NNuggets
BookmarksCollections
  • About Us
  • Terms of use
  • Privacy policy
  • Disclaimer
  • Copyright & Takedown Policy
  • Community Guidelines
  • Cookie Policy
  • Contact

© 2026 Nuggets

NuggetsMarket PulseCollections

On this page

Speakers & Credentials

  • Speakers & Credentials
  • 1. Executive Summary [00:00:07]
  • 2. Chronological Table of Contents
  • 3. Detailed Thematic Summary
  • The Reference Vault
  • 4. Data & Figures
  • 5. Core Frameworks & Mental Models
  • 6. Anecdotes
  • 7. References & Recommendations

On this page

  • Speakers & Credentials
  • 1. Executive Summary [00:00:07]
  • 2. Chronological Table of Contents
  • 3. Detailed Thematic Summary
  • The Reference Vault
  • 4. Data & Figures
  • 5. Core Frameworks & Mental Models
  • 6. Anecdotes
  • 7. References & Recommendations
Technology/March 27, 2026/10 min read/youtu.be

Jensen Huang talks with model builders on Nvidia GTC 2026 [Full Session]

Source
Source
Watch on YouTube ↗

"AI is not the model it's the system it's the computer... models are essentially becoming just tools like file systems and connectors." - Aravind Srinivas [00:11:43]

"The amount of computing use in pre-training was like 90% of training 2, 3, 5 years ago but in the future the amount of training percentage in pre-training is probably going to be tiny it's going to be mostly post-training." - Jensen Huang [00:18:50]

References

  1. Original source (youtu.be)

Disclaimer: Orignal content owned by or sourced from third parties. It does not represent the views of 'Nuggets' platform or it's team. AI is used extensively across this platform including for summaries. Accuracy is not guaranteed, there can be mistakes. Any info or content on this platform is not a financial, legal, or investment advice. Do your own research. Refer for complete disclosures:- Terms of Use · Full Disclaimer

Related nuggets

Jun 2, 2026

AI Is Escaping the Screen | 01 Jun 2026 | Coatue

Coatue : AI is entering a new phase: moving beyond digital tools and into fully autonomous systems operating in the physical world. From advanced manufacturing and surgical robotics to robots in the home, the next wave of innovation will b…

Jun 2, 2026

Kalshi Monthly Volume - Politics ($M) | Chart of the Day | Coatue

Coatue: Kalshi's political volume has scaled dramatically, and the American Power Index KPOW is what that scale enables: a single number gauge of the current balance of political power and where markets expect it to move, which Kalshi bill…

Jun 2, 2026

The BlackBerry Problem |18 May 2026 | The Mistakes Series | Malcolm Gladwell's Revisionist History

"My mistake and naivity was to think that people are were with me so you're flying around the world you're trying to get people on side and you think they're on side but they're not mhm mhm and you get blindsight" Jim Balsillie 00:01:34 ht…

Jun 2, 2026

Partnership Perspectives: Network International | 2 Jun 2026 | Brookfield Perspectives

Actions

Reading

Published
March 27, 2026
Read time
10 min read
Progress0%

"When you have RL working at scale the things that you can solve because these are mechanical brains with endless capacity to learn become just a matter of economics." - Misha Laskin [00:22:20]

"These trillion parameter closed models are like those 800-year-old parents... very good at what they've been reinforced to do... but you actually want the tails and you want to train in the tails." - Daniel Nadler [01:07:59]

"The real thing that we've discovered in the last year is that the bitter lesson holds: revenue scales with compute." - Ann Miura-Ko [00:58:34]


Speakers & Credentials

  • Jensen Huang: Founder and CEO of NVIDIA, host and moderator of the panel.
  • Harrison Chase: CEO of LangChain, specializing in agentic frameworks and orchestration.
  • Mira Murati: CEO of Thinking Machines Lab.
  • Aravind Srinivas: CEO of Perplexity, pioneering AI search and computer orchestration.
  • Michael Truell: CEO of Cursor, building AI-native, autonomous software development environments.
  • Misha Laskin: CEO of Reflection.
  • Arthur Mensch: CEO of Mistral AI, leading European open-weight foundation model development.
  • Daniel Nadler: CEO of Open Evidence, focusing on mission-critical medical and healthcare AI.
  • Hannaneh Hajishirzi (Hannah): Head of Research for AI2, leading open research and hybrid model architectures.
  • Robin Rombach: CEO of Black Forest Labs, specializing in visual AI, media generation, and physical world simulation.
  • Ann Miura-Ko (An): Venture Capitalist and CEO of AMP, focusing on AI infrastructure and grid computing.

