NVIDIA GTC Taipei: Vera Rubin, N1X ARM, and the Era of AI Agents
🔎 Why GTC Taipei 2026 Changes Everything for AI
On June 1, 2026, Jensen Huang took the stage in Taipei for a keynote that redefines the roadmap for the entire AI industry. No teasing, no vague promises: full production of Vera Rubin confirmed, a new ARM chip for PCs unveiled, and a clear-cut stance on the era of autonomous agents.
Taiwan is not a random choice. Every Vera Rubin system contains 2 million parts, assembled by 150 Taiwanese partners. Huang transformed this keynote into an acknowledgment of debt to the ecosystem that makes NVIDIA possible.
Two structural announcements dominate: the Vera Rubin platform for data centers (16 ZFLOPS, NVL72) and the N1X ARM chip for PCs (20 ARM cores + Blackwell RTX 5070 GPU with 6144 CUDA cores, unified LPDDR5X memory). Around this hardware, a vision: AI agents are no longer a concept; they are the deployment model replacing the chatbot.
This article decodes the hardware announcements, their industrial impact, and what the "agent era" concretely means for developers and businesses.
The Essentials
- Vera Rubin in full production: data center platform at 16 ZFLOPS, NVL72 architecture with 7 co-designed chips, direct successor to Blackwell.
- N1 and N1X ARM for PCs: 20 ARM cores, Blackwell RTX 5070 GPU (6144 CUDA cores), unified LPDDR5X memory, targeting the autonomous "AI PC".
- The era of agents declared: Jensen Huang asserts that AI systems must observe, reason, plan, and act with minimal human intervention.
- Dependence on Taiwan: 2 million parts per Vera Rubin system, 150 Taiwanese partners involved in the assembly chain.
Recommended tools
| Tool | Main usage | Price (June 2026, check on site.com) | Ideal for |
|---|---|---|---|
| Hostinger | Web hosting to deploy AI agents | Starting from €2.99/month | Developers launching agentic services |
| Best autonomous AI agents | Comparison of agent frameworks | Free (article) | Choosing a framework before deployment on Vera Rubin |
| Local Ollama AI agents | Running agents locally | Free (open source) | Testing agents on PC N1X before cloud scale |
| LLMs for AI agents | Choosing models for agents | Variable | Selecting GPT-5.5 or Claude Opus 4.7 for agents |
Vera Rubin: 16 ZFLOPS and the end of the Blackwell era
Vera Rubin is in production. Not in preview, not in sampling: in full production. That is the central message of the keynote.
The platform reaches 16 ZFLOPS of performance, a significant leap over Blackwell. The NVL72 architecture — 72 GPUs interconnected via NVLink — is retained and optimized. But the real novelty is the number of co-designed chips: 7 distinct chips jointly designed to form a coherent system, rather than an assembly of generic components.
7 co-designed chips: why it's decisive
Up until Blackwell, NVIDIA primarily optimized the GPU and used standard components for the rest. Vera Rubin marks a turning point: networking, memory, power, cooling — every element is co-designed with the central GPU.
This level of integration explains the 16 ZFLOPS. Performance no longer comes solely from the chip, but from the absence of bottlenecks between components.
Comparison Vera Rubin vs Blackwell
| Feature | Blackwell (2024-2025) | Vera Rubin (2026) |
|---|---|---|
| Raw performance | ~4-5 PFLOPS per GPU | 16 ZFLOPS (NVL72 system) |
| Number of co-designed chips | 2-3 | 7 |
| Interconnect | 5th generation NVLink | Optimized Rubin NVLink |
| Primary target | LLM training and inference | Autonomous AI agents, heavy inference |
| Production | Maturity | Start of full production |
2 million parts and 150 Taiwanese partners
The figure is staggering. A single Vera Rubin NVL72 system contains 2 million individual parts. Huang insisted on this point: the system's complexity exceeds what a single company can master.
150 Taiwanese partners are involved in the assembly. TSMC for chip manufacturing, of course, but also dozens of suppliers for connectors, liquid cooling systems, high-density PCBs, power supplies. NVIDIA invests 150 billion dollars per year in Taiwan, and this keynote was also a demonstration of that commitment.
This has a clear strategic consequence: any attempt to off-shore the assembly chain is illusory in the short term. Vera Rubin is Taiwanese by design.
N1 and N1X ARM: The PC reinvented for AI
The second major announcement concerns PCs. NVIDIA unveils the N1 and N1X chips, ARM processors designed for personal computers. This is a direct offensive move against Apple Silicon and Qualcomm's Snapdragon X chips.
N1X Technical Specifications
The N1X integrates 20 ARM CPU cores, coupled with a Blackwell GPU derived from the RTX 5070 with 6144 CUDA cores. The critical point: unified LPDDR5X memory.
Unified memory means the CPU and GPU share the same address space. For AI, this is a game-changer. No more need to transfer data between system RAM and GPU VRAM — a historic bottleneck on PCs.
