Tsinghua-Backed AI Infra Startup Raises $100M+ to Rebuild Computing Around GPUs

RongXin ZhiYuan Secures Massive Angel Round for GPU-First Architecture

RongXin ZhiYuan, a Beijing-based AI infrastructure startup founded by Tsinghua University alumni, has closed an angel funding round worth several hundred million yuan (estimated at over $100 million). The company is building what it calls the AGC architecture — an AI computer system with the GPU as its core — designed to fundamentally replace the traditional CPU-centered computing paradigm that has dominated for decades.

The round was co-led by the Beijing Green Energy and Low-Carbon Industry Fund and SAIF Partners (Sail Capital). Additional investors include Shunxi Fund, Fuhua Capital, Wanlida Group, Yangtze River Innovation Investment, Shuimu Tsinghua Alumni Fund, and Plum Ventures. Cloud Capital, which participated in the company's earlier seed round, continued its investment and serves as the company's exclusive long-term financial advisor.

Key Facts at a Glance

• Company: RongXin ZhiYuan (容芯致远), headquartered in Beijing
• Funding: Angel round worth hundreds of millions of yuan (~$100M+)
• Core innovation: AGC (AI computer system with GPU as its Core) architecture
• Founder: Shi Xu, Tsinghua University Electronic Engineering graduate with years of experience in chip design and AI
• Problem targeted: Traditional CPU-centered architectures bottleneck AI workloads, limiting GPU utilization and scalability
• Investors: SAIF Partners, Beijing Green Energy Fund, Plum Ventures, and others

Why Traditional CPU-Centered Architectures Are Breaking Down

The explosive growth of AI workloads — from training large language models to running inference at scale — has exposed fundamental limitations in how modern servers are designed. Traditional data center architectures place the CPU at the center of all data scheduling and interaction. Every piece of data flowing between GPUs, memory, and storage must pass through the CPU, creating a critical bottleneck.

In a typical AI server deployment, multiple CPUs are required to coordinate and schedule a relatively small number of GPUs. As clusters scale up, the number of CPUs must increase proportionally, driving up both system complexity and cost. This architectural mismatch becomes increasingly painful as organizations deploy thousands of GPUs for frontier model training.

Founder Shi Xu highlighted this challenge in a recent interview: 'In actual deployment, the traditional architecture simply cannot adapt to the computing demands of the AI era. The CPU becomes the core constraint on data scheduling, GPU-to-GPU communication efficiency is insufficient, and memory cannot achieve unified address space sharing — resulting in low overall compute utilization.'

The AGC Architecture: Putting GPUs at the Center

RongXin ZhiYuan's answer to these challenges is the AGC architecture, which stands for 'AI computer system with the GPU as its Core.' Unlike the conventional approach where CPUs orchestrate all system operations, AGC repositions the GPU as the primary compute and scheduling engine.

This is not merely an incremental optimization. It represents a fundamental rethinking of how AI computing systems are structured. The architecture aims to address several critical pain points simultaneously:

• Eliminating CPU bottlenecks in data scheduling and inter-GPU communication
• Enabling unified memory address spaces across GPU clusters for seamless data sharing
• Improving overall GPU utilization rates, which currently hover at 30-50% in many large-scale deployments
• Reducing the need for excessive CPU provisioning as GPU clusters scale
• Lowering total system cost and complexity for AI infrastructure operators

The concept aligns with a broader industry trend toward GPU-centric or accelerator-centric computing. Companies like NVIDIA have been moving in this direction with technologies such as NVLink and Grace Hopper Superchips, which tightly couple GPUs and reduce reliance on traditional CPU interconnects. However, RongXin ZhiYuan appears to be pursuing this vision at the full system architecture level rather than focusing solely on chip-level integration.

How This Fits Into the Global AI Infrastructure Race

The timing of this funding round reflects the intense global competition to build next-generation AI computing infrastructure. In the United States, companies like NVIDIA, AMD, and startups such as Cerebras and Groq are all working to optimize how AI workloads are processed. NVIDIA's dominance — with its H100 and upcoming B200 GPUs — has made GPU architecture the central battleground of the AI era.