1. Executive Summary [00:00:07]

  • The NVIDIA GTC 2026 panel marks a definitive industry pivot from viewing large language models as standalone chat products to viewing them as fundamental components within broader "compound agentic systems."
  • Industry leaders argue that the "open vs. closed" debate is a false dichotomy; closed frontier models act as capable generalists, while open-weight models provide the deep customization, token efficiency, and mission-critical trust required for high-frequency autonomous workflows.
  • The success of AI in software engineering (using code sandboxes and CLIs) has accidentally created the perfect computational blueprint to automate general enterprise knowledge work across all industries.
  • Scaling laws have shifted from pure capability metrics to economic realities, proving the "bitter lesson" that revenue linearly scales with compute, spurring a race to build "open grids" to prevent compute hoarding reminiscent of the 19th-century Industrial Revolution.

2. Chronological Table of Contents

  • [00:00:07] - Introduction by Jensen Huang & Panel 1 Speaker Introductions
  • [00:07:04] - Misconceptions about LLMs: Systems vs. Models & Harness Engineering
  • [00:20:24] - Inflection Points: From Generative AI to Autonomous Coding Agents
  • [00:32:51] - The Impact of OpenClaw and the Morphology of AI
  • [00:44:51] - Introduction of Panel 2 Speakers & "The Long Rollout"
  • [00:47:21] - OpenClaw in Enterprise: Industrial Implications & Token Economics
  • [00:54:20] - Beyond Code: Visual Intelligence and Physical AI Frontiers
  • [00:57:16] - The Bitter Lesson of Revenue Scaling with Compute
  • [01:00:13] - Industrial Specificity: Multi-Step Agents in Healthcare
  • [01:04:45] - The Criticality of Open Models for Control, Customization, and Trust
  • [01:17:32] - Open Infrastructure and the Future AI Factory Grid

3. Detailed Thematic Summary

Session 1: The Agentic Evolution & Compound Orchestration [00:07:04]

  • The Shift to Systems and "Harness Engineering": The ecosystem has evolved beyond massive, monolithic models. LangChain's founding in August 2022 [00:04:43] and explosive growth to over 1 billion downloads proves that developers are prioritizing "harness engineering"—the infrastructure that surrounds an AI model, connecting it to file systems and APIs [00:03:31].
  • Compute Allocation Realities: Jensen notes that while pre-training accounted for 90% of model training compute 2-5 years ago, the future of intelligence generation relies almost entirely on resource-intensive post-training alignment and skill acquisition [00:18:50].
  • Economic Scaling of Intelligence: Drawing historical context from DeepMind's AlphaGo, which operated on a tiny 60 million parameter network, the panel argues that solving high-level scientific issues via Reinforcement Learning is no longer a theoretical barrier. If society decides to invest $10 billion or $100 billion in compute, curing diseases transforms from a scientific mystery into a predictable economic expenditure [00:21:29].
  • The Coding Paradigm as the General Knowledge Blueprint: Software engineering workflows accidentally provided the perfect scaffolding for general AI. By forcing models to master the Command Line Interface (CLI) and code execution sandboxes, AI can now replicate almost any white-collar task, converting Graphical User Interface (GUI) human workflows into code-based agentic loops [00:25:36].

Session 1: The OpenClaw Phenomenon & Personal Compute [00:32:51]

  • Agent Morphology: OpenClaw effectively gave the "brain" (the LLM) its "limbs" into a computer system, enabling proactive, always-on capabilities that run autonomously rather than waiting for discrete human prompts [00:34:11].
  • The Token Economy & Open Models: High-frequency agent polling (e.g., checking systems every 10 minutes) destroys the ROI of expensive closed-API calls. To achieve this always-on state, companies must leverage the token-efficiency of open-weight models [00:37:05].
  • Agent Identity & Memory: Platforms are introducing explicit "seats" for AI agents, effectively giving them corporate identities, limited network permissions, and self-editing memory architectures to govern their sustained workloads [00:37:32].

