect token pricing on frontier models to drop another 30–50% in the 12 months after Rubin ships at scale.

Conclusion

Computex 2026 was not a single blockbuster announcement — it was four, stacked. Rubin reshapes data centers. RTX 60 reshapes laptops. DGX Spark reshapes developer desks. GR00T N2 reshapes robotics. Taken together, NVIDIA is not just selling chips — it is selling the entire compute layer of the AI economy, from a humanoid robot's hand to a hyperscaler's rack. The companies that move first to ship local-AI features, fine-tune on-desk, and prototype with GR00T are the ones who will be unrecognizable by Computex 2027.

Authoritative references

For deeper research and primary documentation on the topics covered above, these are the sources we cross-checked while writing this guide:

• SEC EDGAR filings
• Reuters Technology
• Stanford AI Index Report