AMD has launched a new generation of AI chips and infrastructure designed to compete directly with NVIDIA, and Chinese AI companies like DeepSeek, Qwen, and Moonshot AI are among the first to deploy them at scale. At the AMD Advancing AI Conference on July 23, CEO Lisa Su announced the Instinct MI455X GPU, the Venice server CPU, and the Helios rack-scale AI infrastructure, all built on cutting-edge 2nm and 3nm process technology. The announcement reveals a significant shift in how enterprise AI is being deployed, with open-weight models from Chinese developers playing a central role in AMD's strategy.

What Makes AMD's New Hardware Different From NVIDIA?

AMD's Helios rack represents a fundamental rethinking of AI infrastructure design. According to independent testing by Signal65, AMD's previous-generation MI355X GPU approaches or even exceeds NVIDIA's B200 in real-world workload throughput while offering better cost-performance. The new MI455X GPU packs 320 billion transistors using TSMC's 2nm and 3nm process technologies and includes 432GB of HBM4 memory, a type of ultra-fast memory designed specifically for AI workloads.

When comparing Helios directly to NVIDIA's Vera Rubin rack, AMD measured significant advantages across multiple dimensions. Helios achieves 20 terabytes per second of memory bandwidth, 20 petaFLOPS of computing power in FP4 precision, and 190 gigabytes per second of scale-out bandwidth, which AMD executives described as "the highest measured performance numbers ever publicly released by a GPU manufacturer". On the CPU side, AMD's Venice processor delivers 2.2 times the throughput performance of NVIDIA's Vera CPU and leads by 20 percent in per-core performance without any optimization tuning.

Why Are Chinese AI Companies Adopting AMD Hardware?

The most striking aspect of AMD's announcement is the prominent role of Chinese AI models and companies. During the conference, AMD demonstrated its ROCm.ai software stack running MiniMax M3 and DeepSeek-V4 Pro, two of the most advanced open-weight models developed in China. The company also highlighted partnerships with major Chinese tech firms including ByteDance, Tencent, Alibaba, and Moonshot AI, as well as support for models like Kimi, Xiaomi MiMo, GLM, and Qwen.

This partnership strategy reflects a broader trend in enterprise AI deployment. AMD conducted internal testing on two real-world scenarios: autonomous threat detection and running personal AI agents. By routing some requests to cutting-edge cloud models and others to open-weight models running on AMD's EPYC CPUs and MI350P GPUs, the company achieved a 43 percent reduction in token costs and increased response speed for on-premise workloads by up to 3 times. This hybrid approach, combining proprietary cloud models with locally-deployed open-weight models, is becoming the standard way enterprises manage AI costs and latency.

How to Deploy Open-Weight Models Efficiently on Enterprise Hardware

• Hybrid Architecture Strategy: Combine cutting-edge cloud models for complex reasoning tasks with open-weight models running locally on enterprise servers, reducing token costs by approximately 43 percent while maintaining response speed.
• Memory and Compute Matching: Use AMD's MI350P GPU with 144GB of HBM3E memory to support models with up to approximately 260 billion parameters, enabling deployment of large open-weight models like DeepSeek-V4 Pro without cloud dependency.
• Cost-Performance Optimization: Leverage AMD's rack-scale infrastructure to achieve 4.2 times better token generation per dollar per second compared to NVIDIA's RTX Pro 4000, making open-weight model deployment economically viable for enterprises.

The MI350P represents a particularly important development for enterprises seeking to run open-weight models locally. This air-cooled GPU features 144GB of HBM3E memory, approximately 4 terabytes per second of memory bandwidth, and a maximum power draw of around 600 watts. A single card can support models with up to approximately 260 billion parameters, making it suitable for running large open-weight models like Qwen or DeepSeek without relying on cloud infrastructure.

What's Next for AMD's AI Roadmap?

AMD has outlined an aggressive three-year roadmap for AI infrastructure. The MI500 series GPU is expected to launch next year with inference throughput surging to over 2,000 times that of the MI300X, along with next-generation HBM memory and new copper and optical interconnect technologies. The MI600 series is currently in development and scheduled to launch in 2028.

On the CPU side, the seventh-generation Venice successor, called Florence, will adopt a new generation of process nodes and memory, shipping in 2028. The eighth-generation Ravenna is in development and expected to arrive in 2030. For rack-scale infrastructure, Helios 500 will arrive next year featuring the Verano CPU and MI500 series GPUs, while Helios 600 is scheduled for 2028 with the Ferrara CPU and MI600 series GPUs.

AMD also announced a technical partnership with Cerebras Systems to create a decoupled AI inference solution combining AMD's Helios rack with Cerebras' wafer-level engine. When the two computing engines work together, tokens per watt per second are expected to increase by 5 times, addressing one of the most pressing challenges in AI deployment: energy efficiency.

For on-premise local AI deployment, AMD launched Gorgon Halo, a high-end upgrade of the Ryzen AI Max 400 series featuring 192GB of unified memory and capable of running 300-billion-parameter large models locally. The company also introduced the Ryzen AI Embedded X100 series processors and the Kria AI robot developer platform, signaling AMD's ambition to support the full spectrum of AI applications, from data center inference to edge computing and robotics.

The prominence of Chinese open-weight models in AMD's announcements underscores a significant shift in the global AI landscape. As enterprises seek to reduce costs and maintain data privacy, locally-deployed open-weight models from companies like DeepSeek, Qwen, and Moonshot AI are becoming increasingly competitive with proprietary cloud-based alternatives. AMD's hardware and software stack are now optimized to support this deployment pattern, potentially reshaping how enterprises approach AI infrastructure investment.