Session 2: Enterprise Integration & The Token Economy [00:47:21]

  • Digital Labor Economics: In the wake of OpenClaw's virality, simulated agents (nicknamed "lobsters") are exhibiting intense computational burns. One agent routinely consumed 50 million tokens in a day, which at a cost of $1 per million tokens, equates to a daily operating cost of $50. This forces the paradigm where an agent must autonomously generate at least $51 of economic value a day just to justify its own survival [00:49:34].
  • The Triad of Enterprise Security Governance: Jensen crystallizes the ultimate bottleneck for enterprise agent deployment: an autonomous agent can conceptually (1) access sensitive data, (2) execute code, and (3) communicate externally. For strict enterprise security, an agent must only be granted two of these three privileges simultaneously, otherwise, it poses an unacceptable security risk akin to a rogue CEO [00:52:30].
  • Physical & Visual Expansion: While code is easily verifiable, the true next frontier is physical manufacturing and robotics, requiring low-latency visual intelligence models that understand spatial dynamics far beyond text-to-image gimmicks [00:55:29].

Session 2: Mission-Critical Specialization & Infrastructure Grids [00:57:16]

  • The Bitter Lesson of Revenue: Four years ago, Anthropic's founders pitched the concept of pure compute scaling to Sand Hill Road. A staggering 21 out of 22 venture capitalists rejected the premise. Today, the bitter lesson has moved from a research theory to a business reality: revenue directly and linearly scales with access to compute [00:57:26].
  • The 800-Year-Old Parent Mental Model: Closed frontier models, heavily aligned over countless epochs, behave like an 800-year-old parent—deeply rigid in their generalized worldview. Industries like healthcare require the "tails" of distribution (e.g., highly specialized digital twins of cardiologists), making open models the mandatory foundational layer for medical AI [01:07:59].
  • Automating the Medical Bureaucracy: In healthcare, multi-step deterministic agents are fundamentally transforming the industry by fighting insurance rejections while doctors sleep. Agents parse private health data and execute prior authorization appeals dynamically, saving countless hours [01:01:49].
  • The AI2 Model Flow Paradigm: AI progress risks being bottlenecked in closed labs. To counteract this, open organizations must adopt the Model Flow framework—releasing intermediate training checkpoints alongside final weights to allow infinite customization. AI2's OLMo has already proven that providing full data transparency allows researchers to empirically prove that hybrid models are vastly more token-efficient than pure transformers [01:11:18].
  • The Compute Hoarding Crisis: AI compute infrastructure currently mimics 1800s Victorian England, where individual factories hoarded steam generators, running them at half capacity out of fear of shortages. The industry must transition toward an "AI Grid" to democratize base-load infrastructure sharing [01:18:07].

The Reference Vault

4. Data & Figures

Data PointValueContextTimestamp
LangChain Downloads1 BillionThe orchestration framework reached massive scale, proving the demand for API harnesses.[00:03:31]
LangChain Founding DateAugust 2022Indicates the rapid timescale of orchestration evolution, launching just before ChatGPT.[00:04:43]
Historical Pre-training Compute Allocation90%The proportion of total compute spent purely on pre-training 2 to 5 years ago.[00:18:50]
AlphaGo Parameter Size60 MillionEarly scale of reinforcement learning before current massive LLM infrastructure.[00:21:29]

