Date: 2026-03-17 Source: Jensen Huang GTC 2026 Keynote (2026-03-16, San Jose), multi-source cross-survey Perspective: For practitioners building Agentic AI systems daily (Claude Code / LangChain / Multi-agent orchestration)
The surface narrative of GTC 2026 is that the era of training is over and the era of inference has begun. The underlying logic is that NVIDIA, through an open ecosystem design, is turning the model layer into a replaceable commodity, anchoring itself as the sole infrastructure provider.
Jensen Huang spent over two hours building a complete argument: data centers are transforming from file warehouses into Token Factories, where the core metric has shifted from storage and bandwidth to how many tokens are produced per watt. He compared OpenClaw (a third-party open-source Agent framework) to Linux, predicting that every SaaS will become AaaS (Agent-as-a-Service), and engineers’ future offers will specify an annual token budget.
Placing these statements back into the decision space makes the strategic intent clear.
The “Token Factory” metaphor has clear strategic considerations. Factories produce standardized, volume-priced industrial goods. In this framework, GPT-5 and Claude 4 become recipes on an assembly line. Recipes are important, of course, but the one who monopolizes the production line takes the largest profit. Through a seemingly neutral economic metaphor, Huang defines the model layer as a replaceable link in the supply chain: whoever can produce tokens at the lowest cost and highest efficiency wins. This is a role only NVIDIA can play.
Comparing OpenClaw to Linux is an extension of the same logic. OpenAI’s path is to build a closed ecosystem with closed models and APIs. NVIDIA embraces OpenClaw, a model-agnostic orchestration layer, with the goal of making the Agent framework layer free and open-source. When applications all run on the free OpenClaw and the underlying models become pluggable APIs, the only thing left with pricing power is the hardware and the inference stack. This is consistent with Google’s strategy for Android.
To understand what NVIDIA did, one must first understand the paths they could have chosen. Behind the GTC 2026 announcements lie at least five key decision points, each with distinct paths.
NVIDIA could have just sold chips (pure hardware vendor), built a complete inference cloud service (competing with AWS Bedrock), or created a middle layer: an open-source inference operating system. They chose the middle layer, which is Dynamo 1.0.
Constraint Logic: Building a cloud service would put them in direct competition with AWS, Azure, and GCP—NVIDIA’s largest customers (the top five supercomputing vendors account for 60% of revenue). Selling only chips would mean cloud providers reap all the benefits of software optimization, turning NVIDIA into a mere component supplier. The middle layer is the only position that allows profiting from both sides. Dynamo is open-source, so theoretically anyone can use it, but the optimal implementation of disaggregated serving and KV cache pinning depends deeply on NVLink bandwidth, HBM4 memory hierarchy, and GPU-to-GPU communication topology. Open-source code, tied to hardware. It’s the same model as Android’s AOSP being open-source while Google Play Services remains closed.
What was sacrificed: Direct software revenue. If Dynamo were closed-source, they could sell licenses like VMware. NVIDIA judged that the network effect of an inference OS is more valuable than software revenue: if everyone uses Dynamo and optimizations are targeted at NVIDIA hardware, the reason to buy NVIDIA becomes even stronger.
NVIDIA has the most GPUs and the capability to train top-tier models, but they chose “good enough.” Nemotron 3 Super is a MoE with 120B total parameters / 12B active, scoring 60.47% on SWE-Bench. The official recommended deployment pattern itself explains the positioning: Nano for simple tasks, Super for planning, and closed-source frontier models for hard problems.
The core of the constraint conflict: A platform entering the model layer directly threatens customer relationships. If NVIDIA trained a model that surpassed Claude, Anthropic would seriously evaluate AMD and Google TPUs because their hardware supplier would have become a direct competitor. The 120B/12B MoE design of Nemotron shows that the optimization goal is inference efficiency (running fast on their own hardware); the model’s capability itself is secondary. A 60.47% SWE-Bench score is enough to prove that open-source models are viable on NVIDIA hardware while maintaining a safe distance from Claude and GPT-5.
True priority ranking: Maintain hardware monopoly > Enter the model market.
NVIDIA could have built their own Agent framework (competing with LangChain), acquired an existing one, or embraced community projects and added a security layer. They chose the third path: building NemoClaw (OpenShell sandbox + Privacy Router + Agent Toolkit) on top of OpenClaw.
