NVIDIA GTC: The AI industry's Spring Festival Gala — full of anticipation, but leaving disappointed?

On March 16, 2026, NVIDIA Founder and CEO Jensen Huang delivered a keynote speech at GTC 2026, covering topics such as the 20th anniversary of the CUDA platform, inflection points in inference and explosive computing power demand, Vera Rubin system architecture, Groq integration, OpenClaw proxy revolution, and physical AI and robotics.

1. Key Points of GTC 2026

1) Data Center Revenue Outlook: From 2025 to 2027, cumulative data center revenue is expected to reach $1 trillion (last year’s GTC forecast was $500 billion for 2025-2026), meeting expectations. The market consensus has already risen above $1 trillion, with more anticipation for the company to provide clear order information.

2) Performance and Cost: In terms of tokens/watt (throughput) and token speed (intelligence), NVIDIA ranks as the highest globally; its token cost is the lowest worldwide.

3) Data Centers as “Token Factories”: Each factory is limited by power (e.g., 1GW) and needs to manage token production throughput and speed.

Tokens will be subdivided like commodities: free tier (high throughput, low speed) -> $3/million tokens -> $6/million tokens -> $45/million tokens -> $150/million tokens (top-tier low latency, high bandwidth compute).

For example, a 1GW data center allocates 25% power to each tier: Grace Blackwell can generate 5 times the revenue of Hopper, Vera Rubin can increase it fivefold.

4) Vera Rubin: Building on six previous chip classes, a new Groq 3 LPU has been added.

① Vera Rubin: 100% liquid-cooled (hot water cooling at 45°C), all cables removed, installation time reduced from two days to two hours;

② CPO (Co-packaged Optical) Spectrum-X switch: fully mass-produced, co-developed with TSMC;

③ CPU: The world’s only data center CPU using LPDDR5, sold separately, becoming a billion-dollar business;

Vera CPU Tray for Agentic workloads: each Vera Compute Tray integrates 8 Vera processors, each with 88 cores, supporting 8-channel LPDDR5x memory, with a single socket supporting 1.2TB/s memory bandwidth. The CPU Tray also includes 2 BF4-DPU modules.

④ Vera Rubin: Launched and running on Microsoft Azure (first rack). NVIDIA’s supply chain can produce thousands of systems weekly, with gigawatt-level AI factory capacity per month;

⑤ Rubin Ultra: Rubin is a horizontal slide-in cabinet, while Rubin Ultra is designed to be vertically inserted into new racks like Kyber, with 144 GPUs within an NVLink domain, replacing copper cables with NVLink switches behind the mid-plane.

5) Groq 3 LPU (New Chip): Using both Groq and HBM, as expected.

Technology from the acquired Groq team, with Groq LP30 manufactured by Samsung, expected to ship in Q3.

A single Groq chip has 500MB SRAM vs. Rubin’s 288GB; Groq alone cannot handle large models’ parameters and KV Cache.

Solution: Introduced Dynamo software to decompose inference steps:

1. Pre-fill stage: Also called Prefill, processing batch user prompts mainly via computation, done on Vera Rubin;

2. Attention decoding: Calculating relationships between current tokens and historical tokens (KV Cache), involving both compute and storage, also on Vera Rubin, frequently reading HBM memory;

3. Feedforward network (FNN): After establishing context via attention, the FNN predicts the next token based on previous tokens, “articulating” the output.

Each layer reads model weights; a single read handles one token. Parameters stored in HBM, with compute units waiting for data transfer—this is the true “memory wall.”

By splitting decoding into software-defined stages, the model’s “context memory” remains in HBM, while most parameters are transferred to Groq’s SRAM, enabling low-latency access and solving inference speed issues.

Rubin and Groq are tightly coupled via Ethernet, with RDMA connections reducing interaction latency by about half.

6) Feynman: New GPU + LP40 (LPU) + Rosa CPU (named Rosalind) + BlueField-5 + CX10.

Kyber copper cable scale-up + Kyber CPO scale-up (supporting both copper and CPO simultaneously). Even at the Feynman stage, hybrid support for copper and CPO is planned.

Although NVIDIA favors CPO long-term, customers tend to maximize copper cable deployment before switching to CPO for easier deployment and maintenance.

