Before Jensen Huang's keynote on March 16, we examine the investment implications of Vera Rubin, silicon photonics, and the $690B CAPEX cycle.
NVIDIA GTC 2026 will be held from March 16-19 at the McEnery Convention Center in San Jose. Jensen Huang CEO's keynote will begin on Monday, March 16 at 11:00 AM (PT), and is estimated to be attended by up to 27,000+ people from over 190 countries worldwide. FACT
In an interview just before GTC, Jensen Huang foreshadowed that he would unveil "a chip the world has never seen before." Over 700 sessions will cover Physical AI, AI Factory, Agentic AI, inference optimization, and more, with major announcements also scheduled from global partners such as Deutsche Telekom, SK Group, and Oracle. FACT
1. Detailed Unveiling of Vera Rubin Architecture — TSMC 3nm process, equipped with HBM4 memory, 50 PFLOPS based on FP4 (2.5x compared to Blackwell). Mass production entry was confirmed at CES 2026, and it is expected to reduce Agentic AI inference token costs by 1/10 compared to Blackwell. FACT
2. Blackwell Ultra GB300 Performance Update — GB300 AI server mass production is underway, and mass shipments have begun through Supermicro and others. 3.6 million units in backlog, sold out until mid-2026. FACT
3. Silicon Photonics / 3D CPO — NVIDIA has developed the world's first 3D stacked silicon photonics engine. The Quantum-X InfiniBand 115Tb/s (144 ports × 800Gb/s) switch is scheduled for release in early 2026, and the Spectrum-X Ethernet CPO is scheduled for release in the second half of 2026. 5x improvement in power efficiency, 10x improvement in network resilience. FACT
4. Feynman Architecture Roadmap — TSMC A16 1.6nm process, first silicon photonics on-chip application. Scheduled for release in 2028, but the possibility of early sample exhibition at GTC is being raised. RUMOR
5. Physical AI / Robotics / Autonomous Driving — Cosmos World Model, Isaac Lab-Arena, Omniverse Digital Twin, Alpamayo Autonomous Driving Portfolio, etc. are covered as key topics. FACT
The U.S. economy is in a late expansion phase. The S&P 500 is maintaining near all-time highs at 6,781, but Late Cycle signs are appearing, such as small-cap to large-cap rotation being detected. FACT
Fed benchmark interest rate is maintained at 3.50~3.75%, and the probability of a freeze at the March FOMC (3/18) is over 92% based on CME FedWatch. The market is pricing in 2 rate cuts (50bp) within 2026, but the timing has retreated to the second half of the year. FACT
The January CPI recorded the lowest in 8 months at +2.4% YoY, and the core CPI is also continuing its downward trend at +2.5%. The consensus for the February CPI to be announced today (3/11) is +2.5%, and the impact of tariffs may begin to be reflected. FACT
The combined CAPEX of the Big 5 (Microsoft, Alphabet, Amazon, Meta, Oracle) in 2026 will increase by +36% YoY to $660~690B. The proportion of AI-dedicated infrastructure amounts to approximately 75%, or $450B. FACT
However, hyperscalers are investing approximately 90% of their operating cash flow into CAPEX, resulting in the highest capital intensity on record. The possibility of Amazon's FCF turning negative and the forecast of Meta's FCF decreasing by 90% raise questions about sustainability. NARRATIVE
If even one of the hyperscalers lowers its CAPEX guidance, or if the February CPI surges to 2.8% or higher, the macro signal will turn bearish.
The Trump administration converted the export policy of NVIDIA H200 to China to "case-by-case review" in January 2026, imposing a 25% tariff + 50% volume cap. China has approved the import of 400,000+ units of H200 to ByteDance, Alibaba, Tencent, etc., and the order size is estimated to be up to $14 billion. FACT
A Trump-Xi Jinping summit is scheduled for 3/31~4/2 in Beijing, so there is a possibility that the AI semiconductor export license system will be re-examined. NVIDIA has conservatively not included China data center sales in its Q1 FY27 guidance. FACT
NVIDIA recorded FY26 revenue of $215.9B (+65% YoY) and data center revenue of $197.3B (+75%). The Q1 FY27 guidance of $78B exceeds AMD's annual revenue on a quarterly basis alone. The AI GPU market share is estimated to be approximately 80~90%. FACT
AMD MI400 is scheduled for release in 2026 with 432GB HBM4, FP4 40 PFLOPS specs. While touting a 1.5x advantage in memory capacity, the CUDA ecosystem gap remains a key barrier to entry when competing directly with Vera Rubin (FP4 50 PFLOPS). Meanwhile, AMD has increased its presence by signing a large-scale contract with Meta worth $60 billion. FACT
The Custom ASIC camp is growing rapidly. Broadcom's Q1 AI revenue was $8.4B (+106%), and Custom ASIC grew by +140%. Google Ironwood TPU v7 (2026 H2), Amazon Trainium3 ($10B+ run rate), and Microsoft Maia 200 are all emerging as GPU alternatives. FACT
However, the growth of Custom ASIC reflects the expansion of the overall AI compute market pie rather than replacing NVIDIA. CAPEX $690B is a scale that can accommodate both GPUs and ASICs, and NVIDIA's general-purpose GPU + CUDA ecosystem remains irreplaceable for startups, enterprises, and research institutions. INFERENCE
Bottleneck #1: TSMC CoWoS Packaging (Most Severe) — Expansion to 127,000 wafers per month is scheduled by the end of 2026, but it is insufficient compared to demand. NVIDIA occupies 55% of the total capacity, making it difficult for competitors to secure supply. Expansion takes 18~24 months, making short-term resolution impossible. FACT
Bottleneck #2: HBM Memory (Severe) — SK Hynix dominates the market with 50%+, with all 2026 supply sold out and prices increased by