Video summary

【直播｜AI即時中字翻譯】輝達 NVIDIA GTC Taipei 2026 黃仁勳主題演講

Main summary

Key takeaways

Technology

Tech-focused summary (NVIDIA GTC Taiwan 2026 keynote)

1) “Useful/agentic AI” arrives → AI becomes a profit/GDP driver

  • The keynote frames a shift from earlier generative AI to agentic AI—systems that observe, reason, plan, act, and use tools.
  • Jensen argues useful AI is now measurable, citing:
    • GitHub commit growth as evidence that coding activity is accelerating—attributed to agents increasing productive software output rather than directly replacing software engineers.
  • Core economic claim: AI compute demand rises because tokens become economically valuable
    • “Tokens are profitable units of revenue.”

2) New computing pattern: agents + harnesses + tools + runtime

  • Agent architecture
    • LLM for reasoning/thinking
    • A harness to orchestrate the agent loop (tool routing, memory management, context handling, etc.)
    • Tools (browsers, databases, data engines, spreadsheet-like workflows, CAD, and more)
    • Memory, split into:
      • Working memory (short-term)
      • Long-term memory
    • Additional emphasis on KV caching, and the complexity of memory retrieval/ontology
  • The orchestration layer is described as the “operating system” for agentic systems.

3) CUDA-X libraries as “tools for agents”

  • NVIDIA positions CUDA-X compute libraries as agent-operable building blocks.
  • Mentioned examples include:
    • cuLitho (computational lithography)
    • COOP (optimization)
    • cuDSS (sparse solvers)
    • AIQ (document research)
    • Aerial (AI RAN)
    • Warp (differentiable physics)
    • Parabricks (genomics)
  • Agents can use these more effectively with added “skills”—tool invocation guidance (manual-like instructions for how the AI should call the tool).

4) “Vera Rubin” system: multi-rack podscale agentic supercomputer (in full production)

  • NVIDIA introduces Vera Rubin specifically for agentic AI (not just inference).
  • Architecture themes:
    • Disaggregated + distributed + heterogeneous execution across model/harness/tools/runtime
  • Multi-rack/pod scale components:
    • Vera GPUs with MVLink72 for token generation throughput
    • Vera CPUs for orchestration, tool access, and memory/KV cache management
    • BlueField-4 DPUs for security/processing offload
    • High-bandwidth storage/networking components
  • Confidential computing focus:
    • Encryption at rest, in motion, and in use
  • Readiness claim:
    • “Vera Rubin is in full production,” with emphasis on Taiwan supply-chain scale.
  • Additional hardware notes:
    • Modular trays/switch trays
    • Liquid cooling (hot-liquid cooling mentioned)
    • Integration described as reducing external cabling (“no cables/hoses/fans” via internal PCB/midplane approach)

5) “Vera CPU”: CPUs redesigned for the agentic loop (low latency, high single-thread performance)

  • NVIDIA contrasts traditional CPU optimization (human workloads, throughput, virtualization, cloud “rent by the hour”) with agent needs.
  • In agent systems, CPUs bottleneck latency and tool responsiveness.
  • Key Vera CPU requirements:
    • Extremely strong single-threaded performance (IPC / instructions-per-clock)
    • World-class bandwidth (per core and total)
    • Interconnect/fabric at “speed of light” scale
    • Energy efficiency to enable many cores without starving token generation
  • Messaging/performance:
    • Claimed speedups on real workloads (e.g., SQL acceleration; real-time stream processing using NYSE telemetry)
  • Ecosystem/readiness:
    • Transition risk from x86 to Vera framed as lower because CUDA/stack integration is already in place and “qualified” in NVIDIA’s ecosystem.

6) “DSX” AI factory blueprint: shifting from building chips to building AI factories

  • NVIDIA positions compute as the economics engine:
    • Compute = revenues
    • Performance per watt = revenues
  • DSX is presented as an end-to-end AI factory stack:
    1. DSX SIM: digital twin / Omniverse-based simulation for power, cooling, and network validation pre-deployment
    2. DSX OSS: operations/provisioning/monitoring/remediation for multi-tenant resilient capacity
    3. DSX Max LPS: higher GPU deployment under power budgets; liquid cooling at 45°C; improved power utilization
    4. DSX Flex: grid-interactive operation using real-time grid signals
  • Quantitative claims:
    • Existing factories overprovision power by up to 40%
    • Max LPS reduces wasted power and increases GPU deployment
  • Core promise:
    • Faster time-to-first-token/inference/training, higher tokens-per-watt, and improved reliability/lifetime via integrated design + simulation.

