AMD(AMD.US) Stepping into the era of rack-level AI infrastructure! Strong alliance with Tianzhou Technology(CLS.US) to build Helios computing cluster

Bloomberg News has learned that AMD (AMD.US) has announced a major partnership with Celestica (CLS.US) to deeply collaborate on bringing AMD’s new Helios rack-scale AI computing infrastructure platform to the global AI data center market, targeting the NVIDIA NVL72 rack-level AI platform. For AMD, which aims to steadily chip away at NVIDIA’s (NVDA.US) dominant 90% share in the multi-trillion-dollar AI core computing cluster market, Helios is crucial for its revenue and profit prospects. AMD is gradually shifting its competitive focus toward full cabinet-level systems similar to NVIDIA’s NVL72, with the large-scale launch of Helios AI clusters by the end of 2026 essentially positioning itself in direct competition with NVIDIA’s “rack-level AI infrastructure.”

Both companies stated in a release that upon the platform’s launch, Celestica will be responsible for the research, design, and manufacturing of the vertically scalable high-performance network switches within AMD’s Helios rack-scale AI cluster architecture.

AMD Helios is a comprehensive, high-performance, open-standard AI rack-level infrastructure platform designed specifically for ultra-large-scale AI training and inference tasks in data centers. Rack-scale AI architecture is currently the most popular form of clustered computing, where the entire rack—rather than individual CPU/GPU servers—serves as the fundamental computational unit for massive AI workloads. It integrates core computing capabilities such as AI GPUs/ASICs, high-performance network architecture, and liquid cooling units into a single AI infrastructure system, enabling efficient training of large language models (LLMs) or handling enormous AI workloads based on large models.

The two companies also indicated that these vertically scalable switches will utilize the most advanced network chip architectures to enable high-speed interconnection among the next-generation AMD Instinct MI450 series AI GPUs, providing cutting-edge computational capabilities optimized for large-scale AI infrastructure clusters.

“Helios rack-scale AI solutions represent a new blueprint for AI computing infrastructure, allowing customers to deploy large-scale AI data centers with the performance, efficiency, and flexibility needed for the next generation of enormous AI workloads,” said Forrest Norrod, Executive Vice President and General Manager of AMD Data Center Solutions.

They also noted that they are collaborating to support one-click, high-efficiency deployment of Helios in cloud platforms, enterprise organizations, and large research environments. Thanks to the latest developments in scaling Helios production capacity, on Monday, after the US stock market closed, Celestica’s stock rose about 3%, AMD’s stock temporarily increased over 3%, and ultimately closed up 1.7%.

It is expected that AMD Helios rack-scale AI infrastructure will begin mass shipments to major cloud clients such as Microsoft and Amazon by the end of 2026.

Joint Forces to Counter NVIDIA’s Blackwell Series

As AMD and Celestica accelerate the market rollout of the Helios rack-scale AI platform, they are also working together with several leading tech giants to oppose NVIDIA’s vertically integrated AI infrastructure solutions. Previously, AMD announced partnerships with Hygon and Broadcom to supply open, rack-scale AI infrastructure for high-performance computing clusters and large AI data centers, aiming to accelerate global “Sovereign AI” research.

Hygon will be among the first system suppliers to adopt AMD’s Helios rack-scale AI architecture. Based on deep collaboration with ASIC and high-performance network infrastructure leader Broadcom, AMD and Hygon will integrate a customized high-performance scale-up switch within HPE Juniper Networking racks. This large AI computing system aims to simplify deployment of larger AI infrastructure clusters and offer a more cost-effective and energy-efficient alternative to NVIDIA’s Blackwell series.

Essentially, Helios is AMD’s response to NVIDIA’s Blackwell NVL72/GB200 NVL72 line of rack-level AI infrastructure. Both are built around 72 GPUs+CPUs+high-speed interconnects+liquid cooling+rack-level system engineering as the fundamental units for AI workloads, rather than individual servers. AMD defines Helios as an open rack architecture based on OCP Open Rack Wide, targeting large-scale training and inference; NVIDIA’s GB200 NVL72 is a liquid-cooled rack platform consisting of 36 Grace CPUs and 72 Blackwell GPUs. In other words, Helios is not just “another batch of MI450 GPUs,” but AMD’s first true full-rack system to directly compete with NVIDIA’s NVL72 ecosystem.

Compared to AMD’s previous AI GPU series, Helios offers a significant performance leap. AMD’s official benchmarks state that Helios can deliver up to 36 times the performance of the previous generation AMD AI computing platform. This indicates AMD’s shift from selling high-performance GPU cards to selling entire AI factory racks—packaging GPUs, CPUs, NICs, liquid cooling, network topology, and ROCm into a complete AI solution.

The biggest selling points of Helios are memory capacity and open interconnects. AMD states that the 72-GPU Helios can deliver up to 2.9 exaFLOPS FP4, 1.4 exaFLOPS FP8, 31TB HBM4 memory, 1.4PB/s aggregate memory bandwidth, and 260TB/s scale-up interconnect bandwidth. Compared to NVIDIA’s GB200 NVL72, Helios is more aggressive in memory capacity, raw scale-up bandwidth, and open rack design, making it more attractive for long-context, large-parameter models, and bandwidth-sensitive training/inference systems. AMD even claims that Helios’s memory capacity exceeds NVIDIA’s next-generation Vera Rubin platform by 50%.

Why Did AMD Choose Celestica?

AMD’s partnership with Celestica is driven by practical reasons: the bottleneck in rack-scale AI systems is no longer just the GPU but also high-speed scale-up switches, liquid cooling engineering, manufacturing yield, delivery capabilities, and supply chain resilience.

AMD explicitly states that Celestica is responsible for the research, design, and manufacturing of Helios scale-up networking switches based on UALoE, which are critical for the stable operation of MI450 clusters in large-scale AI environments. Celestica’s value lies not just in manufacturing ordinary components but in bridging AMD’s transition from a chipmaker to a systems provider—engineering and productizing open compute cluster architectures and delivering them according to hyperscalers’ requirements. This aligns with NVIDIA’s recent emphasis on full systems, high-performance networking, and operational software integration.

Celestica, headquartered in Canada, is an electronics manufacturing services (EMS/ODM) and infrastructure solutions provider. It not only handles traditional hardware assembly but also plays a key role in designing, manufacturing, and integrating AI data center infrastructure products such as network switches, servers, rack-level solutions, and high-bandwidth network components. As cloud providers and large tech companies like Google, Meta, and Amazon massively build AI data centers, the demand for high-speed network connectivity, custom hardware, and rack-level integrated solutions has surged. Celestica is a core supplier of high-performance network switches, servers, ASIC/TPU modules, and integration services for these large data centers. The company’s stock price increased by 220% in 2025.