Nvidia to unveil AI inference chips, new CPU at GTC 2026

Nvidia will unveil new AI inference chips and a new CPU product at GTC 2026 in San Jose from March 16 to 19.

The company is expected to detail performance metrics and production timelines for Vera Rubin, an accelerator focused on inference rather than training.

Nvidia is also set to announce a new CPU aimed at managing and processing data for agent-based workloads.

Dion Harris, Nvidia’s head of AI infrastructure, told CNBC that CPUs are becoming a bottleneck for scaling AI agent tasks.

GTC will also showcase physical AI use cases such as robotics and factory automation.

Reports say Samsung Electronics and SK Hynix may supply HBM4 memory for Vera Rubin.

🔗 Source: Chosun Daily

🧠 Food for thought

Implications, context, and why it matters.

Nvidia’s new platform is a fully integrated, six-chip architecture that can be built into rack-scale systems

“Vera Rubin” refers to the NVIDIA Rubin platform, which bundles six co-designed chips that run together as an AI-supercomputer architecture 1.
The NVIDIA Vera CPU uses 88 custom “Olympus” Arm cores, handling orchestration and data movement for AI factory workloads 2.
Rubin can split inference work across chips. The main Rubin GPU pairs with High Bandwidth Memory (HBM4), while the Rubin CPX GPU variant uses GDDR7 for the initial compute-heavy “context” stage of long-context inference 3.
Other parts include NVIDIA NVLink 6 Switch (a high-speed chip-to-chip interconnect switch for linking accelerators), NVIDIA ConnectX-9 SuperNIC (a high-performance network interface card for AI data centers), NVIDIA BlueField-4 DPU (a data processing unit that offloads networking and security tasks from the CPU) and NVIDIA Spectrum-6 Ethernet Switch (a data-center network switch) 1.

Nvidia’s one-year cycle pressures customers and reshapes the memory market

Annual launches, with Rubin close behind Blackwell, can turn data-center gear into “disposable infrastructure” and squeeze customer capital expenditure plans 4.
That schedule raises the stakes across the supply chain, especially for High Bandwidth Memory (HBM) 4.
TrendForce (a market research firm covering the electronics supply chain) cites reports saying NVIDIA wants HBM4 speeds above 10Gb/s, which exceeds the 8Gb/s JEDEC standard. The same reports expect Samsung Electronics and SK hynix as suppliers, while Micron may provide HBM4 for mid-tier accelerators such as Rubin CPX rather than the flagship Vera Rubin 5.
The faster cadence targets custom silicon from cloud providers, yet it may slow enterprise adoption when hardware can age out within a year 4.