GF Securities Outlook AI Industry "Spring Festival Gala": Nvidia 2026 GTC Opens Next Month, System Breakthrough in Computing Power Bottleneck

IT Home, February 27 — GF Securities (Hong Kong) Brokerage Limited released a major preview report on February 25, providing an early look at NVIDIA’s 2026 GTC (GPU Technology Conference), revealing key highlights of this AI industry “Spring Festival Gala.”

The annual NVIDIA GTC event originally focused on GPU (Graphics Processing Unit) developer conferences, but with the rise of artificial intelligence, it has evolved into one of the most important annual industry trendsetters in the global AI sector.

The event will be held from March 16 (Monday) to March 19 (Thursday), 2026, in San Jose, California, with a keynote speech on the opening day.

This year’s GTC will shift focus from single-chip performance improvements to showcasing how NVIDIA employs a series of “combination punch” technologies to systematically address three core bottlenecks: AI inference, computing power expansion, and energy efficiency. Based on the blog content, the main highlights for NVIDIA at this event include:

• Rubin: This year, Rubin’s focus is no longer on “parameter showcase,” but on a “large-scale commercial deployment timetable.”
• Groq acquisition debut: After NVIDIA’s approximately $20 billion acquisition of Groq, GTC 2026 will be the first major showcase of the integrated entity.
• Feynman: This chip is expected to utilize advanced packaging (such as SoIC) with Logic-on-Logic 3D stacking, using TSMC’s A16 (1.6nm) process technology.

IT Home also provides the original report below: GTC 2026 Preview: Key Highlights on LPX, CPO, and PCB Latest Developments

Following our earlier GTC preview released on February 12, we now offer more detailed information about GTC 2026. We anticipate that the CEO’s keynote will cover topics including LPX (also called LPU), AI infrastructure distribution forecasts, VR200NVL72, NVL576, CPO, Alpamayo, among others. Nonetheless, we expect the focus to be on interference, enhanced expansion, and CPO.

Comments:

• LPX rack aims to strengthen NVIDIA’s product capabilities in inference: As mentioned, LPX (also known as LPU, Language Processing Unit) is expected to feature SRAM-based on-chip memory, enabling fast token generation and ultra-low latency, thereby elevating NVIDIA’s position in inference.

Before NVIDIA’s non-exclusive licensing agreement with Groq in December 2025, the rack design planned to include 64 Groq LPUs, using RealScale chip-to-chip interconnect technology.

For GTC 2026, the enhanced LPX rack is expected to host 256 LPUs distributed across multiple PTH PCBs based on 52L M9Q-glass, with each PCB valued at about $200.

Despite this, PCB manufacturers and liquid cooling (cold plate) companies are expected to benefit.

• VR200NVL72 to enhance NVIDIA’s product leadership and inference capabilities: Although VR200 is not new, Rubin is expected to leverage HBM4 (compared to Google v8AX’s HBM3e) and excellent system design to boost NVIDIA’s leadership, achieving 5x/3.5x performance improvements in inference/training, surpassing GB300.

We also believe Samsung’s HBM4 technology is making progress, helping Rubin stay on schedule. The company is expected to reaffirm plans for CPX used in inference pre-filling.

However, due to GDDR7 shortages, the CPX chip design is likely to switch to HBM4, with smaller capacity than the standard Rubin.

• NVL576 on mid-layer board with hybrid CCL: NVIDIA is expected to showcase its NVL576 architecture again, using mid-layer board designs.

As Rubin Ultra shifts to 448G Serdes technology, the mid-layer board is likely to use a hybrid of PTFE substrate and Q-glass M9 to improve signal transmission. Several options are under study, including:

• 78-layer PTFE mid-layer boards;
• 78-layer boards with 36 layers of PTFE and 41 layers of Q-glass IM9;
• 104/144-layer mid-layer boards.

Previously, we estimated the value of mid-layer content at about $25,000 per unit; with the introduction of higher-value PTFE-based CCL and more complex manufacturing requirements, we now expect the value to increase by at least 20-25%.

• NVL576 for optical interconnects replacing cable boxes: As mentioned in our “Key Moves for NVIDIA’s Scale Expansion in 2027” report (to be published in December 2025), NVIDIA plans to introduce CPO/NPO technology for Rubin Ultra NVL576 in late 2027.

Specifically, this applies to scale expansion interconnects within the NVL576 architecture. In this architecture, compute and switch trays are expected to continue relying on mid-layer connections, while inter-tray links may gradually shift to optical interconnects based on CPO or NPO. NVIDIA is expected to launch its Scale-Up optical interconnect solution at GTC 2026.

• CPO scale expansion into higher capacity domains and main beneficiaries: According to our “CPO Debate: Strengthening as NVIDIA Accelerates Its Roadmap” report, NVIDIA may launch a new generation of scale-expanding CPO switches (covering Ethernet and Infiniband markets).

We believe this CPO design significantly improves thermal performance, offering far better cost-effectiveness than previous generations. Suppliers are expected to accelerate deployment in late 2026/2027.

Driven by NVIDIA’s strong push and bundling strategies, we have raised our estimates for NVIDIA’s scale-expanding CPO switches to 20k/100k units (previously 20/80k), though penetration remains relatively low.

Main beneficiaries include suppliers of FAU, CW lasers, mixers, InP substrates, and connectors (such as Lumentum, Browave, Sumitomo, AXTI, and GLW).

We also expect Chinese FAU suppliers to become key providers, with optical modules/mixers benefiting from increased production.

Recommendations

• Buy NVIDIA (NVDA) due to strong recent quarterly performance, full VR200 lineup, and improved financial outlook for non-frontline CSPs (like OpenAI).

• Buy Lumentum (LITE), a leader in CPO and CW laser markets; Browave (3163TT, NR), as reported in multiple sources (e.g., “KeyMove plan for December 8 to become CPO”), with ASP (around $10,000) and market share of CPO mixers exceeding expectations.

• Taiflex (8039TT, NR) and leading local CCL suppliers are currently testing samples of PTFE-CCL. We remain optimistic about PCB products (per unit $30,000) and drilling processes (mid-layer ASP of over $2 vs non-drilled products at $0.2), which could significantly increase mid-layer content.

Risks

1. AI demand slowdown; 2) macroeconomic uncertainties; 3) competition.