NVIDIA GTC Taipei 2026 Keynote | Live

Technological concepts & product/strategy highlights

1) “Useful AI” and agentic AI as the new computing wave

• Jensen frames AI’s progression as: generative AI → useful AI → agentic AI.
• Key claim: agentic systems are arriving and can do productive work (not just generate text).
• Agents are described as an architecture that includes:
  • LLM(s) for “thinking”
  • a harness for orchestration (routing context → plan → tool calls → execution)
  • tools (e.g., browsers, databases, compilers, data engines)
  • memory (working memory / “KV caching” and long-term retrieval)
  • a runtime/security harness that makes the whole system operate safely.

2) Tokens as the economic driver

• “Tokens” are presented as profitable units that drive demand:
  • If AI systems generate more output, companies want to generate more tokens
  • This increases compute demand, which (per Jensen) is reflected in ecosystem activity in Taiwan.
• NVIDIA’s repeated metric focus: performance per watt = revenue per watt (compute as revenue).

3) New “computing pattern”: from apps to agents

• Traditional model: launch app → click/type → get output.
• New model: express intent, and the AI generates code or uses tools to produce the output.
• Breakthrough emphasized: tool use and tool orchestration, enabling agents to function like software coworkers.

4) NVIDIA CUDA-X libraries as “tools for agents”

• NVIDIA positions its library collection as an agent “toolbelt,” calling them CUDA-X.
• Examples named:
  • cuLitho (computational lithography)
  • cuOpt (decision optimization)
  • cuDSS (direct sparse solvers)
  • AI-Q (deep research across documents)
  • Aerial (AI RAN)
  • PhysicsNeMo (differentiable physics)
  • Parabricks (genomics)
• “Skills”/documentation concept: libraries ship with usage “skills” so agents can learn how to use them effectively.

5) Disaggregated + distributed agent runtime (rack/data center orchestration)

• Agents are described as an ultimate disaggregated/distributed computing model, where:
  • LLM context/logic runs on GPUs
  • tool execution can involve CPUs
  • security harness runs on a DPU (BlueField)
  • orchestration runs on CPUs
• Memory complexity is highlighted as a major challenge:
  • KV caching, compression, retrieval of structured vs unstructured data
  • “ontology/relationships” for retrieval

---

Major product announcements & infrastructure features

6) Vera Rubin: the next-generation agentic supercomputer (full production)

• Vera Rubin is positioned as:
  • not just a GPU system
  • a multi-rack, pod-scale agent processing system
• Claims include:
  • “Vera Rubin in full production”
  • supply chain expansion and faster assembly:
    • assembling a rack reduced from ~2 hours to 5 minutes (relative to Grace Blackwell assembly time per the narration)
• Architectural components described across a rack/pod:
  • Vera NVL72 GPU rack/tray for high-throughput token generation
  • Vera CPU rack (liquid-cooled) for orchestrating harnesses/tools and critical-path work
  • Vera BlueField-4 for storage processing and security/context memory
  • ConnectX-9 / SuperNICs and DPUs
  • liquid-cooled bus bars and modular tray design for scalability/resiliency
  • a specialized Ethernet switch concept (co-packaged optics mentioned)

7) Vera CPU: CPUs re-designed for agent latency/interaction

• Central message: older CPUs were built for “humans”; agents need nanosecond-level responsiveness.
• Vera is described with four design requirements:
  1. very high single-thread performance / IPC (instructions per clock)
  2. high bandwidth per core
  3. very high total bandwidth (including a “fabric” connecting cores)
  4. energy efficiency (so CPU capacity scales without stealing token-generator power)
• Specific implementation details mentioned:
  • Olympus core architecture for modern data center workloads (Python/tool/sandbox execution)
  • LPDDR5X with multi-error correction while maintaining bandwidth
  • scalable coherence fabric (monolithic mesh unifying cores)
  • “LPDDR5X vs x86” comparison claims:
    • ~40% lower peak memory latency vs x86 (per narration)
  • agent sandbox performance improvements (stated as 1.8× vs x86 in the narration)
• Real-world workload speedups cited:
  • SQL running 3× faster
  • real-time stream processing ~6× faster (example: NYSE telemetry)

