At the 2025 World Artificial Intelligence Conference (WAIC), Shenzhen-based AI pioneer Intellifusion officially announced its strategic shift to focus solely on AI inference chips. The company aims to establish a domestic computing power ecosystem centered around three core scenarios: edge computing, cloud-based large model inference, and embodied intelligence. Leveraging a decade of chip R&D experience, Intellifusion is poised to capitalize on the vast opportunities presented by the AI inference era.

Intellifusion is no stranger to the chip landscape. Over the past decade, it has continuously iterated and upgraded its NPU chips, progressing from the first-generation edge SoC Nova100 in 2015 to the latest DeepEdge10 series in 2023. The DeepEdge10 series, which has garnered significant attention, represents the pinnacle of domestic chip technology. It employs a domestic 14nm Chiplet process and RISC-V cores, pioneering the 'Compute Bricks' architecture that enables flexible chip combinations akin to building blocks, spanning a compute power range of 8T to 256T. Tailored for diverse scenarios, the series comprises four chip products: the compact 10C for AIoT devices, the 10 Standard and Max versions for edge multimodal large models, and the Edge200 specialized in accelerating inference for hundred-billion-parameter large models.

Intellifusion's decisive 'All In' approach is underpinned by four pioneering achievements in domestic process technology: the first D2D Chiplet commercial design, the first C2C Mesh MoE inference architecture, the first 128T compute power commercial platform, and the first AI chip integrating domestic RISC-V cores. These technological breakthroughs have been successfully applied and yielded tangible results. Currently, the DeepEdge series chip platform supports nearly 10 mainstream large models, including DeepSeek-R1 and Tongyi Qianwen, achieving widespread adoption in robots, edge gateways, and servers. For instance, the Shenqiong X6000 accelerator card, equipped with DeepEdge200, can effortlessly run hundred-billion-parameter models. The DeepSeek inference all-in-one machine, based on this chip, achieves an industry-leading throughput of 500 tokens/s when running the full 671B MoE model.

Industry data highlights the AI industry's rapid transition from the 'training era' to the 'inference era'. An IDC and Inspur Information joint report reveals that AI inference workloads accounted for 67% in 2024 and are projected to rise to 73% by 2028. "Training is akin to generating electricity, while inference is like using it," metaphorically explained Chen Ning, Chairman of Intellifusion. "In the future, there may be just a handful of foundational large models, but the number of terminals accessing these models will surge exponentially." This trend is evident in the terminal market, where intelligent devices, from Tesla's in-car Grok4 voice interaction system to AI learning cameras with sales exceeding 50,000 units, increasingly rely on inference chips.

Intellifusion's robust growth momentum underpins its strategic transformation. Financial reports indicate that the company's revenue reached 900 million yuan in 2024, marking an 81.3% year-on-year increase. In the first quarter of 2025, the growth rate surpassed 160%. In June this year, Intellifusion secured a 1.6 billion yuan compute power service order, providing approximately 4,000 PFLOPS of heterogeneous compute power. The announcement of its secondary listing plan in Hong Kong in early July underscores its ambition to venture into the international market. Looking ahead, Intellifusion is fully dedicated to developing the second-generation 'Compute Bricks' architecture, aiming to elevate the compute power of a single chip to several thousand TOPS through innovations such as near-memory computing and 3D hybrid bonding, continuing to spearhead the domestic AI inference chip landscape.