Kael Zhang
NVIDIAGTCAgentic AIVera RubinRTX SparkCosmos 3

NVIDIA GTC Taipei 2026: Seven Launches Signal the Arrival of the Agentic AI Era

Kael Zhang

At 11:00 AM on June 1, Taipei Music Center.

NVIDIA CEO Jensen Huang stood on stage and delivered a two-hour keynote that the industry has called “the most important day for the PC industry in 40 years.”

There was only one core thesis: AI has moved from “generating content” to “getting work done,” and Agentic AI is the signature of this shift.

Seven flagship products were announced in a single event. This was not a product refresh—it was a systemic declaration of a new paradigm.


The Seven Launch List

No.ProductCore Positioning
1Vera RubinFully mass-produced rack-scale AI supercomputer
2Vera CPUNVIDIA’s first self-developed data center processor
3Nemotron 3 UltraWorld’s first open model with SSM+MoE hybrid architecture
4RTX Spark + N1XRedefining the PC ecosystem in partnership with Microsoft
5Cosmos 3Open-source foundation model for physical AI
6Alphamayo 2Open reasoning model for autonomous driving
7Isaac GR00THumanoid robot development platform

Core Thesis: Token Economics

Throughout the keynote, Huang anchored one logic: compute is revenue, and tokens per watt is profit margin.

He used a set of data to refute the “AI will reduce employment” narrative:

“That is complete nonsense. AI is not taking jobs—it is exponentially amplifying the output of every engineer.”


Vera Rubin: The Supercomputer Built for Agents

Vera Rubin has officially entered full-scale production. This is the most complex end-to-end rack-level system in NVIDIA history:

ComponentSpecification
GPUVera Rubin NVL72 (NVLink 72 interconnect)
CPUVera CPU (self-developed Olympus architecture)
NetworkConnectX-9 + world’s first 200Gb CPO optical Spectrum-X switch
SecurityBlueField-4 security processor (confidential computing standard)
StorageNew agent-native memory system architecture
Assembly TimeCompressed from 2 hours to 5 minutes

The supply chain is twice the scale of Grace Blackwell, involving 150 partners in Taiwan.


Vera CPU: NVIDIA Enters the CPU Market

This was the most strategically significant announcement of the entire keynote.

Huang’s argument: “The number of agents in the future will far exceed the number of humans, and they are extremely impatient with latency. Traditional CPUs designed for ‘humans’ are inherently ill-suited to nanosecond-level heterogeneous computing. This is a massive new market that NVIDIA has created out of thin air.”

Benchmark data:

Key architectural advantage: monolithic mesh interconnect eliminating chiplet overhead, with core bandwidth of 3.6 TB/s.


RTX Spark: Redefining the PC

Jointly developed by NVIDIA and Microsoft over three years, marking the first comprehensive PC revolution in 40 years.

SpecificationData
GPUBlackwell RTX, 6144 CUDA cores, 1 PFLOP AI compute
CPU20-core Grace (co-developed with MediaTek, codename N1X)
Memory128GB unified memory
ProcessTSMC 3nm, 70 billion transistors
InterconnectNVLink unified interconnect

Demo scenario: running the Hermes framework locally, connecting to cloud-based Claude Sonnet, AI autonomously invokes Rhino modeling, Blender rendering, and Flux 2 image generation—transforming a sketch into a professional 3D render, with full self-correction throughout the process.

Huang’s prediction: in ten years, home AI supercomputers could be as common as home theater systems.


Cosmos 3: The Foundation Model for Physical AI

The biggest pain point for physical AI is data—the vast majority of video is third-person perspective, while robots need first-person.

Cosmos 3’s solution: fusing autoregressive and diffusion Transformer architectures, unifying pixels, actions, sounds, and language into a single processing pipeline, directly generating synthetic video that conforms to real-world physical laws.

As the “virtual mentor” for robots, Cosmos 3 is fully open-source—model, data, and training methods all publicly available.

Built on Cosmos 3, NVIDIA also announced:


The Nature of Competitive Moats Has Shifted

Huang invited 11 leaders from the open-model ecosystem—including Mistral, Perplexity, Cursor, and Thinking Machines Lab—to join him on stage.

He offered a new perspective: “Proprietary and open are not opposing concepts. You can be both proprietary and open.”

Cursor CEO Michael Truell’s assessment was even more direct: the industry is witnessing the rise of a third category of company—neither pure foundation-model labs nor pure application builders, but new AI companies that integrate models, systems, and products into a single stack.

NVIDIA’s competitive moat is no longer just about single-GPU performance. It is the full-stack integration and extreme co-design capability spanning from chip to rack, from software to operations, from cloud to edge PC.


The Core Signal

Looking back at the entire keynote, there is only one core signal:

NVIDIA is no longer a GPU company. It is the irreplaceable infrastructure monopoly of the agent era.

When the cost of a single AI factory approaches the $100 billion scale, whoever helps customers deploy faster, operate more efficiently, and generate profit over longer cycles will win this era.


Sources