---
title: "NVDA GTC: AI's Big Show — High Hopes In, Disappointed Out?---"
type: "Topics"
locale: "zh-CN"
url: "https://longbridge.com/zh-CN/topics/39303579.md"
description: "On Mar. 16, 2026, Jensen Huang, founder and CEO of NVIDIA, delivered the keynote at GTC 2026.Core themes included the 20th anniversary of the CUDA platform, the inference inflection and a surge in compute demand, the Vera Rubin system architecture, Groq integration, the OpenClaw agent revolution, and physical AI and robotics..."
datetime: "2026-03-17T07:33:51.000Z"
locales:
  - [en](https://longbridge.com/en/topics/39303579.md)
  - [zh-CN](https://longbridge.com/zh-CN/topics/39303579.md)
  - [zh-HK](https://longbridge.com/zh-HK/topics/39303579.md)
author: "[Dolphin Research](https://longbridge.com/zh-CN/news/dolphin.md)"
---

> 支持的语言: [English](https://longbridge.com/en/topics/39303579.md) | [繁體中文](https://longbridge.com/zh-HK/topics/39303579.md)


# NVDA GTC: AI's Big Show — High Hopes In, Disappointed Out?---

On Mar 16, 2026, Jensen Huang, founder and CEO of NVIDIA, delivered the GTC 2026 keynote. Core topics spanned the 20th anniversary of the CUDA platform, the inference inflection and the surge in compute demand, the Vera Rubin system architecture, Groq integration, the OpenClaw agent revolution, and physical AI and robotics.

**I. GTC 2026 Key Takeaways**

**1) Data center revenue outlook**: For 2025-2027, NVIDIA guided to cumulative data center revenue of $1tn (vs. $500bn for 2025-2026 at last year’s GTC), broadly in line with expectations. The sell-side already models above $1tn and was looking for firmer color on orders and backlog.

**2) Performance and cost**: NVIDIA leads globally on both tokens per watt (throughput) and token speed (latency/intelligence). Its token cost is also the lowest.

**3) Data centers as ‘token factories’**: Each factory is power-constrained (e.g., 1 GW) and must manage token throughput and speed. Tokens will stratify like commodities across tiers: free (high throughput, low speed) -\> $3 per mn tokens -\> $6 per mn -\> $45 per mn -\> $150 per mn (top-tier, low-latency, high-bandwidth compute).

In a 1 GW data center, each 25% power tranche maps to one tier. Grace Blackwell can generate 5x the revenue of Hopper, and Vera Rubin can add another 5x uplift.

**4) Vera Rubin: Groq 3 LPU added to the prior six-chip lineup**.

**① Vera Rubin: 100% liquid-cooled (45°C hot-water cooling), all cables eliminated**, cutting install time from two days to two hours.

**② CPO (co-packaged optics) Spectrum-X switches**: now in full mass production, co-developed with TSMC.

**③ CPU**: the only data center CPU using LPDDR5, sold standalone and set to become a multi-billion-dollar business. The Vera CPU Tray targets agentic workloads and integrates eight Vera processors per compute tray, each with 88 cores and eight channels of LPDDR5x, delivering 1.2 TB/s memory bandwidth per socket. Two BF4-DPUs are integrated on the CPU Tray.

**④ Vera Rubin on Microsoft Azure (first rack)**: now live. NVIDIA’s supply chain can ship thousands of systems per week, enabling mn-GW scale AI factory capacity on a monthly basis.

**⑤ Rubin Ultra**: while Rubin sleds slide horizontally into racks, Rubin Ultra is vertically inserted into the new Kyber rack. 144 GPUs sit in one NVLink domain, with NVLink switches behind the midplane replacing copper cables.

**5) Groq 3 LPU (new chip): Groq plus HBM, as expected**

**Technology comes from the acquired Groq team; the Groq LP30 is fabbed by Samsung and is slated to ship in Q3**. A single Groq chip has 500 MB SRAM vs. 288 GB on a single Rubin GPU, so Groq alone cannot hold mainstream model parameters and KV cache.

**Solution: the Dynamo software splits inference into stages**:

**1\. Prefill**: batch processing of user prompts, compute-heavy and run on Vera Rubin. **2\. Attention during decode**: computes relationships between the current token and historical tokens (KV cache), balancing compute and memory and also run on Vera Rubin with frequent HBM reads.

**3\. Feed-forward network (FNN) during decode**: after attention resolves context, the FNN computes the next-token distribution and selects the output. Each layer must read model weights and, per read, processes only one token, creating a memory wall as compute stalls waiting for HBM.

By splitting decode, context memory stays on HBM while most model weights move to Groq’s on-die SRAM. SRAM can fetch weights at ultra-low latency, accelerating token emission in inference. Rubin and Groq are tightly coupled over Ethernet, and an RDMA-based connection cuts inter-chip latency by roughly half.

**6) Feynman**: next-gen GPU + LP40 (LPU) + Rosa CPU (named after Rosalind) + BlueField-5 + CX10. Kyber copper scale-up and Kyber CPO scale-up debut together, the first time both copper and CPO scale-ups are supported simultaneously. This means even in the Feynman phase, hybrid copper plus CPO will be supported.

**While NVIDIA is structurally bullish on CPO, customers prefer to push copper to its limits before switching to CPO**, given simpler deployment and maintenance.

**7) Other updates**:

**① Space data centers**: addressing energy constraints, NVIDIA announced Vera Rubin Space-1 to deploy data centers in space. This requires solving for radiative heat dissipation, as there is no conduction or convection in space.

