--- title: "NVDA (Analyst Huddle): IT to become token distributors; 50% of CF to shareholder returns" type: "Topics" locale: "en" url: "https://longbridge.com/en/topics/39326852.md" description: "Below is Dolphin Research's Trans from NVDA's GTC 2026 analyst session. For our readout on the GTC conference, see 'NVDA GTC: The AI gala—high hopes, letdown?'. I. $NVIDIA(NVDA.US) analyst session key takeaways.1) Demand outlook: Committed demand for Blackwell + Rubin through end-2027 exceeds $1tn (POs and firm commitments), excluding new products such as Rubin Ultra, Feyn, a standalone Vera CPU, and Groq. Orders are expected to continue building on this base..." datetime: "2026-03-18T04:38:08.000Z" locales: - [en](https://longbridge.com/en/topics/39326852.md) - [zh-CN](https://longbridge.com/zh-CN/topics/39326852.md) - [zh-HK](https://longbridge.com/zh-HK/topics/39326852.md) author: "[Dolphin Research](https://longbridge.com/en/news/dolphin.md)" --- > Supported Languages: [简体中文](https://longbridge.com/zh-CN/topics/39326852.md) | [繁體中文](https://longbridge.com/zh-HK/topics/39326852.md) # NVDA (Analyst Huddle): IT to become token distributors; 50% of CF to shareholder returns **Dolphin Research's notes from NVIDIA GTC 2026 analyst huddle. For the GTC takeaways, see '**[**NVIDIA GTC: AI's Spring Gala — Hype In, Letdown Out?**](https://longbridge.cn/en/topics/39303579?channel=SH000001&invite-code=294324&app_id=longbridge&utm_source=longbridge_app_share&locale=zh-CN&share_track_id=4157ebca-9c64-4dae-97b9-1939b53307b7)**'.** **I.** $NVIDIA(NVDA.US) **Key takeaways from the huddle** 1\. **Demand outlook**: **Committed demand (POs and firm orders) for Blackwell + Rubin through end-2027 exceeds $1 tn**. This **excludes Rubin Ultra, Feyn, a standalone Vera CPU, Groq and other new products**. Orders are expected to continue building on top of that base. 2\. **Capital returns**: **The company plans to return 50% of FCF via buybacks and dividends**. This ratio does not yet contemplate the 'incremental' demand above the $1+ tn for Blackwell + Rubin. If that incremental demand materializes, there will be room to raise the proportion or size of returns. In 1H26 the company will first honor existing investment commitments, then step up buybacks; **returning capital remains a core focus**. 3\. **GPM logic**: Each generation delivers multi-fold improvements in tokens/sec/W. Customers prefer to pay more for the new generation rather than buy prior-gen cheaper, sustaining margins. This underpins margin durability. 4\. **Cash use priorities**: **(1) Supply chain investments (prepayments, capacity)** to support trillion-dollar shipments; **(2) ecosystem investments**; **(3) shareholder returns**. This sequence reflects growth-first capital allocation. **II. Detailed notes from the analyst huddle** **2.1 Management highlights** 1\. **Three AI inflections and Agentic systems** a. The three inflections: GenAI → inference → Agentic systems (current phase). This frames the shift from content generation to autonomous task execution. b. Agentic systems execute tasks autonomously. Coding is among the most popular applications. Automation moves from assistance to action. c. New engineers now receive a token budget in addition to a laptop. Token budgets are becoming a real enterprise spend line. This formalizes tokens as a cost center. d. Computers are shifting from tools to 'manufacturing equipment', akin to ASML lithography, producing tokens for sale. Compute generates output tokens as the economic unit. 2\. **Open Claw and the software ecosystem** a. Open Claw is an open-source Agent framework. 1.5 mn downloads enable building an Agent with one line of code. Adoption is broad-based. b. NVIDIA is building an enterprise version, Nemo Claw. This targets production-grade deployments. c. Open Claw serves as the operating system for the Agentic computing stack. Every company needs an Open Claw strategy. It standardizes Agent orchestration. d. The $2 tn global IT software industry will integrate AI models (OpenAI, Anthropic, open source) and resell tokens. TAM could expand to $8 tn. Software licenses plus token resale redefine biz. models. 3\. **Demand and customer mix** a. In 2025, Anthropic and Meta SL are added as platform customers. Open-source models become the No. 2 source of token generation. The ecosystem is diversifying. b. Market structure: OpenAI has the largest model footprint, open-source models rank second, Anthropic third, followed by a substantial long tail. The tail is large in aggregate. c. Revenue mix: ~60% from hyperscalers, ~40% from regional cloud, industrial, and on-prem enterprise (Dell, Lenovo, HP, etc.). Non-hyperscaler demand is meaningful. d. The 40% non-hyperscaler segment relies entirely on NVIDIA's full stack. Chips alone cannot reach these customers. Platform sales drive access. 