Official Keynote Closing Video | GTC Taipei 2026
The keynote frames a concrete problem: agent-based AI is moving from novelty to production but demands much more compute, memory correctness, and safety. The presented approach is an integrated stack (Vera/Rubin token optimization, BlueField memory, a faster agent-focused CPU, NVLink Fusion for ASIC interconnect, Nemo Tron Ultra, Nemo Clock and Open Shell guardrails, ESX power optimizations, and GPU/CPU partitioning) that targets inference throughput, memory fidelity, sandboxing, and power efficiency. Evidence in the talk consists of headline performance claims ("cheapest tokens," 10x faster inference, 5x faster work, 50% faster CPU) and feature claims (memory correctness, guardrails, sandboxing, automated code migration), but the transcript does not include benchmark methods or detailed validation. Outcome: the stack is presented as production-ready for large-scale agent workloads and global AI clouds, while the talk does not provide the underlying evaluation details or failure-mode analysis needed for engineering validation.
Key technical points
- Agents have shifted from niche/Movie-like examples to mainstream teams building with them, creating large new compute demands.
- Vera is introduced as a core component for agents, claimed to deliver the "cheapest tokens" for inference cost reduction.
- A claim of 10x faster inference is associated with this new token-optimized approach (presented as a headline performance metric).
- BlueField is described as maintaining agents' memory fidelity—"keeps agents' memory true"—indicating a system-level memory or state-consistency role.
- A CPU improvement of "50% faster" was announced and explicitly described as being built for agents (presented as agent-oriented CPU performance uplift).
- NVLink Fusion is presented as an interconnect/aggregation feature that "blends ASICs smartly," implying heterogeneous ASICs can be interconnected and used together.
- "Everyone's welcome to the NVLink party" implies broad compatibility or openness for multi-ASIC configurations via NVLink Fusion.
- Nemo Tron Ultra is presented as a leading compute element for agent workloads with a claim of 5x faster work throughput.
- Nemo Clock is described as the component that keeps guardrails correct (a runtime/monitoring guardrail mechanism), while Open Shell is described as providing a tight sandbox for execution isolation.
- The keynote claims automated developer operations: "Your code migrated and reviewed all before this song is through," implying automated migration and code review pipelines integrated into the stack.
- The AI stack is described metaphorically as a "five-layer cake for your revenue," indicating a layered architecture spanning hardware, runtimes, and cloud services.
- Global AI clouds are called out as the deployment target, requiring very large power capacity ("lots of gigawatts").
- ESX is presented as a power-management and orchestration element that keeps power lean and optimizes every watt across the system.
- A platform message: RTX 40 availability and the statement "Running anywhere Windows go" indicates desktop/PC availability; compute partitioning guidance is explicit: harnesses run on CPU and models run on GPU.
- Cosmos is framed as building virtual worlds needed by robots and systems that learn to move and reason—implying support for embodied/robotics workloads and learned perception/reasoning stacks.
For engineers this transcript highlights a few actionable system-design signals without experimental detail: prioritize agent-oriented architectures that separate harness/control (CPU) from model execution (GPU); design for stateful agent memory with mechanisms similar to BlueField to preserve memory fidelity; plan for heterogeneous-ASIC topologies and NVLink-style high-speed interconnects for fused workloads; include runtime guardrails and sandboxing (Nemo Clock, Open Shell) as first-class components; expect integrated CI/automation for fast code migration and review; and optimize for power per watt (ESX) when planning for cloud-scale deployments. Note that the talk provides headline performance claims (10x inference, 5x throughput, 50% CPU) and qualitative feature claims but does not include benchmark methodology, validation details, or explicit failure modes, so engineering evaluation and validation must be performed independently before production adoption.
For engineering teams in Japan: the stack emphasizes energy-efficiency and large-scale power planning (ESX + "gigawatts" clouds), which intersects with local data-center design and energy constraints; NVLink Fusion and multi-ASIC support matter for procuring compatible hardware and integrating heterogeneous accelerators in existing facilities; agent memory fidelity (BlueField) and runtime guardrails/sandboxing (Nemo Clock, Open Shell) are critical for deploying autonomous or robotics-related workloads that Japan teams commonly build (Cosmos/robot worlds mention); and the CPU/GPU partitioning guidance (harnesses on CPU, models on GPU) should inform workload placement, benchmarking, and cost modeling. Because the keynote provides performance claims without technical validation details, Japanese engineering teams should plan reproducible benchmarks and validation against local regulatory, energy, and operational requirements before large-scale rollouts.