NVIDIA

Confidence 0.80 · 2 sources · last confirmed 2026-09-04

GPU and AI-accelerator maker founded 1993; led by founder and CEO Jensen Huang. Entered the wiki as a central-subject first appearance via Huang’s interview with Harrison Chase (LangChain YouTube channel, 8 Jul 2026), following the same first-appearance-as-central-subject precedent used for Khan Academy and PwC. NVIDIA had previously surfaced only in passing — named alongside AMD as general-purpose silicon in Hoffman — before this substantive first-party source.

Why NVIDIA matters in this wiki

NVIDIA’s pitch, in Huang’s words, is that “most companies will be built on harnesses” rather than business processes, and that every company’s durable value is its own specialized, proprietary intelligence — which it must own and improve, not outsource. NVIDIA’s role in that world is supplying the open substrate (models + runtime) enterprises specialize on top of, rather than owning the specialization itself. This positions NVIDIA as an infrastructure/substrate vendor in the wiki’s agent-harness vocabulary — adjacent to, but distinct from, LangChain’s harness/framework layer.

NVIDIA Research, and a tension inside the company (added 2026-09-04)

NVIDIA enters the wiki a second time from a different altitude: NVIDIA Research (with Georgia Tech) authored [[2025-06-02-belcak-nvidia-small-language-models-future-agentic-ai|Small Language Models are the Future of Agentic AI]] — the reference statement of the SLM position, and the paper carried to practitioners in Sokolenko’s PyCon DE talk. Authors: Peter Belcak, Greg Heinrich, Shizhe Diao, Yonggan Fu, Xin Dong, Saurav Muralidharan, Yingyan Celine Lin (Georgia Tech), Pavlo Molchanov.

Held next to Huang’s sovereign-AI argument, the two documents pull in opposite rhetorical directions and both are NVIDIA speaking. Research argues the industry over-provisions frontier LLMs for agentic work that a fine-tuned sub-10B model handles, and that the centralised-inference build-out (~USD 57bn) persists largely as inertia — its barrier B1. The CEO argues sovereign owners should build and own frontier-scale intelligence.

They are not strictly incompatible: one is about which model serves each invocation, the other about who owns the stack. And the commercial reading reconciles them — NVIDIA sells the accelerators for centralised frontier inference and the consumer/edge silicon that runs SLMs locally, so “more inference in more places” is the position both documents serve. Worth stating plainly whenever either is cited: neither is a disinterested source on how much compute agentic work needs.

Products referenced in this wiki

• Nemotron (Nemotron 3 Ultra) — NVIDIA’s open-weight large language model family. On an internal Deep Agents benchmark, Nemotron 3 Ultra scores 86% vs. Claude Opus’s 87%, at roughly 10x lower cost than Opus — cited by Huang as evidence that open-weight models are reaching frontier performance at a fraction of the cost. Improved not just by scale but by harness-side tuning (prompts, tools) and, prospectively, by post-training the model inside the LangChain harness.
• OpenShell — a secure, open agent runtime; the deployment layer of the Deep Agents + OpenShell blueprint, providing sandboxing and access control so enterprise IT organizations can safely run agents.
• The Deep Agents + OpenShell blueprint — a joint NVIDIA + LangChain announcement (8 Jul 2026): running LangChain Deep Agents with Nemotron 3 Ultra inside OpenShell, packaging model + harness + runtime + acceleration stack as a reusable enterprise starting point for building domain-specific “super agents.”
• DGX Spark / DGX station — NVIDIA hardware named as deployment targets for enterprise agent systems, alongside cloud and on-prem options.
• Nemotron Coalition — named as a founding-team collaboration between NVIDIA and LangChain on Nemotron Ultra; not yet substantively detailed in the wiki.
• Nemotron-H (2 / 4.8 / 9B) and Hymba-1.5B — the small end of the family, cited in NVIDIA Research’s own SLM paper as evidence that hybrid Mamba–Transformer architectures reach 30B-class instruction-following and code generation at an order-of-magnitude fewer inference FLOPs.
• NVIDIA Dynamo — the distributed inference framework the SLM paper names as what reduces its barrier B1 (sunk investment in centralised LLM serving) “to a mere effect of inertia” by supporting high-throughput low-latency SLM inference in cloud and edge.

People

• Jensen Huang — Founder and CEO. Central subject of the interview but dangling (single-source, deferred) per the person-entity second-source promotion rule — central-subject status on a first appearance does not itself trigger person promotion (precedent: Sal Khan on the Khan Academy source). Promote on a second substantive Huang-authored or Huang-centric source.

Concepts NVIDIA touches in this wiki

• agent-harness — substrate/infrastructure vendor supplying the open model + runtime layers that harnesses like LangChain Deep Agents wrap.
• enterprise-ai-adoption — the “companies built on harnesses, not business processes” thesis and the specialize-after-frontier decision rule.
• dynamic-capabilities — digital-transforming/redesigning-internal-structures and digital-seizing/rapid-prototyping per the source page’s W&W tags.
• small-language-models — NVIDIA Research is the origin of the position; see the tension noted above.
• foundation-models — the device-indexed definition of “small” that the SLM paper contributes to the page’s size axis.

Open questions

• NVIDIA’s own harness/runtime engineering practices beyond the LangChain partnership — the wiki holds only this one vendor-collaboration vantage.
• Nemotron Coalition — named but not substantively detailed; an open ingest target if NVIDIA or LangChain publish more about the collaboration structure.
• Jensen Huang’s independent voice — this source is a fireside interview co-framed by Chase; a Huang solo keynote or NVIDIA-first-party essay would be a useful independent second source for promoting Huang himself.