4\. **TAM expansion: Vera Rubin vs. Grace Blackwell** a. **Vera Rubin adds ~50%** to the addressable market vs Grace Blackwell. This reflects broader system scope. b. New sources: Groq accelerators (~+25% compute spend), storage (the second-largest compute spend), and CPU/tooling (~5%). These layers lift total system cost. c. All unified into a single-rack architecture, 100% liquid-cooled, fully optimized. Standardization simplifies deployments. d. Grace Blackwell focused on inference; Vera Rubin addresses the full Agentic system. Scope shifts from single workload to end-to-end. 5\. **Org. design and annual product cadence** a. Jensen Huang directly manages a 60-person team. Org. structure mirrors the product architecture. This tight alignment accelerates execution. b. Seven chips share a unified tape-out schedule, and the full software stack is in-house (storage, networking, the Dynamo factory OS, etc.). Vertical integration is deliberate. c. CUDA compatibility ensures new systems run prior-gen software on Day 1. This de-risks customer upgrades. d. Owning the full software stack is prerequisite for annual product turns. Outsourcing any link would make it impossible. Control enables cadence. **2.2 Q&A** **Q: When will hyperscalers' API and cloud revenue show upside commensurate with capex?** A: I wish those companies were public so you could see what I see. Historically, no startup has grown revenue by $1–2 bn per week, but that is what is happening now. The $2 tn IT software industry will not be disrupted so much as transformed. Every IT company will integrate OpenAI, Anthropic, and open-source models, building Agents via Open Claw. **100% of global IT will become token distributors for OpenAI and Anthropic. This industry could expand from $2 tn of software licenses to $8 tn while reselling massive token volumes. Raise your revenue expectations for OpenAI and Anthropic**. **Q: How do Agentic AI needs affect Vera Rubin rack architecture and pricing?** A: We unified everything into a single-rack architecture — same power, same cooling, 100% liquid-cooled, fully optimized. To run Agentic systems, you add ~25% on top of GPU compute spend to deploy Groq. Groq uses ~8x the number of chips, with total cost roughly equal to an NVLink 72 rack. **This is not in the $1 tn — if 100% of the $1 tn workloads add Groq, it becomes $1.25 tn**. Adding storage (the second-largest compute spend) and CPU (~5%), Vera Rubin expands TAM by ~50% vs Grace Blackwell. Grace Blackwell solved inference; Vera Rubin solves the full Agentic system. **Q: How will NVIDIA strategically deploy the cash it generates?** A: Priority one is to support growth — long-term supply-chain partnerships, planning with partners, sometimes prepaying or even funding capacity expansion to prepare for $1+ tn of shipments. Second is ecosystem investment; growing CUDA developers and AI-native companies matters. We will still generate significant FCF. As of now, buybacks plus dividends are about 50% of FCF. We will complete some existing investment commitments in 1H, then the 'incremental' portion opens additional buyback capacity. **Q: How do you respond to concerns that NVIDIA captures too much ecosystem value and its margins are unsustainable?** A: Much of what I said yesterday is a fresh lens. People need to understand **token economics**. If we keep delivering multi-fold gains in tokens/sec/W each year, and lift customer ASPs by introducing new token segments, customers will gladly stay with us. Core logic: you are not reselling computers, you are using computers to manufacture tokens. The computers are expensive yet produce low-cost tokens at very high efficiency — you **own the most expensive computers and the lowest unit-cost token output**. That is our job and the source of our margins. Like TSMC wafers and ASML tools — the priciest but best value. The question is simple: do you want to make more money, or buy the cheapest gear? Anyone saying 'my chip is 30% cheaper' — place it in the factory economics and you'll see they do not understand AI. **Q: You expect a 2027 capacity shortfall. Where are the bottlenecks? Are customers pacing buys awaiting next-gen?** A: I told Satya 'buy what you need this year because next year will be better' — those words are mine. On supply, the world is always constrained to some degree. We orchestrate across dimensions and suppliers to keep the system harmonious — not too much, not too little. If I name a specific constraint, I know what you will do. The system is balanced: just-enough power, builders, cables, optics. We optimize daily. Demand is accelerating, and we can support it. **Q: What is the migration path from copper to CPO (co-packaged optics)? How do NVL 576 and NVL 1152 coexist?