N1X vs. Competitors Comparison
| Feature | NVIDIA N1X | Apple M4 Ultra | Snapdragon X Elite |
|---|---|---|---|
| CPU Architecture | ARM (20 cores) | ARM (32 cores) | ARM (12 cores) |
| GPU | Blackwell (6144 CUDA) | Apple GPU (32 cores) | Adreno GPU |
| Memory | Unified LPDDR5X | Apple Unified Memory | LPDDR5X |
| AI Ecosystem | CUDA, TensorRT | MLX, Core ML | Hexagon NPU |
| Availability | Late 2026 | Already available | Already available |
Why the N1X is designed for agents
Huang did not present the N1X as a "fast for gaming" chip. The focus was on the local execution of AI agents. With 6144 CUDA cores and unified memory, the N1X can run models like Claude Sonnet 4.6 or GPT-5 locally without compromise.
This is consistent with the OpenClaw agent configuration which requires a substantial GPU runtime. An N1X PC becomes an autonomous agent node, not just a terminal.
The N1 (standard version, fewer CUDA cores) targets ultrabooks. The N1X targets workstations and "AI creator" PCs. Both share the same ARM architecture, which simplifies software development.
Windows on ARM: NVIDIA's bet
NVIDIA is betting on Windows on ARM. The N1X is presented as a Windows chip, not as a Linux alternative. This is a strong signal: NVIDIA believes Microsoft will succeed in Windows' ARM transition, and wants to be the premium hardware provider for this transition.
According to Fortune, Huang described this moment as the "reinvention of the PC," comparable to the shift from DOS to Windows. The argument is simple: current PCs are not designed for AI. The N1X is.
The era of agents: what Jensen Huang means exactly
"The era of agents" is the phrase that structures the entire keynote. But what does it actually refer to?
Huang defines an AI agent as a system that accomplishes four steps: observe, reason, plan, act. And above all, it does so with minimal human intervention. It is not a chatbot that answers a question. It is a system that receives an objective, analyzes the context, breaks it down into tasks, executes, and adapts.
From chatbots to autonomous systems
The distinction is fundamental. ChatGPT, even powered by GPT-5.5 (agentic score of 98.2 on benchmarks), remains a conversational system if it is not given action capabilities. An agent, on the other hand, has access to tools: APIs, files, databases, terminals.
This is where the 5 patterns of AI agents that work become relevant. The "Reflection" pattern (the agent critiques its own output), the "Planning" pattern (decomposition into subtasks), the "Tool Use" pattern (calling external APIs) — all require an infrastructure that Vera Rubin is designed to serve at scale.
Vera Rubin as agentic infrastructure
Why a 16 ZFLOPS data center for agents? Because agentic is a massively parallel inference problem. An agent that observes a real-time video stream, reasons on each frame, plans actions, and executes them — this requires continuous inference, not the batch processing of training.
Vera Rubin is optimized for this regime. The 7 co-designed chips reduce the latency between data acquisition (observation) and action generation (output). This is very low-latency inference, repeated millions of times per second.
The role of agentic models in this era
The best LLMs for AI agents are not the same as the best LLMs for conversation. The June 2025 agentic ranking shows a clear hierarchy: GPT-5.5 dominates (98.2), followed by Gemini 3 Pro Deep Think (95.4) and Claude Opus 4.7 Adaptive (94.3).
These scores measure the ability to plan, use tools, and maintain coherence over long chains of action. Claude Opus 4.7 Adaptive is particularly interesting: its "Adaptive" mode dynamically adjusts the depth of reasoning based on the complexity of the task — exactly what an agent needs.
An agent deployed on Vera Rubin could use GPT-5.5 for strategic planning and Claude Sonnet 4.6 (81.4) for routine subtasks, all orchestrated by a multi-agent framework. This is the architecture described by collaborative multi-agent systems.
Human intervention: minimal, not zero
Huang was precise: "minimal human intervention", not "zero intervention". Vera Rubin agents are designed to operate with light human supervision — validation of critical actions, adjustment of objectives, management of edge cases.
It is a pragmatic positioning that avoids the trap of AGI promised for tomorrow. Today's agents are advanced automation systems, not conscious entities.
Impact on data centers: what is actually changing
The arrival of Vera Rubin is not just a simple upgrade. It changes the very physics of data centers.
Power, cooling, density
16 ZFLOPS in an NVL72 rack implies unprecedented power consumption and heat generation. The 7 co-designed chips specifically include integrated liquid cooling solutions, suggesting that air cooling is insufficient at this density.
Existing data centers cannot simply "swap" Blackwell racks for Vera Rubin racks. The cooling infrastructure, power supply, connectivity — everything must be rethought. It is a 2-3 year construction cycle for operators.
The CUDA ecosystem as a moat
Despite competition from Grok 4.1 (xAI, score 79) and Kimi K2.6 Moonshot AI in self-host (88.1), NVIDIA is reinforcing its advantage through its ecosystem. CUDA, TensorRT, Triton — developers trained on these tools do not migrate easily.
