Nvidia GTC: The Spring Festival Gala of AI, Full of Expectations, Disappointed to Leave?

On March 16, 2026, NVIDIA Founder and CEO Jensen Huang delivered a keynote speech at GTC 2026, covering topics including 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:
Projected total revenue for data centers from 2025 to 2027 reaches $1 trillion (last year’s GTC estimated $500 billion for 2025-2026), meeting expectations. The market consensus has already raised expectations above $1 trillion, with more anticipation for the company to provide clear order and pipeline 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 also 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 into tiers:
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 that by 5 times.

4) Vera Rubin:
Built upon six previous chip classes, with the addition of Groq 3 LPU.
① Vera Rubin: 100% liquid-cooled (hot water cooling at 45°C), all cables removed, reducing installation time from two days to two hours;
② CPO (Co-packaged Optical) Spectrum-X Switch: Fully mass-produced, co-developed with TSMC;
③ CPU: The only data center CPU using LPDDR5, sold independently, becoming a billion-dollar business;
Vera CPU Tray for Agentic workloads, integrating 8 Vera processors, each with 88 cores, supporting 8-channel LPDDR5x memory, with 1.2TB/s memory bandwidth per socket. The CPU Tray includes 2 BF4-DPU chips.
④ Vera Rubin: Deployed and operational on Microsoft Azure (first rack). NVIDIA’s supply chain can produce thousands of systems weekly, with gigawatt-level AI factory capacity monthly.
⑤ Rubin Ultra: Rubin is a horizontal rack insert; Rubin Ultra is designed to be vertically inserted into new Kyber racks, with 144 GPUs within a single NVLink domain, replacing copper cables with NVLink switches behind the midplane.

5) Groq 3 LPU (New Chip):
Uses both Groq and HBM, as expected.
Technology from the acquired Groq team, with Groq LP30 manufactured by Samsung, expected to ship in Q3.
One Groq chip has 500MB SRAM, while a Rubin chip has 288GB; Groq alone cannot handle large models’ parameters and KV Cache.

Solution:
A software called Dynamo decomposes inference steps into stages:
1. Pre-fill stage: Also called Prefill, processes user prompts in batches, mainly computation, done on Vera Rubin;
2. Attention decoding: Calculates the relationship between current tokens and historical tokens (KV Cache), involving heavy compute and storage, also on Vera Rubin, frequently reading HBM memory;
3. Feedforward network (FNN): After establishing context in the Attention stage, the FNN predicts the next token based on previous tokens, generating speech.
Each layer reads model weights, processing one token at a time. These weights are stored in HBM, but the compute units wait for data transfer from HBM, creating the “memory wall.”

By splitting decoding into software-defined stages, most model parameters are transferred from HBM to Groq’s SRAM, enabling low-latency access and solving slow inference 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 cables scale-up + Kyber CPO scale-up (supporting both copper and CPO simultaneously). Even at the Feynman stage, hybrid support for copper and CPO is maintained.
While NVIDIA remains optimistic about CPO, customers prefer 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 (solving radiation and cooling issues, as space has no conduction or convection, only radiation).
② OpenClaw:
Every SaaS company will become a 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, and others are joining Robtaxi, collaborating with Uber.
In robotics, KUKA, ABB, and many drone platforms are involved.

Overall, this conference clarified that copper and CPO will be used together, and a new server option with Groq’s LPU was introduced. After Groq’s acquisition, market expectations were high; the projected $1 trillion revenue over three years has already been exceeded.

From NVIDIA’s product iterations, recent focus has shifted from microarchitecture innovation to integration and connectivity issues—moving from Hopper to Blackwell, completing the transition from chip sales to system and service sales.
From Blackwell to Rubin, the addition of DPU (NAND chips) and the urgent integration of LPU (SRAM) mainly address AI inference and agent era challenges, particularly the memory wall.

2. NVIDIA’s Current Status: Conference Guidance is Moderate, 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 better-than-expected earnings, the stock has not broken out, mainly due to market concerns:

a) Sustained Capital Spending by Major Firms:
Meta, Google, and others have announced increased 2026 capital expenditures, with the four largest cloud providers expected to spend over $660 billion, up 60%. However, these expenditures are already a high percentage of their revenues.
For example, Meta expects $115-135 billion in 2026, with capex over 50% of annual revenue, leaving limited room for further increases. Despite the outlook, market remains wary of continued growth in capital spending.

b) Market Share in AI Chips:
NVIDIA currently holds over 75% of the AI chip market, with high prices and near-monopoly status prompting cloud providers to seek alternatives.
Beyond Google, Broadcom (AVGO) has secured large orders from Anthropic, OpenAI, and others, with many customers developing in-house solutions. Even with new Rubin products, the market expects NVIDIA’s share to gradually decline.

3) Product Competitiveness:
Google’s TPUv7 approaches NVIDIA’s B200 (mass production in Q4 2024), lagging about a year.
NVIDIA’s Blackwell introduces NVFP4 format, doubling inference performance over FP8, but FP8 already meets most market needs, making TPUv7 a viable alternative.
To counter industry competition, NVIDIA is investing strategically and expanding compute capacity—supporting OpenAI ($30 billion), Anthropic ($10 billion), and providing hundreds of thousands of GPUs to Meta’s new AI labs, with some agreements involving price locks to secure customer demand.

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

Despite strong quarterly earnings, the stock has not risen, mainly because the market fears that once 2027 revenue expectations are priced in, further growth from cloud capital expenditure will be limited.
In theory, NVIDIA’s valuation is based on cloud customers’ high capex, but if that remains flat, revenue from cloud clients will stagnate, leading to a PE ratio of only 13x for 2027 profits, reducing investor interest.

From the GTC keynote, Huang emphasized that “by 2027, data center revenue will exceed $1 trillion,” but market expectations are already higher.
Most of the conference time was spent on product promotion and roadmap, with less new incremental information about the company itself.

Looking ahead, NVIDIA’s PE could rise again if AI applications scale faster and more broadly, or if new growth drivers emerge, such as “Physic AI” or “Space Computing Power.”