1. The Last Six Months Changed Everything
0:00LLM advances across reasoning, tool use, and long-context finally came together in the last six months. AI is no longer a demo — it's genuinely useful. The question has shifted from "Can AI do this?" to "How do we deploy it?"
- Multiple breakthroughs converged simultaneously — reasoning, tool use, long context
- AI crossed the threshold from impressive demo to production-ready utility
- The central challenge is now deployment, not capability
2. Why NVIDIA Invests in Open Agent Ecosystem
0:34Jensen argues AI must be open. NVIDIA's investment thesis centers on domain-specific agents that enterprises build on open foundations. Super agents operate in flywheels — getting better as they process more domain data.
- AI infrastructure must be open for enterprise adoption
- Domain-specific agents are the killer use case, not general chatbots
- Super agents create flywheels — they improve with use
- NVIDIA's role: provide the compute and open model layer
3. How to Specialize Agentic Systems
4:14The recipe: start with intelligence that's "good enough" (Nemotron), combine it with a harness like LangChain, and add domain-specific data. The breakthrough is post-training the model inside the harness, not separately.
- Three ingredients: base model + harness + domain data
- Post-training inside the harness is the key innovation
- Nemotron provides the "good enough" intelligence foundation
- LangChain serves as the orchestration harness
4. Nemotron 3 Ultra: Near-Frontier at 10x Lower Cost
5:54Nemotron 3 Ultra scores 86% on agentic benchmarks vs Opus at 87% — near-frontier performance. DeepSeek and MiniMax sit at 82-83%. The kicker: Nemotron costs 10x less to run.
- 86% vs 87% (Opus) on agentic benchmarks — essentially at parity
- DeepSeek/MiniMax at 82-83% — meaningful gap below
- 10x cost reduction makes new deployment patterns viable
- Open weights mean enterprises can self-host and customize
5. Cheaper Intelligence Finds Better Answers
6:48Cost-effective inference enables exploring a larger search space. Fast thinking means more exploration paths, which means better answers. It's not about the smartest single call — it's about the most exploration per dollar.
- Cheaper inference = larger search space per dollar
- Fast thinking enables more exploration paths
- More exploration = higher probability of finding optimal answers
- The economics of intelligence change the architecture
6. Frontier vs Open Models
8:50Jensen's practical advice: start with frontier models for general tasks, then add specialized sub-agents using open models for domain-specific work. They're complementary, not competing.
- Frontier models handle general reasoning and orchestration
- Open models power specialized sub-agents for domain tasks
- Complementary strategy, not either/or
- Open models enable customization frontier models can't offer
7. Building Specialized Super Sub-Agents
9:50NVIDIA itself uses Deep Agents + Nemotron for internal tasks: supply chain optimization and chip design. These are "super sub-agents" — built for one job, extremely good at it, and proprietary to the company.
- NVIDIA dogfoods its own agent stack for supply chain and chip design
- Super sub-agents: narrow scope, deep expertise, proprietary data
- Built for one job — not general purpose
- These agents become company crown jewels
8. Companies Built on Harnesses
13:10The future: companies are built on harnesses, not business processes. LangChain becomes the company OS. Every workflow becomes an agent pipeline orchestrated through a harness.
- Harnesses replace traditional business process management
- LangChain as company operating system — orchestrating all workflows
- Every business process becomes an agent pipeline
- The harness is the new enterprise platform
9. Why Open Stacks Empower Enterprises
14:48Company intelligence is IP — it's too important to outsource. Enterprises need open tools they can control, customize, and deploy on their own infrastructure. You can't rent your competitive advantage.
- Company intelligence = intellectual property = competitive moat
- Intelligence is too important to outsource to a third party
- Open tools enable full control over the AI stack
- Self-hosting and customization are enterprise requirements
10. Deep Agents + OpenShell Blueprint Announced
17:25New blueprint announced: Deep Agents + Nemotron 3 Ultra + OpenShell secure runtime. A complete, deployable stack for enterprise agent systems. Deploy anywhere — cloud, on-prem, or DGX Spark.
- Deep Agents: NVIDIA's agent framework for complex multi-step tasks
- Nemotron 3 Ultra: the intelligence layer (open weights)
- OpenShell: secure runtime environment for agent execution
- Deploy anywhere: cloud, on-prem, DGX Spark
11. Runtime, Security, Access Control
18:53Agent deployment needs HR-like onboarding. Just as you wouldn't give a new employee access to everything on day one, agents need access control, sandboxing, and governance frameworks.
- Agent onboarding mirrors HR onboarding — graduated access
- Access control: what data and tools can each agent touch?
- Sandboxing: isolate agent execution environments
- Governance: audit trails, compliance, accountability
12. Why More AI = More Jobs
22:12Engineers aren't typing Python anymore — they're building agents. The new roles: creating evals, benchmarks, guardrails, and agent architectures. More AI creates more work, not less.
- Engineers shift from writing code to building agent systems
- New roles: eval creation, benchmark design, guardrail engineering
- Agent architecture becomes a core engineering discipline
- More AI deployment = more human work to build, maintain, and govern it
🎯 Key Takeaways
- The last 6 months changed everything — AI is finally useful
- Companies of the future are built on harnesses, not business processes
- Nemotron 3 Ultra: 86% vs Opus 87%, at 10x lower cost
- Cheaper intelligence explores larger search spaces, finds better answers
- Start with frontier, add specialized open-model sub-agents over time
- Super agents: domain-specific, proprietary, your crown jewels
- Intelligence is too important to outsource
- New blueprint: Deep Agents + Nemotron 3 Ultra + OpenShell
- Agent deployment needs security/access control like HR onboarding
- More AI = more jobs — building agents, evals, guardrails
- Post-training models inside harnesses is a breakthrough
- Deploy anywhere: cloud, on-prem, DGX Spark