Challenging NVIDIA's Monopoly! Startup Upscale Secures $200 Million in Funding, Self-Developed Chip Directly Targeting NVSwitch

In the AI hardware competition, the network architecture connecting GPUs has become a new battleground. According to AIbase, AI network startup Upscale AI recently announced a $200 million Series A funding round. This round was led by Tiger Global and Premji Invest, aiming to challenge NVIDIA's dominance in rack-level AI system switches.

Upscale AI was established last year, backed by major supporters such as Intel, AMD, and Qualcomm. The company's core product is a custom ASIC chip called "SkyHammer." According to AIbase, this chip is designed for AI workloads and is intended to directly compete with the NVSwitch chips in NVIDIA's NVL72 rack. Although giants like Cisco and AMD are also trying to introduce alternatives (such as UALink), their technologies are still in early stages. However, Upscale plans to bring its products to market later this year.

To improve compatibility, Upscale adopts a memory-semantic load-store network architecture and supports two competing protocols, UALink and ESUN. According to AIbase, the company also plans to reduce management complexity for hyperscale customers by supporting the open-source network operating system SONiC. CEO Barun Kar said that with the new funding, the company will focus on delivering and deploying its products, aiming to achieve "de-NVIDIAization" in the critical area of AI network interconnection.

Key Points:

💰 Massive Funding Entry: Upscale AI secured $200 million in funding to develop an AI network switch chip that challenges NVIDIA's NVSwitch.
🚀 Self-developed SkyHammer Chip: Redesigning the network architecture for AI workloads, planning to launch customized silicon products supporting the UALink protocol within the year.
🌐 Breaking the Closed Ecosystem: By supporting the open-source system SONiC and various industry protocols, teaming up with allies like AMD and Qualcomm to counter NVIDIA's NVLink closed ecosystem.

2.4 Trillion-Parameter Giant Arrives! Baidu Launches ERNIE Bot 5.0: Native Full-Modal Technology Reshapes AI Interaction

Baidu officially launched ERNIE Bot 5.0, featuring 2.4 trillion parameters, achieving a transition from multimodal fusion to native full-modal. It adopts native full-modal unified modeling technology, jointly training text, images, videos, and audio within a unified architecture, differing from the industry's common post-hoc integration approach.

Rise of Enterprise AI Computing Infrastructure: 1.5-Year ROI, Hardware Configuration Becomes Critical - Kingston Launches Full-Stack Solution

As AI becomes an essential tool for enterprises, building dedicated computing infrastructure has become a trend. More and more companies are choosing to build their own local AI workstations instead of relying on cloud-based APIs, mainly due to considerations of data security, cost control, and business stability. The initial investment in self-built hardware can typically be recouped within 1.5 to 2.5 years, showing significant economic advantages. Enterprises need to match different scales of computing solutions based on the complexity of their tasks.

Challenging NVIDIA's Monopoly! Startup Upscale Secures $200 Million in Funding, Self-Developed Chip Directly Targeting NVSwitch

Related Recommendations

2.4 Trillion-Parameter Giant Arrives! Baidu Launches ERNIE Bot 5.0: Native Full-Modal Technology Reshapes AI Interaction

Rise of Enterprise AI Computing Infrastructure: 1.5-Year ROI, Hardware Configuration Becomes Critical - Kingston Launches Full-Stack Solution

Kimi from Moonshot AI Makes Debut at Davos: Surpassing U.S. Proprietary Models with 1% Computing Power, Engineering Thinking Becomes the Key for China's AI Breakthrough

Microsoft Releases Rho-alpha Model: Giving AI Robots Human-like Flexibility and Understanding

Saying No to AI Spam Infringement: 800 Creative Professionals, Including Scarlett Johansson, Join a Joint Protest Against Technological Theft