Vera Rubin extends CUDA with new primitives for agentic: long context management, hardware-level multi-agent orchestration, priority inference scheduling. These are software features that rely on the 7 co-designed chips and have no equivalent among competitors.
Deployment cost: the math is brutal
A Vera Rubin NVL72 system with 2 million parts, assembled by 150 suppliers — the cost is astronomical. Initial deployments will go to hyperscalers (Microsoft, Google, Meta) and major research labs.
For smaller companies, the N1X offers an alternative: running agents locally with models like GPT-5 (78.1) or Claude Opus 4.6 (84.7) in self-host mode, via solutions like Ollama en local. The data center in the PC, in a way.
The scientific dimension: Vera Rubin and the observation of the sky
The name "Vera Rubin" is not chosen at random. It pays tribute to the astronomer who discovered evidence for the existence of dark matter. And the connection with science is deeper than a mere nod.
The Vera C. Rubin Observatory and AI
The Vera C. Rubin Observatory, whose scientific capabilities are documented in this study on the LSST impact, will generate 20 terabytes of data per night. Processing this stream in real time requires precisely the type of infrastructure that NVIDIA builds.
The observatory's preliminary data, published in Data Preview 1, already show the complexity of the processing required: detection of trans-Neptunian objects, tracking of interstellar comets like 3I/ATLAS observé par le Rubin Observatory, searching for dark matter via the Snowmass2021 program.
AI agents on Vera Rubin could automate the analysis of this data: an agent that observes the telescope stream, identifies anomalies, plans follow-up observations, and triggers alerts — exactly the "observe, reason, plan, act" pattern described by Huang.
Target fields and computational optimization
The study on the recommended target fields for commissioning of the Rubin Observatory illustrates another aspect: the computational optimization of observations. Choosing where to point the telescope, when, and for how long — that is a complex planning problem, exactly the type of task where an AI agent excels.
❌ Common mistakes
Mistake 1: Confusing Vera Rubin the chip and Vera Rubin the observatory
The Vera C. Rubin astronomical observatory and NVIDIA's GPU platform share the same name but have no direct technical link. The confusion is common in online comments. The observatory produces the data; the NVIDIA chip might one day process it, but they are two distinct entities.
Mistake 2: Thinking the N1X replaces an RTX 5090
The N1X integrates a GPU derived from the RTX 5070 (6144 CUDA cores), not a 5090. It is an SoC chip for AI PCs, not a dedicated graphics card. For pure gaming, an RTX 5090 remains superior. The N1X shines in scenarios where the CPU and GPU must collaborate closely via unified memory — typically local AI inference.
Mistake 3: Deploying agents without suitable infrastructure
The most costly mistake for companies: building agents with GPT-5.5 or Claude Opus 4.7, then trying to run them on existing Blackwell infrastructure hoping it "will work." Agentic requires continuous low-latency inference. Blackwell was optimized for training. Vera Rubin is optimized for agentic inference. It is not the same workload.
Mistake 4: Ignoring the Taiwan factor in planning
Planning a Vera Rubin deployment without factoring in the dependence on 150 Taiwanese partners means ignoring the central geopolitical risk. A single system contains 2 million parts. The supply chain is extraordinarily complex and geographically concentrated.
❓ Frequently Asked Questions
What is the difference between N1 and N1X?
The N1 is the standard version intended for ultrabooks, with fewer CUDA cores. The N1X is the high-performance version with a full Blackwell GPU (6144 CUDA cores) and more LPDDR5X memory, targeting AI workstations.
When will Vera Rubin systems be available?
Production is announced as "complete" by June 1, 2026. The first deployments at hyperscalers are expected in the second half of 2026. General availability for smaller companies should follow in 2027.
Can AI agents run locally on an N1X PC?
Yes, that is precisely the targeted use case. With 6144 CUDA cores and unified LPDDR5X memory, the N1X can run models like Claude Sonnet 4.6 or GPT-5 locally. Frameworks like Ollama already enable this type of deployment on less powerful hardware.
Are all 7 co-designed chips of Vera Rubin manufactured by TSMC?
NVIDIA has not detailed the exact manufacturing breakdown. TSMC certainly produces the main compute chips, but some co-designed chips (cooling, power, connectivity) likely come from other specialized Taiwanese suppliers.
Does the age of agents mean the end of developers?
No. Huang speaks of "minimal human intervention," not a total absence. Developers build the agents, define the guardrails, and validate critical actions. The profession is evolving toward the orchestration of autonomous systems, not disappearing.
✅ Conclusion
GTC Taipei 2026 marks a tipping point: NVIDIA is no longer just selling GPUs, but complete systems designed for the agentic era — from the Vera Rubin data center (16 ZFLOPS, 7 co-designed chips, 2 million parts) to the N1X ARM PC (unified memory, 6144 CUDA cores). Le硬件 is ready. What remains is building the agents that justify this infrastructure. For developers, the message is clear: train on the meilleurs agents IA autonomes now, because the hardware to deploy them at scale is arriving.