Competition in Agent frameworks is already fierce; building their own would offend all players like LangChain, CrewAI, and AutoGen. Embracing a community project and adding an enterprise security layer is politically safe. The key to Linux’s success was that no single company owned it; Red Hat made money by adding enterprise features on top. NemoClaw follows the same strategy.
A hidden power position worth watching: Privacy Router. This router decides which requests go to the local Nemotron model and which go to the cloud-based Claude or GPT. On the surface, it’s a privacy feature; in reality, it’s a traffic scheduler. Whoever controls the routing logic controls the traffic distribution for model providers, similar to the logic of Safari’s default search engine: Google pays Apple tens of billions of dollars annually to remain the default because the default route is pricing power. NVIDIA has not yet clearly specified the default behavior, degree of openness, or configurability of the routing policy for Privacy Router.
NemoClaw is currently in early alpha, and NVIDIA’s own description is “expect rough edges.”
The Vera CPU could have been a general-purpose server CPU (competing directly with EPYC/Xeon), a secondary co-processor for the GPU, or a specialized processor for specific workloads. NVIDIA chose the narrowest path: 88 Arm Olympus cores, specifically optimized for Agent sandboxes, supporting 22,500 concurrent sandboxes per rack.
This choice implies a clear technical judgment: the bottleneck of Agentic AI is shifting from GPU inference to CPU-side sandbox execution. Amdahl’s Law is at work: GPUs are efficient at token generation, but many Agent operations are CPU-intensive serial tasks (compiling code in a sandbox, running a browser, querying a database, executing API calls). This matches the daily experience of using Claude Code: the model’s response might take 2-3 seconds, but waiting for bash commands to execute, tests to run, and linters to return results often takes 10-30 seconds. In multi-agent systems, where each agent needs its own isolated execution environment, the problem is magnified exponentially.
What was sacrificed: The general-purpose server CPU market. AMD EPYC’s market share in data center CPUs is still growing, and NVIDIA could have competed for it. Choosing specialization means a much smaller addressable market but dominance in the agent sandbox niche.
The member list of the Nemotron Coalition is worth examining: Mistral (France), Sarvam (India), Thinking Machines Lab (Mira Murati’s new company, former OpenAI CTO), Reflection AI, Cursor, LangChain, Perplexity, and Black Forest Labs.
From a geopolitical perspective, this list is very strategic. Members cover Europe’s demand for digital sovereignty (Mistral), India’s need for local models (Sarvam), and Silicon Valley teams independent of OpenAI (Thinking Machines, Reflection). As the world’s largest AI hardware supplier, NVIDIA’s interest is best served by maintaining multi-polar competition in the model layer. If any single model company were to unify the market, NVIDIA’s bargaining power would decrease. Supporting various global competitors and ensuring the model layer remains fragmented is the optimal strategy for NVIDIA.
The first project is a foundation model co-developed by Mistral and NVIDIA, which will become the basis for Nemotron 4. Cursor contributes real evaluation data from coding scenarios, LangChain builds the agent harness and observability, and Perplexity contributes production experience in search and reasoning. Each member provides differentiated application-layer value to NVIDIA’s hardware ecosystem.
Jensen Huang announced that software optimization alone (disaggregated serving + KV cache pinning) could achieve a 7x performance improvement on existing Blackwell hardware.
The other side of this number: if pure software optimization can bring a 7x improvement, the actual utilization of the GPUs heavily purchased by the industry over the past few years might only be around 15%. At least part of the compute shortage in the last two years was caused by crude system architectures and a lack of inference-layer optimization. GPUs are indeed in short supply, but the degree of shortage has been exaggerated.
This also means the source of future performance dividends is changing—from mindless hardware stacking to system-level optimization of memory bandwidth and routing. For AI practitioners, the operational value of understanding Dynamo’s disaggregated serving architecture (splitting encode, prefill, and decode into independent workers) may be higher than upgrading hardware.
Note the specific context of the 7x figure: running DeepSeek R1-0528 (FP4) on GB200 NVL72, with 1k/1k input/output, targeting an interactive speed of about 50 tok/sec/user. Improvements on other benchmarks range from 1.5x to 4x, with the largest gains appearing in MoE models and scenarios with KV cache reuse.
