NVIDIA’s rival, AMD, has asserted itself in the AI race again. The company has partnered with Meta to deploy up to 6 gigawatts of AMD Instinct GPUs to power the social media giant’s next phase of AI infrastructure build-out.
The first phase will deliver 1 gigawatt of capacity, with shipments beginning in the second half of 2026. The deployment will feature a custom AMD Instinct GPU based on the MI450 architecture, paired with 6th Gen AMD EPYC ‘Venice’ CPUs and powered by the ROCm software.
The infrastructure will be built on AMD’s Helios rack-scale architecture, co-developed with Meta through the Open Compute Project.
This follows Meta’s recent announcement of a multi-year partnership with NVIDIA, raising an obvious question: what sets AMD’s rack apart?
As AI infrastructure discussions increasingly focus on scale, sovereignty, and openness, AMD sees its Helios rack-scale architecture as a foundation for the next generation of AI infrastructure.
In an exclusive interaction with AIM, Archana Vemulapalli, Corporate Vice President of Global Commercial Sales at AMD, and Mahesh Balasubramanian, Senior Director of Data Centre GPU Product Marketing, discussed the company’s strategy around open ecosystems and integrated AI systems.
Vemulapalli says Helios emerged from AMD’s work within the Open Compute Project (OCP), an initiative launched by Meta in 2011 to make data centre hardware designs open and shareable. At the time, proprietary hardware designs drove up costs and slowed down innovation. OCP aimed to change that.
From a hardware standpoint, Vemulapalli mentions that Helios scales up to 72 AMD Instinct MI400-series GPUs (such as MI450 or MI455X) in a single rack and delivers 2.9 exaFLOPS of FP4 performance for inference and 1.4 exaFLOPS of FP8 for training.
https://youtu.be/PKYNu3LMOng
Open Software
AMD’s open philosophy goes beyond hardware infrastructure. “That’s not only on the hardware side. It’s also on the software side,” Balasubramanian says.
He further adds that AMD’s ROCm software stack is built on open components, giving both proprietary and open model developers full-stack access. This allows companies to innovate at every layer of the system while contributing improvements back to the broader ecosystem, so others can build on that work and accelerate innovation further.
Speaking about ROCm’s competition with CUDA, Balasubramanian says developers do not need additional layers of code to run models on ROCm. “There’s no coding required. Everything runs out of the box,” he says, encouraging developers to try the platform themselves.
He adds that models available on platforms like Hugging Face are supported on ROCm from day one, including newly launched versions. The open software stack, he notes, supports frameworks such as PyTorch and JAX, ensuring immediate compatibility.
The India Play
AMD recently partnered with TCS to bring Helios to India. Through its subsidiary, HyperVault AI Data Centre Limited, TCS will work with AMD to co-develop a rack-scale AI infrastructure design built on the Helios platform to support India’s national AI initiatives.
“TCS has done amazingly well. Just look at their track record in the services space. They’ve really set the industry standard among many other great Indian companies in terms of how to run services globally,” Vemulapalli says.
She adds that TCS is now taking a thoughtful approach towards shaping India as an AI hub, with a strong focus on building data centre capacity.
Highlighting the company’s broader experience, including work across Tata Sons brands, she added, “For us, it’s about taking their experience and expertise and bringing our experience and expertise together to service both global as well as Indian customers as best we can.”
Training vs Inference
Globally, AMD expects inference to outpace training in the coming years.
“Globally, we do see that the inference market is going to outpace training,” Balasubramanian says.
In India, that trend may be even more pronounced. “There’s a lot of conversation about building SLMs, not LLMs. With the population that can benefit from AI, inference is going to be a very, very critical part of that.”
Speaking about agentic workloads, Balasubramanian says AMD has a strong structural advantage in x86 computing. “AMD has a huge advantage on x86 with our EPYC CPUs,” he says, adding that most enterprise applications, including CRM and ERP systems, already run on it.
Talking about new MCP servers coming up, he explains that EPYC CPUs accelerate data access, while adding Instinct GPUs on top of that stack further strengthens performance.
Speaking of its competition, Intel, he says, “At the end of the day, there are only two major players in x86. One is Intel, which does not have GPUs. One is AMD, which has GPUs.” He adds that AMD’s Linux capabilities, performance, power and area efficiencies, and full-stack solutions provide further advantages.
Balasubramanian explains that AMD is developing solutions to run agentic AI workloads directly on consumer devices. As an example, he points to Liquid AI’s demonstration at the Consumer Electronics Show, where agentic workloads ran on personal hardware rather than in the cloud.
Liquid AI’s Liquid Foundation Models are built to be efficient, using less memory and reducing computing overhead while allowing quick fine-tuning for specific tasks. Running on AMD hardware, these models demonstrate how advanced AI capabilities can move beyond large data centres and function effectively on consumer systems.
“We build our compute in all these different form factors and dimensions, and for different use scenarios,” Vemulapalli concludes.
Your reaction
Discussion
What's your take on this story?
No comments yet — be the first to share your view.