Research Finder
Find by Keyword
The New Multi-vendor AI Triangle: Zyphra, AMD Silicon & ROCm Software, IBM Cloud
Zyphra ZAYA1 proves the trio can train frontier MoE models faster, cheaper, with reduced lock-in risk vs NVIDIA CUDA/NVLink
19/11/2025
Key Highlights:
- AMD announced that Zyphra successfully trained ZAYA1, a large-scale Mixture-of-Experts (MoE) foundation model using an all-AMD stack on IBM Cloud.
- The ZAYA1 model delivers competitive performance against leading open models across reasoning, mathematics, and coding benchmarks.
- The project leveraged AMD Instinct MI300X GPUs, Pensando networking, and the ROCm open software stack - the GPU's 192 GB of high-bandwidth memory enabled efficient large-scale training, avoiding complex sharding and achieving over 10x faster model save times.
- I believe that this collaboration demonstrates a multi-vendor approach solidifying AMD's position as a viable alternative to NVIDIA silicon and CUDA software - offering an option where enterprises face supply constraints driving high GPU pricing.
The News
AMD today announced that Zyphra successfully trained its ZAYA1, a large-scale Mixture-of-Experts (MoE) foundation model, entirely on AMD Instinct MI300X GPUs, Pensando networking, and the ROCm open software stack. This achievement showcases the scalability and efficiency of AMD's hardware for complex, production-scale AI workloads, with ZAYA1 delivering competitive performance against leading open models.
The collaboration between AMD, Zyphra, and IBM in system design, utilizing 128 compute nodes each outfitted with eight MI300X GPUs and eight Pensando Pollara 400 interconnects, set the foundation for robust, scalable training operations. This cluster, jointly engineered with IBM under a multi-year agreement announced in October 2025, leverages IBM Cloud's high-performance fabric and storage for enhanced security, reliability, and power efficiency in GenAI workloads. At the time, IBM Cloud GM Alan Peacock noted that, "Scaling AI workloads faster and more efficiently is a key differentiator in achieving ROI."
AMD announced this milestone in an independent release probably to highlight the company’s growing role in enabling advanced AI applications and efficiency at scale. Zyphra, valued at $1B after its Series A, is positioning itself as an open-source superintelligence lab. This joint accomplishment - backed by an AMD-stack on an IBM Cloud cluster - underscores the initiative's scale, with analysts viewing it as a catalyst for multimodal AI models that could disrupt proprietary training pipelines.
For more details, see the official AMD press release: https://ir.amd.com/news-events/press-releases/detail/1268/amd-powers-frontier-ai-training-for-zyphra
Analyst Take
From my perspective as a former tech executive, this announcement from AMD is one more arrow in the company’s quiver of proof points in the ongoing AI stack arms race. My initial observation is that AMD is working beyond alternative-source-competitor towards a solution democratizing access to high-performance AI training, thus offering a compelling alternative to the dominant NVIDIA ecosystem. The contrarian view here is that while AMD touts open software and competitive performance, the true test will be the ease of migration and the breadth of the developer ecosystem, which still heavily favors NVIDIA. Recent reports echo this, noting ROCm's HIP portability reduces lock-in but demands reworking custom kernels for AMD's architecture - reportedly slower than CUDA initially - prompting hybrid strategies among developers.
What Was Announced
AMD's announcement centers on Zyphra's successful training of the ZAYA1 MoE foundation model using a full AMD stack, including Instinct MI300X GPUs, Pensando networking, and the ROCm open software platform. The MI300X GPUs, with their 192 GB of high-bandwidth memory, are architected to facilitate efficient large-scale training by avoiding the need for complex expert or tensor sharding, which simplifies the training architecture and enhances throughput. Zyphra reported over 10x faster model save times due to AMD-optimized distributed I/O, further enhancing training reliability and efficiency. The ZAYA1-base model, with 8.3 billion total parameters and 760 million active, is designed to deliver competitive or superior performance to leading open models across various benchmarks, including reasoning, mathematics, and coding. This collaboration also involved IBM, which provided its Cloud's high-performance fabric and storage architecture for the large-scale training cluster. IBM's contributions, including its robust infrastructure for rapid GPU scaling, were pivotal in de-risking end-to-end pretraining on AMD, as highlighted in the October joint announcement.
Market Analysis
My analysis suggests this development significantly strengthens AMD's competitive positioning in the burgeoning AI infrastructure market. The all-AMD platform demonstration of training a large-scale MoE model provides the company with a tangible proof point for enterprises seeking NVIDIA-ecosystem alternatives. Especially in a market where challenging allocations drive sky-high GPU pricing. AMD's recent earnings (Q3 2025) reflect market interest in an alternative, the company’s data center revenue reached a record $4.3B (+22% YoY) from 5th Gen EPYC CPUs and Instinct accelerators. This exceeded estimates and from my perspective presages a similar full-year growth result amid ongoing hyperscale wins.
The AMD release about this achievement supports broader industry movement, based on a desire for open-source software and hardware flexibility to foster innovation while simultaneously reducing vendor lock-in and supply chain risks. AMD's UALink Consortium involvement spotlights this trend by promoting open interconnects for flexible AI infrastructure, demonstrated in collaborations with Astera Labs and Marvell. According to recent reports, the demand for AI accelerators is projected to increase significantly, with analysts expecting AMD's revenue to continue record-setting rises this year as demand for AI accelerators increases across data centers. This collaboration with Zyphra, and use of IBM's dedicated cluster - the first time a large-scale AMD full-stack training platform has been integrated on their cloud - validates a hybrid ecosystem approach, reducing vendor lock-in while accelerating multimodal AI innovation. It signals the ongoing strategic effort by AMD to build a robust ecosystem around its Instinct GPUs and ROCm software, challenging the established order. The focus on efficiency and scalability with fewer active parameters for competitive results illustrates AMD's edge in efficient scaling and throughput for AI training.
Looking Ahead
Based on what I am observing, the success of Zyphra's ZAYA1 model on AMD's platform and IBM’s cloud demonstrates the continued maturation of an AI hardware landscape where performance and efficiency must be available beyond a single vendor. My analysis of the market suggests that enterprises are increasingly prioritizing open, scalable, and cost-effective solutions for their AI workloads. Recent analyses suggest AMD-based stacks can deliver roughly 20–40% lower hardware cost in many configurations versus comparable NVIDIA systems, but at the expense of extra setup and tuning. The key trend I'll be monitoring is how AMD continues to attract developers into its software ecosystem by building the ROCm capabilities to truly rival the deeply entrenched CUDA incumbent. The company’s ongoing quest to credibly offer the complete high-performance stack, from chips to to software, is crucial as AMD continues moving beyond a component supplier to capture a larger share of the rapidly expanding AI market.
Stephen Sopko | Analyst-in-Residence – Semiconductors & Deep Tech
Stephen Sopko is an Analyst-in-Residence specializing in semiconductors and the deep technologies powering today’s innovation ecosystem. With decades of executive experience spanning Fortune 100, government, and startups, he provides actionable insights by connecting market trends and cutting-edge technologies to business outcomes.
Stephen’s expertise in analyzing the entire buyer’s journey, from technology acquisition to implementation, was refined during his tenure as co-founder and COO of Palisade Compliance, where he helped Fortune 500 clients optimize technology investments. His ability to identify opportunities at the intersection of semiconductors, emerging technologies, and enterprise needs makes him a sought-after advisor to stakeholders navigating complex decisions.