Every platform shift redistributes value. If NVIDIA’s vision is realized, besides AMD and Intel, two directions are worth watching.
First, cloud service providers. When Dynamo becomes the inference OS and NemoClaw’s Privacy Router can intelligently schedule between local sandboxes and the cloud, the value-added service layer of cloud vendors (managed inference, model hosting) will be compressed. CNBC’s analysis also points out that NVIDIA is trying to expand from a GPU supplier to a full-stack supplier for the entire AI factory: compute, networking, storage, inference software, Agents, and robotics. This full-stack ambition directly squeezes the value space of cloud vendors.
Second, Meta’s Llama. The Nemotron Coalition’s impact on Anthropic and OpenAI is limited because their moat is in model capability rather than inference efficiency. But the impact on Llama is substantial. If Nemotron 4’s inference efficiency on NVIDIA hardware is significantly better than Llama’s, users choosing open-source models will lean toward Nemotron because most people’s inference hardware is NVIDIA GPUs. NVIDIA is shifting from “all open-source models run well on our hardware” to “our own open-source models run best on our hardware.”
Throughout the keynote, Jensen Huang hardly mentioned AGI or the next milestone of Scaling Law. Instead, he spent a lot of time on Uber self-driving cars, Disney robots, and industrial software.
This choice reflects a judgment: the marginal utility of simply increasing model parameters is diminishing. To support the $1 trillion infrastructure demand in 2027, the narrative focus must shift from “creating stronger models” to “creating economic value with existing models.” The cash flow of AaaS provides more support for valuation than the vision of AGI.
Another signal: if pre-training scaling were still the main theme, the statement “the era of training is over” would not hold. The reason Huang can confidently announce the entry into the inference era is the underlying assumption that the marginal gains of training have flattened enough that inference optimization has become a larger source of value. This narrative favors NVIDIA: inference is always-on continuous consumption, while training is a one-time investment. Inference era = continuous hardware procurement.
Vera Rubin Platform: 7 chips, 5 racks. Core chip: Rubin GPU (336B transistors, TSMC N3, 288GB HBM4, 22 TB/s bandwidth, 50 PFLOPS FP4). Compared to Blackwell: 5x inference, 3.5x training, 10x energy efficiency, 1/10 token cost. An NVL72 rack houses 72 Rubin GPUs + 36 Vera CPUs, with 260 TB/s internal interconnect. Shipping in H2 2026. Azure is already running the first unit, as confirmed by Satya Nadella.
Vera CPU: 88 Olympus cores (Arm v9.2), 1.5TB LPDDR5X, 1.2 TB/s bandwidth. 256 CPUs per rack, 22,500+ concurrent sandboxes. 4x higher density and 2x higher energy efficiency than x86 solutions.
Dynamo 1.0: Open-source distributed inference OS (github.com/ai-dynamo/dynamo). Core capabilities: disaggregated encode/prefill/decode, KV cache pinning, agent hints routing, NIXL fast GPU-to-GPU data transfer. Already integrated with vLLM, SGLang, and TensorRT-LLM. Kubernetes-native deployment (AWS EKS, Azure AKS, GKE, OCI).
Nemotron 3 Super: 120B total parameters / 12B active (MoE), 1 million token context, Hybrid Mamba-Transformer. SWE-Bench Verified 60.47%, PinchBench 85.6% (best open model for OpenClaw Agent scenarios), RULER 1M 91.75%. Throughput is 2.2x higher than GPT-OSS-120B. Open weights, training data, recipes, and evaluation pipelines.
NemoClaw: Early alpha status. OpenShell sandbox + Privacy Router + Agent Toolkit. One-command deployment. Supports GeForce RTX, DGX, and cloud. Co-developed with OpenClaw founder Peter Steinberger.
The Nemotron 3 series also includes Ultra (flagship), Omni (multimodal), VoiceChat (real-time voice), and Nano (edge).
Physical AI: Uber robotaxis will hit the road in the Bay Area/LA in 2027, with plans to cover 28 cities across four continents by 2028. Disney’s Olaf robot (trained with Newton physics engine + Isaac Lab) walked on stage with Jensen. Cosmos 3 unifies synthetic world generation, visual reasoning, and action simulation. GR00T N2 preview performance on new tasks and environments is more than 2x higher than leading VLA models.
