admin The Helios system, unveiled at AMD’s Advancing AI event at San Francisco’s Moscone Center, is now in full production and is scheduled to begin shipping before the end of the third quarter of 2026. It marks AMD’s most ambitious effort to sell complete computing systems rather than individual chips to data centre operators.
Helios combines 72 Instinct MI455X graphics processing units with sixth-generation EPYC processors, codenamed Venice, and AMD’s Pensando networking technology. The design allows thousands of accelerators to work together on the training and operation of large artificial intelligence models.
AMD said the platform could deliver as much as 30 per cent more inference performance per dollar than Nvidia’s competing Vera Rubin NVL72 system under selected workloads. Such comparisons depend heavily on software, power, model size and deployment configuration, making independent testing important once commercial systems become available.
Chief executive Lisa Su described rack-scale computing as the central battleground for the next phase of artificial intelligence. Data centre customers are increasingly buying integrated clusters containing processors, accelerators, memory, networking equipment and software rather than selecting each component separately.
The shift has favoured Nvidia, whose tightly connected hardware and CUDA software ecosystem became the industry standard for training foundation models. AMD is seeking to counter that advantage with an open architecture supporting established programming frameworks and networking standards.
OpenAI plans to deploy Helios systems at scale during 2026 and 2027, providing AMD with a prominent customer capable of validating the platform under demanding workloads. Microsoft, Meta Platforms, Oracle and several cloud infrastructure operators have also expanded their use of AMD accelerators.
Anthropic has agreed to deploy up to two gigawatts of AMD’s MI450-generation computing capacity, starting with one gigawatt during the first half of 2027. AMD could invest as much as $5 billion in the artificial intelligence company as deployment targets are met.
The agreement includes joint engineering work aimed at optimising Anthropic’s Claude models for AMD hardware. Such partnerships can help chipmakers improve software performance while giving AI developers additional sources of computing capacity.
AMD also announced a partnership with Cerebras Systems to support disaggregated inference, an approach that assigns different stages of an AI workload to specialised processors. Helios systems would handle high-throughput processing, while Cerebras hardware could accelerate latency-sensitive generation tasks.
The joint service is expected to become available through Cerebras Cloud later this year. Splitting workloads across platforms could reduce costs, though it creates technical challenges involving data movement, scheduling and communication between chips from different suppliers.
The MI455X accelerator is manufactured by Taiwan Semiconductor Manufacturing Company and supports high-bandwidth memory designed to handle large AI models. Venice processors use AMD’s Zen 6 architecture and are intended to improve the number of simultaneous tasks that can be completed within each server rack.
AMD is positioning the processor as a throughput-focused alternative to Nvidia’s Arm-based Vera central processing unit. Nvidia has emphasised faster single-threaded performance and lower latency, while AMD is promoting the ability of its x86 chips to handle larger numbers of parallel requests.
Su expects the global computing market to approach $2 trillion by 2030 as artificial intelligence spreads across cloud services, robotics, industrial systems and autonomous software. AMD separately estimates that the market for AI accelerators could reach about $1.4 trillion by the end of the decade.
The company’s roadmap includes annual updates to its rack-scale products, with newer Instinct accelerators and server processors planned for each generation. Florence processors are expected in 2028, followed by the Ravenna family around 2030.
AMD’s challenge extends beyond hardware performance. Nvidia’s CUDA platform has been used by developers for nearly two decades and is supported by a broad library of tools, applications and trained engineers. Moving workloads to another platform can require extensive software changes and testing.
AMD has accelerated development of its ROCm software stack, improving support for popular AI frameworks and large language models. The company has also acquired specialised software and infrastructure businesses to strengthen its ability to deliver complete systems.
Its $4.9 billion acquisition of ZT Systems added server design and data centre integration expertise. Other purchases have expanded AMD’s capabilities in networking, compilers and artificial intelligence software.
Nvidia remains the dominant supplier of advanced AI accelerators and continues to introduce new systems on a rapid schedule. Its scale, software advantage and relationships with major cloud providers present formidable barriers for challengers.