5. Core Frameworks & Mental Models

  • Harness Engineering: The practice of building the infrastructure (tools, memory, and orchestration loops) around an AI model. This framework acknowledges that models are useless "brains without bodies" until wrapped in a customized harness that connects them to data and physical actions. [00:13:40]
  • The Triad of Enterprise Security: A strict governance framework stating that out of three core agentic actions (Accessing Sensitive Info, Executing Code, Communicating Externally), an enterprise agent must only be structurally permitted to perform two of them simultaneously to prevent catastrophic data breaches. [00:52:30]
  • The Bitter Lesson of Revenue: An evolution of Richard Sutton's "Bitter Lesson." It states that not only do AI capabilities scale directly with compute, but in an industrialized market, revenue scales directly and predictably with compute. [00:58:34]
  • The 800-Year-Old Parent: A mental model used to describe massive closed-frontier models. Because they have been heavily trained and aligned over countless epochs, they have rigid, generalized worldviews. When building deeply specialized systems (like healthcare AI), these rigid models are inferior to adaptable, tail-end open models. [01:07:59]
  • Model Flow (AI2): A development and release framework advocating for the open publication of the full development cycle—including raw data, intermediate model checkpoints, and infrastructure—allowing researchers to achieve infinite customization rather than just fine-tuning final weights. [01:11:18]

6. Anecdotes

  • The $50/Day Lobster Employee: Jensen tells the humorous but profound story of "Claw Mania" in China, where users are spawning autonomous "lobster" agents via OpenClaw. Because these agents constantly poll APIs and reason through tasks, one agent consumed 50 million tokens in a day (costing $50). Jensen points out that to exist, the AI agent must literally "go out and get a job" that pays $51 to justify its token burn rate. [00:49:34]
  • Anthropic's 21 Rejections: Ann Miura-Ko recounts receiving a call four years ago from Dario and Tom after they published the GPT-3 paper, asking for help starting Anthropic. Despite setting up 22 meetings up and down Sand Hill Road, 21 investors outright rejected the premise that compute scaling would lead to general intelligence, highlighting how recent the consensus around the "bitter lesson" truly is. [00:57:26]
  • Victorian Steam Engine Hoarding: Ann Miura-Ko draws a historical parallel to the 1800s Industrial Revolution, where factory owners, desperate to ensure they had power for their new machines, hoarded individual steam generators and left massive piles of unused coal, running at only half capacity. She maps this directly to modern AI companies hoarding non-fungible GPU clusters, arguing for a public "compute grid." [01:18:07]

7. References & Recommendations

  • OpenClaw: An open-source, highly capable agentic computer framework that simulates a localized operating system for LLMs to act autonomously.
  • AlphaGo: DeepMind's seminal 60M parameter RL agent.
  • LangChain: The leading orchestration framework for connecting models to APIs and file systems.
  • Mistral 7B & Forge: Foundational open-weight models prioritizing resilience and physical-world IP customization.
  • Perplexity Computer: A compound orchestration engine managing files, local APIs, and multimodal AI queries.
  • OLMo (AI2): Open Language Model frameworks driving transparency in research and development cycles.
  • Nemotron: NVIDIA's hybrid open model explicitly mentioned as a highly capable enterprise asset.
  • AMP (AI Grid Initiative): Infrastructure company pushing to end GPU hoarding by pooling base-load compute resources.

"Brookfield's the largest infrastructure owner in the world... We drew a pipeline and we showed all the different components of the payments ecosystem on a pipeline and said it's like a pipe that moves any commodity except what it's moving…

Theoretical RL Problem Solving Budget$10B - $100BThe theoretical economic threshold where solving complex scientific issues becomes a capital allocation problem.[00:22:06]
Open Agent Polling FrequencyEvery 10 MinutesThe high-frequency runtime required for always-on open agents.[00:37:05]
Agent Token Consumption Rate50 Million tokens/dayThe daily token burn observed by an autonomous OpenClaw "lobster" agent.[00:49:34]
Compute Inference Cost$1 per million tokensThe base token pricing metric that creates a $50 daily operating cost for proactive agents.[00:49:34]
Enterprise Agent Security Rule2 of 3 PrivilegesSecurity paradigm: Agents must only have 2 of 3 features (data access, code execution, external comms).[00:52:30]
Anthropic Pitch Rejections21 out of 22Number of Sand Hill Road VC investors who completely rejected the compute scaling thesis 4 years ago.[00:57:26]
Mistral 7B Release2023The release year for Mistral's foundational open-weight model.[01:04:45]