Industrial Software: Cadence, Dassault, PTC, Siemens, and Synopsys integrate CUDA-X and Omniverse. Honda’s aerodynamic simulation accelerated by 34x; Samsung/SK hynix use GPUs to accelerate lithography. NVIDIA is expanding from an AI chip supplier to the computational substrate for physical manufacturing.
DLSS 5: Releasing in Fall 2026, shifting focus from frame rate improvement to visual fidelity. AI-enhanced textures, lighting, and environmental details.
Vera Rubin Space-1: A computing platform for low Earth orbit, with partners Axiom Space, Starcloud, and Planet Labs.
Feynman Architecture Roadmap Preview: The next generation after Vera Rubin, including new GPUs, LP40 LPUs, and the Rosa CPU (named after Rosalind Franklin).
93% of analysts have a buy rating, with an average target price of around $267 (45%+ upside), but on the day of the keynote, the stock price rose 4.3% before pulling back, closing down 0.70% at $181.93.
CNBC’s analysis attributes this to market structure: option hedging activities and market maker operations pinned the stock price at its current level. Bernstein believes that based on a 2027 EPS of $10.68, a 17x PE is already very cheap. Stifel’s evaluation is more precise: the $1 trillion figure validated rather than raised existing expectations. Everyone already knew NVIDIA would do well; GTC provided confirmation rather than incremental information.
Deeper macro tension: the Iranian oil crisis is pushing up energy prices, and an AI bull market needs a stable economic environment, low-cost data center electricity, and expectations of interest rate cuts. These two forces are moving in opposite directions, and the market is waiting for the contradiction to be resolved.
The bottleneck is moving. The existence of the Vera CPU is a hardware-level signal: the problem NVIDIA is spending real money to solve is the efficiency of tool execution and sandbox management, not model inference speed. If you are designing an agent architecture, your energy should shift more toward efficient tool-calling patterns and sandbox lifecycle management. Current Docker-container-based agent sandbox solutions (including Claude Code’s sandbox) may need a fundamental redesign. The 22,500 concurrent sandboxes imply a pattern closer to serverless functions (created, executed, and destroyed in seconds) rather than long-running containers.
The inference stack is getting thicker. Dynamo’s disaggregated serving means inference has become a distributed systems problem. Encode, prefill, and decode execute on different workers, KV cache is reused across requests, and routing is dynamically adjusted based on agent hints. This has a direct impact on the design of multi-agent systems: KV cache pinning means multi-turn agents should maintain session affinity rather than being randomly routed each time.
The layered pattern of Super + Nano + frontier models is worth trying immediately. Nemotron 3 Super’s 60.47% SWE-Bench score and 1 million token context are sufficient for most structured coding and planning tasks. Routing simple requests to Nano, medium complexity to Super, and calling Claude/GPT for extreme scenarios could theoretically cut inference costs by more than half. The prerequisite is that the routing logic is smart enough, which is an interesting engineering problem in itself.
NVIDIA’s ecosystem lock-in is deepening, but embracing it remains a rational choice for now. As the inference stack is built on Dynamo + NVIDIA GPUs, migration costs will continue to rise. However, alternative solutions (AMD software stack, the openness of Google TPUs) are not yet mature. The pragmatic approach is to embrace the NVIDIA ecosystem for performance advantages while maintaining an abstraction layer for the inference backend in your architecture to leave room for future migration.
Watch the actual adoption curve of OpenClaw. Jensen’s Linux analogy is bold, but for OpenClaw to become an industry standard, it needs ecosystem advantages: enough tools, integrations, tutorials, and community support to make not using OpenClaw a choice that requires justification. Currently, NemoClaw is still in early alpha. More than what NVIDIA says, it’s worth tracking OpenClaw’s GitHub star growth and PR activity.
This report synthesizes information from multiple sources, including NVIDIA official press releases, developer blogs, SemiAnalysis InferenceX benchmarks, CNBC/Sherwood News/Seeking Alpha market analysis, Latent.Space/The Decoder technical analysis, and Tom’s Hardware hardware reviews, and was written after cross-verification.