Quick Summary
- Confirmed pre-keynote: NemoClaw open-source agent platform launches Monday; hardware-agnostic, full source access
- Expected: Feynman AI reasoning chip details, Vera Rubin consumer timeline, and Jensen's "world-surprising chip" tease
- This article is a hub: deep-dive articles will be linked as they publish throughout March 16-17
GTC 2026 starts Monday, March 16. Jensen Huang's keynote runs at 11 AM PT, and based on pre-show filings, partner announcements, and NVIDIA's own teasers, this is shaping up as the most consequential GTC for local AI builders since the original H100 launch. We're covering it here in real time.
This article is structured as a hub. The top section covers what we know before the keynote. Sections below will be updated as announcements land on Monday.
What We Know Before the Keynote
NemoClaw: Open-Source Agents Land Monday
NVIDIA confirmed NemoClaw ahead of the show. It's an open-source enterprise AI agent platform — hardware-agnostic, full source code, no proprietary API dependency. Partners announced include Salesforce, Cisco, Google, Adobe, and CrowdStrike, which signals this is aimed squarely at enterprise automation workflows.
The hardware-agnostic positioning is significant. Unlike CUDA-only tooling, NemoClaw is confirmed to run on AMD, Apple Silicon, and third-party AI accelerators. For local AI builders, this matters: a major enterprise agent framework that doesn't require NVIDIA hardware is a direct validation of the "your own GPU rig runs production agents" thesis.
Full breakdown: NVIDIA NemoClaw: Run Enterprise AI Agents on Your Own GPU Rig
The Feynman Architecture
NVIDIA's roadmap — last publicly updated at Hot Chips 2025 — placed a chip codenamed Feynman after Rubin in the AI accelerator sequence. Pre-show speculation from semiconductor analysts positions Feynman as an AI reasoning-specific chip rather than a general-purpose GPU architecture. The framing: as AI shifts from pure token generation toward multi-step reasoning and agentic tasks (long chain-of-thought, tool use, reflection loops), the compute profile changes. Reasoning workloads are more serial, more branch-heavy, and benefit from different memory hierarchies than parallel transformer inference.
Feynman details will be updated after the March 16 keynote.
What this means for local builders depends entirely on whether Feynman has a consumer path. If it's data center only (like H100), the impact is indirect — it sets the ceiling that future RTX consumer chips target. If NVIDIA introduces a Feynman-derived consumer SKU, that changes the upgrade calculus for anyone holding off on a new GPU.
Vera Rubin: Consumer Timeline
Blackwell is in the market on the data center side (B100, B200) and landing in consumer form as the RTX 50 series. Vera Rubin is the architecture after Blackwell — named after the astronomer who pioneered evidence for dark matter, maintaining NVIDIA's tradition of naming accelerator architectures after scientists.
Pre-show reports suggest NVIDIA will detail Vera Rubin's timeline at GTC. The question for local builders: will the RTX 60 series based on Vera Rubin land in 2027 or 2028, and what VRAM configurations are expected?
Vera Rubin consumer timeline will be updated after the March 16 keynote.
Jensen's "World-Surprising Chip"
Jensen teased a "world-surprising chip" in pre-show media appearances. This kind of language from NVIDIA's CEO typically precedes one of three things: a new architecture announcement ahead of schedule, a compute density claim that rewrites performance-per-watt expectations, or a new product category entirely.
Given the Feynman and NemoClaw pre-announcements, the most plausible interpretation is either a Feynman early tape-out reveal or a new inference-specific chip class targeting the growing agentic workload market.
"World-surprising chip" details will be updated after the March 16 keynote.
Monday Live Updates
This section will be updated during and after the March 16 keynote. Check back after 11 AM PT.
Hardware Announcements
This section will be updated after the March 16 keynote.
Software and Platform Announcements
[UPDATE: will be filled after March 16 keynote — NemoClaw details, developer tooling, CUDA updates]
Local AI Impact Summary
[UPDATE: will be filled after March 16 keynote — specific callouts for what matters to local LLM builders]
What to Watch For
Before Monday, here are the specific questions we'll answer in the live coverage:
Does Vera Rubin have a consumer GPU path in 2026? If RTX 60 series is a 2026 product, anyone buying an RTX 50 series now should know. If it's 2027+, current GPU purchases are fully justified. For context on current RTX 50 series pricing implications, see RTX 5060 Ti pricing and what it means for builders.
What VRAM do Vera Rubin consumer GPUs offer? The 12-24GB ceiling on current RTX 50 consumer cards is the primary limitation for local LLM inference. Whether Vera Rubin breaks this ceiling — or whether we're waiting for another generation — is the most actionable question for builders.
Is NemoClaw actually runnable on consumer hardware? "Hardware-agnostic" can mean many things. We'll be looking at minimum VRAM requirements, whether it runs with quantized models, and how the setup process compares to Ollama.
Feynman: data center only or consumer path? Sets the 2027-2028 upgrade roadmap for anyone building today.
Related coverage: Tenstorrent QuietBox 2 — the alternative for builders who don't want to wait on NVIDIA's roadmap. For a pre-keynote deep-dive on NemoClaw specifically, see NVIDIA NemoClaw: Run Enterprise AI Agents on Your Own GPU Rig.
For context on how current RTX 50 series hardware performs for local LLM inference before the next architecture arrives, see our best 16GB GPU for local LLMs comparison. If you're evaluating whether to buy now or wait for Vera Rubin, the AMD vs NVIDIA for local LLMs guide covers the current ecosystem landscape that any new architecture will need to compete with.
This article was last updated March 13, 2026. Keynote coverage will be added beginning March 16 at approximately 12:30 PM PT.