7) Other Information:

① Space Data Centers: Addressing energy shortages, NVIDIA announced Vera Rubin Space-1, planning to deploy data centers in space (radiation shielding and heat dissipation are challenges, as space has no conduction or convection, only radiation);

② OpenClaw: SaaS companies will become GaaS (Agent-as-a-Service). Proxy systems can access sensitive info, execute code, and communicate externally—requiring enterprise-grade security. NVIDIA partnered with OpenClaw founder Peter Steinberger to launch NemoClaw (enterprise security reference design), integrating OpenShell tech, including network guardrails and privacy routers, connecting to SaaS policy engines;

③ Physical AI and Robotics: In autonomous driving, BYD, Geely, Hyundai, Nissan, etc., join Robtaxi, collaborating with Uber. In robotics, KUKA, ABB, and many drone platforms are involved.

Overall, this launch clarified that copper and CPO will be used together, and a new server option with Groq LPU has been added. After Groq’s acquisition, market expectations have been high; even the three-year $1 trillion revenue target has been exceeded.

From NVIDIA’s product iterations, recent years focus less on microarchitecture innovation, shifting from Hopper to Blackwell mainly on integration and connectivity. NVIDIA has transitioned from chip sales to system and service sales.

From Blackwell to Rubin, the addition of DPU (NAND chip) and the urgent inclusion of LPU (SRAM) mainly address AI inference and agent era, tackling the memory wall.

2. NVIDIA’s Current Status: The Conference Outlook Is Mild, Need for a “Growth Story”

NVIDIA’s stock has hovered between $170 and $200 over the past six months. Despite increased capital expenditure from major cloud providers and consistently strong earnings, the stock has not broken out upward, mainly due to market concerns:

a) Sustained Capital Spending by Major Firms: Meta, Google, etc., have announced increased 2026 capital expenditure, with top four cloud giants expected to spend over $660 billion in 2026, up 60%. However, their share of revenue is already high.

For example, Meta expects $115-135 billion in capital expenditure in 2026, over 50% of annual revenue, with limited room for further increase. Despite optimistic outlooks, market worries about continued growth in capital spending persist.

b) AI Chip Market Share: NVIDIA currently holds over 75% of the AI chip market. Its high prices and near-monopoly structure prompt cloud providers to seek alternatives.

Beyond Google, Broadcom AVGO has secured large orders from Anthropic, OpenAI, and others, with multiple clients developing in-house solutions. Even with new Rubin products, market expects NVIDIA’s market share to gradually decline.

3) Product Competitiveness: Currently, Google TPUv7’s performance in FP8 and related areas is roughly close to NVIDIA’s B200 (mass production in Q4 2024), lagging about a year behind.

NVIDIA’s Blackwell introduces NVFP4 format, doubling inference performance over FP8. But FP8 already meets most market needs; TPUv7 is effectively an alternative.

To counter industry competition, NVIDIA is expanding capacity through strategic investments and supply chain locking, such as $30 billion in OpenAI, $10 billion in Anthropic, and providing hundreds of thousands of GPUs to Meta’s new MSL AI lab—some agreements include price locks to secure customer demand.

Given these market concerns, NVIDIA’s valuation remains relatively low. Based on data center revenue of $1.15 trillion (above the company’s guidance of $1 trillion) for 2025-2027, and a current market cap of $4.4 trillion, the projected PE ratio for FY2028 (close to 2027) is about 13x, assuming a 64% two-year CAGR in revenue, 72% gross margin, and 18% tax rate.

NVIDIA reported strong quarterly earnings, but the stock did not rise. This is because, after fully pricing in 2027 revenue expectations, the market worries that with cloud providers’ capital expenditure exceeding 50%, further increases are limited.

In theory, if cloud providers maintain high capital spending, NVIDIA’s revenue from cloud customers would stagnate, leading to a low valuation for post-2027 earnings. Currently, NVIDIA’s PE is about 13x, with little enthusiasm for further growth.

Looking at GTC’s content, Huang’s projection of “over $1 trillion in data center revenue by 2027” is already priced higher by the market.

Most of the conference time was spent on product promotion and strategic planning, affecting the supply chain (mixing CPO and copper, LPU and HBM sharing different tasks), with limited new information about the company’s own growth.

For NVIDIA’s PE to rise again, beyond larger-scale AI deployment, new growth curves are needed, such as “Physic AI,” “Space Computing,” etc.

Source: Dolphin Research

Risk Disclaimer

Market risks are present; investments should be cautious. This article does not constitute personal investment advice and does not consider individual user goals, financial situations, or needs. Users should evaluate whether the opinions, views, or conclusions herein are suitable for their circumstances. Investment is at their own risk.