** A: We should use copper as long as possible — copper's limit is ~1 meter. You saw us scale from NVLink 72 to Rubin Ultra's NVLink 144; the backplane supports this. If we can extend from 144 to 288, we will — copper is easier to make and more reliable, and humanity has used it for a long time. **We are at 100% copper now. Next-gen Ultra will have two options: all copper or copper + CPO. The following 1152 scale will be all CPO** due to copper's physical distance limit. Even as NVLink moves to CPO, Ethernet expansion and storage will still use copper. With five rack types and rising total demand, copper connector consumption will continue to grow. **Q: Growth outlook by token segment? Will token cost deflation slow?** A: **Token costs will keep falling every year** — from Grace Blackwell to Rubin to Rubin Ultra. At the same time, token intelligence rises and throughput increases. You must normalize by 'tokens/sec/W' — your data center is a fixed size, so comparisons must be normalized. Segment mix depends on brand and target customer. **If your biz. is search, most is on the free tier; if Agent coding, it is high-end; if target users earn $50–70k, token pricing is mid-tier**. All segments are growing exponentially because we are still early. **Q: What do state space models and hybrid architectures mean for NVIDIA? Does Agentic AI need new model architectures?** A: We run all AI models — full Transformer, discrete tokens, continuous diffusion, state space, hybrids — our architecture embraces all. NVIDIA is widely used because whatever researchers invent tomorrow, I guarantee it runs well on CUDA, as we have all required compute elements. Nemotron-3I uses a hybrid architecture to handle ultra-long context. You will have lifelong conversations with AI; managing relevant conversational memory is frontier research. Hybrids allow ultra-long context without quadratic compute blowup. We open-sourced it for everyone to use — to advance AI, not to compete with anyone. **Q: How concentrated is downstream AI usage? Do hyperscalers + frontier labs account for 80%+ of real usage?** A: I split this three ways. First, the end models: OpenAI is largest, open-source models collectively second, Anthropic third, with a substantial long tail. Add physical AI models (robotics VLMs, etc.), and model types are numerous. Second, where compute happens: some build in-house chips (we compete), some host NVIDIA customers (CUDA runs only on NVIDIA), some are NCP infrastructure customers (needing full systems), some deploy on-prem via Dell/HP/Lenovo, and some at the edge (RAN, robotics, ADAS, satellites). Compressed, it is a 60%/40% split. **The 40% slice relies 100% on NVIDIA's full stack — they do not buy chips; they buy platforms, need confidential computing, and global deployment**. Within the 60%, part is competitive and part is demand we bring to the clouds. OCI will not build chips, nor will CoreWeave — our position there is solid. **Q: Future trajectory of training compute? How will training vs. inference shares evolve?** A: Training is moving from pre-training to post-training. Pre-training is memorization and generalization — 'high school' for AI. Post-training teaches skills: RL, executable validation, tool use, structured APIs, etc. **I estimate post-training intensity at roughly a million times pre-training** — there are many skills, rollouts are long, and models must grow. In coming gens, pre-training will shift from internet data to synthetic data, adding multimodal and physical-world motion. The boundary between inference and training will blur — like humans, learning and applying are continuous. My hope is **99% of global compute ultimately goes to inference, where tokens convert to economic value** — no one pays for learning; they pay for tokens produced. That is why we pivoted hard to inference last year. Two years ago, the market said 'NVIDIA is good at training; inference is easy and anyone can do it' — inference is extremely hard. Look at the chart; inference is thinking, working, executing tasks — how could it be easy? **Q: Does the $1+ tn demand outlook include Rubin Ultra?** A: Absolutely not. **It covers Blackwell + Rubin only**. It excludes Rubin Ultra, Feyn, or any subsequent products. **Risk disclosure and statement:**[**Dolphin Research Disclaimer & General Disclosure**](https://support.longbridge.global/topics/misc/dolphin-disclaimer) ### Related Stocks - [Dell Technologies Inc. (DELL.US)](https://longbridge.com/en/quote/DELL.US.md) - [AutoNation, Inc. (AN.US)](https://longbridge.com/en/quote/AN.US.md) - [LENOVO GROUP (00992.HK)](https://longbridge.com/en/quote/00992.HK.md) - [Lenovo Group Limited (LNVGY.US)](https://longbridge.com/en/quote/LNVGY.US.md) - [LENOVO GROUP-R (80992.HK)](https://longbridge.com/en/quote/80992.HK.md) - [NVIDIA Corporation (NVDA.US)](https://longbridge.com/en/quote/NVDA.US.md) - [Vera Therapeutics, Inc. (VERA.US)](https://longbridge.com/en/quote/VERA.US.md) ## Comments (3) - **星走天下 · 2026-03-18T07:39:51.000Z**: 🖥 - **只为你倾心 · 2026-03-18T07:26:31.000Z**: Come in to learn - **Dolphin Research** (2026-03-18T08:34:19.000Z): More communication🤝, leads to true knowledge.