7) AI clouds ecosystem examples (customers using NVIDIA infrastructure)

  • NVIDIA lists AI cloud providers and notable customers (including examples such as CoreWeave, Nebius, Nscale, plus regional providers across Korea, India, Singapore, Indonesia, Taiwan).
  • Distinction emphasized:
    • Companies can “stand up an AI cloud,” but differentiation increasingly comes from AI factory infrastructure (provisioning, reliability, throughput, cost).

8) Agent toolkits for enterprises: “models + harnesses + tools/skills + runtime”

  • NVIDIA announces an NVIDIA agentic toolkit for enterprise AI, requiring:
    1. Models (open models; positioned as “smarter/cheaper/faster”)
    2. Harness for orchestration
    3. Tools with skills (including CUDA-X as tools)
    4. Runtime / OS-like layer for sandboxing and secure execution
  • Runtime/security:
    • NVIDIA OpenShell: enterprise sandbox runtime with policy grounding, privacy, and identity/permission controls
    • Adoption claimed by major software/platform companies (e.g., Red Hat, Canonical, Microsoft mentioned)
  • Harness example:
    • OpenHermes (agentic harness/routing/orchestration concept)

9) “Neotron 3 Ultra” and open model direction (agent intelligence)

  • Neotron 3 Ultra is announced as an open model:
    • Hybrid architecture: SSM (state space) + Mixture of Experts
    • Claims:
      • 5× faster
      • ~30% cheaper vs “most cost-effective” comparable open models
    • Openness emphasis:
      • Model, training data, and training scripts/data provided to enable adaptation.

10) Physical AI + robotics direction: Cosmos 3, AlphaMo/AlphaDrive, and Isaac Groot

  • Cosmos 3 (physical AI foundation model for robots)
    • Open frontier model for perception → planning → simulation → policy assistance
    • Hybrid world-model approach that can generate synthetic data/video to close the training loop
  • Alpaca/AlphaMo “2” reasoning autonomous vehicles
    • Positioned as an open model enabling multi-vendor AV stacks via NVIDIA Hyperion runtime + Halos OS
  • Robotics platform:
    • Isaac Groot: reference humanoid robot plus Isaac robotics software stack (simulation, teleop, training, deployment) to reduce research setup time for universities/labs.

11) “New PC” for the age of agents: RTX Spark + Windows platform for agents

  • NVIDIA and Microsoft are positioned as reinventing the PC after ~40 years.
  • RTX Spark agent PC concept:
    • Agent runs continuously in a sandbox locally, while models/tools may come from the cloud
    • Developers can build secure agent workflows with tool access (example walkthrough: house-design agent using Rhino → Blender + photoreal generation)
  • Third-party app enablement noted:
    • Adobe positioning: Photoshop/Premiere “agent-friendly” via MCP server
    • “Twice as fast” style claim mentioned.

Product/initiative list (as described)

  • Vera Rubin: multi-rack podscale agentic supercomputer (in full production)
  • Vera CPU: agent-optimized low-latency CPU for orchestration + KV/memory pressure
  • MVLink72: interconnect targeting GPU-side token generation throughput
  • BlueField-4 DPU: security and offload in the agent stack
  • DSX (AI factory stack):
    • DSX SIM (digital twin via Omniverse)
    • DSX OSS (operations)
    • DSX Max LPS (power + liquid cooling optimization)
    • DSX Flex (grid-aware energy flexibility)
  • CUDA-X + “skills” for agent tool use
  • NVIDIA Agent Toolkit for Enterprise AI:
    • OpenShell runtime (secure sandbox)
    • OpenHermes harness
    • CUDA-X tools/skills + models
  • Neotron 3 Ultra (open hybrid model)
  • Cosmos 3 (physical AI world model)
  • Isaac Groot (reference humanoid robot + robotics dev platform)
  • RTX Spark / agent PCs + Microsoft “Windows platform for agents”

Main speakers/sources

  • Jensen Huang (Jensen Wong / NVIDIA founder & CEO) — primary speaker for most of the keynote.
  • Additional showcased segments/voicelines include NVIDIA/partner presenters (e.g., DSX presented as a hosted segment within the broadcast), but Jensen remains the dominant source.

Original video