**② OpenClaw**: every SaaS company will evolve into GaaS (Agent-as-a-Service). **Enterprise agents can access sensitive data, execute code, and communicate externally, requiring enterprise-grade security**. NVIDIA and OpenClaw founder Peter Steinberger launched **NemoClaw** (OpenClaw’s enterprise security reference design), integrating OpenShell tech with a network guardrail and a privacy router, and connecting to SaaS policy engines.

**③ Physical AI and robotics**: on autonomous driving, BYD, Geely, Hyundai, Nissan and others joined Robotaxi and partnered with Uber. In robotics, KUKA, ABB and several robot and UAV platforms were highlighted.

Overall, beyond clarifying that copper and CPO will co-exist, the main addition was an LPU option from Groq inside the server stack. This was well anticipated after the Groq acquisition, and even the three-year $1tn revenue marker is already below where the market sits.

NVIDIA’s recent product cadence shows the focus has shifted away from micro-architecture breakthroughs. From Hopper to Blackwell, the emphasis was on composition and connectivity, completing the transition from selling chips to selling systems and services.

From Blackwell to Rubin, the additions of DPU and the newly integrated LPU (SRAM) address the memory wall as AI moves into the inference and agent era. The goal is to relieve memory bottlenecks rather than chase raw core-level innovation.

**II. NVIDIA near term: muted guide, needs a new growth story**

NVIDIA’s shares have been range-bound at $170-200 over the past six months. Despite rising downstream capex and repeated beats, the stock has not broken out due to several concerns.

**a) Sustainability of hyperscaler capex**: Meta and Google both raised 2026 capex, and the four core cloud players could exceed $660bn in 2026 (+60% YoY). **Notably, capex as a share of revenue is already at elevated levels**.

**For Meta, 2026 capex is guided to $115-135bn, taking capex/revenue above 50%**, leaving limited room to push higher. Even with bigger 2026 envelopes, the market still worries about sustained growth beyond that.

**b) AI chip market share**: NVIDIA holds 75%+ share, and high pricing with a near-monopoly structure pushes clouds to seek alternatives. Beyond Google, AVGO has secured large orders from Anthropic and OpenAI, and multiple customers are pursuing in-house designs. Even with Rubin, consensus expects NVIDIA’s share to drift lower.

**3) Product competitiveness**: Google’s TPU v7 is roughly on par with NVIDIA’s B200 in FP8 domains (B200 mass production in Q4 2024), leaving TPU about one year behind. NVIDIA introduced NVFP4 in Blackwell to double inference vs. FP8, but FP8 already satisfies most current needs, making TPU v7 a viable alternative. To counter competition, NVIDIA is using strategic investments and capacity expansion to lock in the supply chain, e.g., a $30bn investment in OpenAI and $10bn in Anthropic tied to deployments, plus millions of GPUs for Meta’s new AI lab MSL, with some pricing concessions to secure demand.

Given these concerns, **valuation screens relatively undemanding**. Based on Dolphin Research’s est. $1.15tn data center revenue during 2025-2027 (above company guidance), the current $4.4tn market cap implies roughly 13x PE on FY2028 net income (close to 2027 calendar), assuming a 64% 2-yr revenue CAGR, 72% GPM, and 18% tax rate.

Despite a strong beat last quarter, the stock did not rally. That is because 2027 revenue is seen as largely priced in, and with capex/revenue already above 50% at key customers, further capex expansion looks constrained.

**In theory, as a second-derivative supplier to clouds, if customer capex holds flat at a high level, NVIDIA’s cloud revenue growth could fall to zero**. The market is hesitant to assign a high multiple beyond 2027, leaving NVIDIA at about 13x on 2027 profits and dampening appetite to build positions.

**From this GTC, the ‘$1tn+ cumulative data center revenue by 2027’ message is not new**. The market had already shifted to a higher number.

**More time was spent on product marketing and roadmap**, with greater implications for the supply chain (copper and CPO co-exist; LPU and HBM split workloads) than for NVIDIA’s own incremental drivers. Company-specific new information was limited.

**For a PE re-rating, beyond faster, broader AI app rollout, NVIDIA needs fresh growth curves**, such as ‘Physical AI’ and space-based compute.

**For the full GTC recap by Dolphin Research, see** [**‘NVIDIA (GTC Trans): LPU Unpacks Inference; Compute Factories Head for Space’**](https://longbridge.com/en/topics/39303796)

**Risk disclosure and disclaimer:** [**Dolphin Research Disclaimer and General Disclosure**](https://support.longbridge.global/topics/misc/dolphin-disclaimer)

### 相关股票

- [Samsung Electronics  (SSNGY.US)](https://longbridge.com/zh-CN/quote/SSNGY.US.md)
- [Vera Therapeutics (VERA.US)](https://longbridge.com/zh-CN/quote/VERA.US.md)
- [Bluebird Bio (BLUE.US)](https://longbridge.com/zh-CN/quote/BLUE.US.md)
- [NVIDIA (NVDA.US)](https://longbridge.com/zh-CN/quote/NVDA.US.md)
- [Jiangsu Azure Corporation (002245.CN)](https://longbridge.com/zh-CN/quote/002245.CN.md)

## 评论 (5)

- **格陵兰大 · 2026-03-17T23:52:50.000Z**: NVIDIA's costs are so low, why are the tokens of American large model companies so expensive?
- **ＴＳＬ · 2026-03-17T14:35:45.000Z**: Garbage.. losing money every day
- **MelodyLL · 2026-03-17T13:59:20.000Z**: No spoilsport
- **心向阳光5700 · 2026-03-17T07:37:10.000Z · 👍 1**: Not a bit disappointing at all
  - **Dolphin Research** (2026-03-17T11:30:46.000Z): There isn't much incremental information about the company itself.