8) DSX AI Factories: blueprint + simulation + operations stack

The video includes a long product narrative around NVIDIA’s DSX:

• DSX Sim (Omniverse blueprint / digital twin):

  • partners design/validate a Vera Rubin AI factory layout
  • simulate power/cooling, network design, integration testing, and changes in a digital environment

• DSX OS:

  • provisions, operates, monitors, remediates infrastructure
  • supports multi-tenant, resilient, AI-ready capacity

• Power/cooling/throughput features highlighted:

  • AI factories may overprovision power by up to 40% today; DSX MaxLPS aims to reduce waste.
  • DSX MaxLPS lets operators safely deploy more GPUs under same power budget
  • hot liquid cooling around 45°C (claimed water/energy efficiency)
  • dynamic power allocation to shift power rack-to-rack (“recover stranded watts”)
  • in-rack power smoothing to reduce spikes
  • DSX Flex:
    • cooperate with the grid using real-time grid signals

• Economic framing:

  • “Compute is revenue” and “lowest-cost tokens”
  • DSX aims at faster time to first token / inference / training and higher tokens per watt

• Hardware/software emphasis:

  • DSX is portrayed as an end-to-end co-designed stack for maximum profitability and reliability.

9) NVIDIA Agent Toolkit for Enterprise AI (Open models + harness + runtime)

• Jensen lists “4 things” for enterprise agents:

  1. Models
  2. Harness (orchestration)
  3. Tools/skills (e.g., CUDA-X libraries with agent-consumable skills)
  4. Runtime (secure “operating system” for agents)

• Key runtime/security product:

  • OpenShell:
    • sandboxing agents under enterprise security policies
    • privacy and identity protection emphasized
    • positioned as open source and being adopted (Red Hat, Canonical, Microsoft mentioned)

• Named components/harnesses:

  • Claude Code and Codex as agent tools
  • harness examples: OpenShell (runtime) and agent harnesses like OpenClaw, Hermes (mentioned as additional harnesses)

10) Chip design example: Cadence + NVIDIA agents + Nemotron

• A “design verification agent” is described as speeding up the RTL verification loop:

  • Codex orchestrates
  • Nemotron + secure environment using NVIDIA OpenShell
  • uses subagents for RTL generation, testbench creation, regression, debug
  • uses Cadence tools (e.g., Xcelium, JasperGold) in loops

• Performance claim:

  • verification cycles over 40× faster
  • “weeks to hours” (explicitly repeated)

11) Nemotron 3 Ultra (open model) for agent workflows

• Announcement of Nemotron 3 Ultra:
  • 5× faster and ~30% cheaper
  • hybrid architecture: SSM (State Space Models) + Mixture of Experts
  • claims: open model system includes training data and scripts (“open models at their best”)
• Mentions ongoing work toward Nemotron 4.

---

“Reviews/guides/tutorials” style content

• This keynote is not presented as a tutorial/review format (no typical product review scorecards), but it does include an “implementation blueprint” view:
  • DSX is effectively an operational guide: simulate → validate in digital twin → provision/operate with DSX OS.
  • The Agent Toolkit is a conceptual guide: models + harness + tools/skills + runtime (OpenShell).

---

Additional platform directions mentioned (AI beyond the data center)

RTX Spark: reinvention of the PC for agents

• RTX Spark: reinvention of the PC for agents (RTX + unified memory + Grace CPU + NVLink)
• Includes partner ecosystem (e.g., Microsoft/MediaTek)
• Agent-friendly updates mentioned:
  • Adobe Photoshop/Premiere
  • MCP server support

Physical AI

• Cosmos 3 (physical-world omni-model): world model/simulator/action-conditioned future prediction
• Alpamayo 2 Super (reasoning autonomous vehicle model): tied to DRIVE Hyperion + Halos OS
• Isaac GR00T humanoid robotics platform:
  • humanoid robot reference for research
  • simulation/teleoperation/data generation stack

---

Main speakers / sources

• Jensen Huang (NVIDIA founder & CEO) — primary speaker
• Janine — stage host/introducer (spoken line: “Welcome to the stage…”)
• Additional source elements: VIDEO NARRATION segments (DSX, Vera Rubin, Vera CPU details, Agent Toolkit, Cadence verification, Cosmos/Alpamayo/Isaac GR00T announcements)