Tech • IA • Crypto
Official Keynote Closing Video | GTC Taipei 2026
9/10TL;DR
Nvidia used Computex to unveil a sweeping push toward AI “agents,” new chips and networking, and infrastructure aimed at faster, cheaper inference at massive scale.
KEY POINTS
AI agents move to the center
The company framed the next phase of AI as “useful” agents that can plan, code, and act on behalf of users and businesses. It emphasized that individuals and small teams can now build companies from home, but that scaling these agents requires significantly more compute and efficient inference.
New silicon and faster inference
A new processor line, highlighted alongside Rubin, was positioned as purpose-built for agents, promising around 50% CPU performance gains and up to 10× faster inference with lower token costs. The pitch focused on reducing latency and operating expenses for real-time AI applications.
Networking and memory with BlueField and NVLink
BlueField data processing units were presented as critical for maintaining agent “memory” and data flow across systems. NVLink Fusion extends high-speed interconnects to custom ASICs, aiming to unify heterogeneous compute and let partners join a broader NVLink ecosystem.
Software stack: NeMo and guardrails
NeMo and NeMo Guardrails were positioned as core to building, aligning, and governing AI systems, with NeMo “Tron Ultra” cited for substantial productivity gains. Sandboxing and security layers were emphasized to keep autonomous systems controlled and auditable.
End-to-end AI “five-layer” stack
Nvidia described an integrated stack spanning models, orchestration, infrastructure, networking, and applications. The goal is to convert AI capability directly into revenue by streamlining deployment from development to production.
Data centers at gigawatt scale
The company pointed to global AI clouds operating at multi‑gigawatt levels, with systems designed to optimize every watt. Efficiency claims focused on performance per watt improvements and tighter integration across compute, networking, and storage.
PC and edge AI push
Consumer and enterprise PCs powered by RTX 40 GPUs were highlighted as a major platform shift, enabling local model execution alongside CPU-based orchestration. The approach blends on-device inference with cloud acceleration.
Robotics and synthetic data
Tools such as Cosmos and advanced simulation were presented as ways to train robots and autonomous systems using synthetic worlds. Emphasis was placed on perception, reasoning, and learning skills that translate from simulation to real environments.
CONCLUSION
Nvidia’s announcements underscore a strategy to own the full AI stack, from chips to software, while accelerating the shift toward autonomous agents and large-scale, energy-efficient AI infrastructure.