In China, the stakes are equally high. U.S. export restrictions on advanced AI chips have forced Chinese companies and research institutions to rethink their approach to AI infrastructure from the ground up. Rather than simply acquiring the most powerful GPUs available, Chinese startups like RongXin ZhiYuan are exploring how to maximize the performance of available hardware through architectural innovation.

This 'systems-level thinking' approach has historical precedent. Google's development of the TPU (Tensor Processing Unit) demonstrated that purpose-built architectures could outperform general-purpose GPU setups for specific AI workloads. RongXin ZhiYuan's AGC architecture follows a similar philosophy — optimizing the entire system stack rather than relying solely on more powerful individual chips.

The company's Tsinghua University pedigree also matters. Tsinghua has emerged as one of China's most prolific sources of AI and semiconductor talent, producing founders behind companies like Zhipu AI, MiniMax, and numerous chip design startups. The university's electronic engineering department, where founder Shi Xu studied, has deep expertise in chip architecture and systems design.

What This Means for the AI Industry

RongXin ZhiYuan's approach has significant implications for multiple stakeholders in the AI ecosystem:

For cloud providers and data center operators, a GPU-centric architecture could dramatically reduce the total cost of ownership for AI infrastructure. If GPU utilization rates can be pushed from the current 30-50% range toward 70-80% or higher, the economic impact would be substantial — potentially saving billions of dollars across the industry.

For AI model developers, better GPU utilization and more efficient inter-GPU communication could accelerate training times and reduce costs. Training a frontier large language model currently costs tens of millions of dollars, with a significant portion of that expense attributable to inefficient hardware utilization.

For the semiconductor industry, the rise of GPU-centric architectures may reshape demand patterns. If systems require fewer CPUs per GPU, it could shift the balance of power between CPU manufacturers like Intel and AMD and GPU-focused companies like NVIDIA.

Key implications include:

• Cost reduction: Lower CPU-to-GPU ratios mean lower overall system costs
• Performance gains: Unified memory and direct GPU communication could boost training throughput by 2-3x
• Scalability: GPU-centric designs may scale more efficiently to clusters of 10,000+ GPUs
• Energy efficiency: Fewer idle CPU cycles means lower power consumption per unit of useful compute
• Competitive dynamics: Chinese companies could partially offset chip restrictions through superior system architecture

The Funding Landscape for AI Infrastructure Startups

RongXin ZhiYuan's angel round — reportedly worth several hundred million yuan — is notably large for an early-stage company. This reflects the enormous appetite among investors for AI infrastructure plays, particularly in China where the sector is viewed as strategically critical.

Globally, AI infrastructure startups have attracted record levels of venture capital in 2024 and 2025. In the U.S., CoreWeave raised $7.5 billion in its IPO, while Lambda Labs secured $800 million in funding. Chinese AI chip companies like Cambricon and Enflame have also raised significant capital.

The participation of the Beijing Green Energy and Low-Carbon Industry Fund is particularly noteworthy. It signals that GPU-centric architectures are being viewed not just as performance optimizations but as sustainability plays — more efficient compute means lower energy consumption per AI operation, a growing concern as data center power demands surge worldwide.

Looking Ahead: Challenges and Opportunities

While RongXin ZhiYuan's vision is ambitious, the company faces significant challenges. Building a fundamentally new computing architecture requires deep expertise across chip design, system software, networking, and developer tools. The company will need to demonstrate that its AGC architecture delivers meaningful real-world performance gains, not just theoretical improvements.

Competition is fierce from both established players and well-funded startups. NVIDIA's own push toward tighter GPU integration through technologies like NVSwitch and GB200 NVL72 racks already addresses some of the same pain points. AMD's Instinct MI300X takes a chiplet-based approach to integrating compute and memory.

However, if RongXin ZhiYuan can deliver on its GPU-centric vision, the market opportunity is enormous. The global AI infrastructure market is projected to exceed $300 billion by 2027, and even a small share of that market would represent a massive business.

The next milestones to watch include the company's first product announcements, benchmark results comparing AGC performance to traditional architectures, and partnerships with major cloud providers or AI companies in China. With substantial funding now secured, RongXin ZhiYuan has the resources to move from concept to prototype — and potentially reshape how the world builds AI computing systems.