NVIDIA GTC Preview: Potential Launch of New LPU Inference Chip
NVIDIA's GTC 2026, dubbed the "Oscars of AI," is scheduled to take place from March 16 to 19 in San Jose, USA. Market expectations are high that the "unprecedented" new chip previously teased by CEO Jensen Huang is highly likely to be a new AI inference chip integrating LPU technology from Groq, which NVIDIA acquired. The chip aims to specifically optimize inference speed and latency for AI models.
Next-Gen GPU "Feynman" Details Revealed: 1.6nm Process and High Power Consumption
At GTC, NVIDIA is expected to unveil its next-generation data center GPU architecture—the "Feynman" platform, named after physicist Richard Feynman. This chip will be the first to utilize TSMC's A16 process node and may feature backside power delivery. However, its power consumption is projected to exceed 1000W, demanding more advanced liquid cooling solutions. The product is targeted for a 2028 launch.
Computing Demand Shifts to Inference, Creating Definite Opportunities for Domestic Chips
With the explosion of AI Agent applications, the global computing demand structure is rapidly shifting from model training to model inference. In this context, domestic models like DeepSeek V4 are accelerating their migration to本土 computing platforms such as Ascend and Cambricon, opening up vast growth opportunities for Chinese AI chip manufacturers.
Domestic First: Optical Interconnect GPU Supernode "LightSphere" Achieves Commercialization
At AWE 2026, Shanghai INESA, in collaboration with Lightelligence, BirenTech, and ZTE, officially released the "LightSphere 128-node Commercial Edition". This solution, centered around Lightelligence's silicon photonic OCS optical switching chip and utilizing BirenTech's liquid-cooled GPU modules, represents a domestic breakthrough in original optical interconnect technology.
NVIDIA Invests Heavily ($4 Billion) in Optical Interconnect Technology
To address communication bottlenecks in future hyperscale AI computing centers, NVIDIA recently signed multi-year strategic agreements with leading optical communication firms Coherent and Lumentum. Beyond procurement commitments worth billions of dollars, NVIDIA will also invest $2 billion in each company to secure production capacity and technological R&D in optical communication products.
Domestic "Render+Inference" GPU Launched, Supporting 3A Games and LLMs
Domestic GPU manufacturer MetaX announced at AWE 2026 the official public sale of its "Render+Inference GPU" product line, the Lisuan eXtreme series, based on its proprietary TrueGPU architecture. This graphics card series can not only run 3A titles like Black Myth: Wukong smoothly but also support the local deployment of large models like DeepSeek.
NVIDIA Invests $2 Billion in AI Cloud Provider Nebius
NVIDIA announced on March 11 an investment of $2 billion in AI cloud service provider Nebius. This move aims to deepen collaboration across the full AI technology stack and further expand its AI computing ecosystem by investing in cloud infrastructure partners, offering customers more diverse access to computing power.
Computing Hardware Stocks Rally on GTC Expectations
In the A-share market this week, sectors related to computing hardware showed strong performance. As of the close on March 10, several stocks, including Everbright Photonics, hit their daily upside limit or rose over 10%. Market sentiment was primarily boosted by the anticipation of the upcoming GTC conference, with investors eagerly awaiting breakthroughs in next-gen GPU parameters like Rubin and "Feynman," as well as technologies like CPO switches and liquid cooling.
Domestic GPU Firms Release 2025 Financial Reports, Showing Explosive Potential
Recently, leading domestic GPU companies like Cambricon, Moore Threads, and MTT (MetaX) released their 2025 financial reports. While realistic gaps compared to international giants remain, these reports also reveal the astonishing explosive potential of domestic computing power, with the trillion-yuan market driven by domestic AI infrastructure capital expenditure opening doors for these enterprises.
NVIDIA Launches Open-Source LLM Nemotron-3-Super
On the eve of GTC, NVIDIA launched an open-source large model named Nemotron-3-Super on March 11. This move not only showcases its technical strength in the AI software layer but is also seen as a crucial step to further improve its CUDA ecosystem, attract developers, and provide underlying support for enterprise-level AI applications.