AMD has put a 72-GPU rackscale AI system at the center of its most direct challenge yet to Nvidia’s data-center dominance, pairing its new Instinct MI455X accelerators with EPYC “Venice” CPUs, Pensando networking, and an open-rack design called Helios. The announcement matters because it changes the terms of competition: AMD is no longer presenting an eight-GPU server alternative, but a coherent, exaFLOP-class AI infrastructure design intended to compete rack-for-rack with Nvidia’s Vera Rubin NVL72 platform. AMD’s Helios specifications describe a 72-accelerator system with up to 31TB of HBM4 memory, 2.9 exaFLOPS of FP4 compute, and 260TB/s of scale-up bandwidth.
For years, the AI hardware race has centered on individual accelerators: the GPU, its memory capacity, its compute throughput, and the software needed to use it. That framing is increasingly incomplete. The largest AI training jobs and the most demanding inference deployments depend on how dozens—or thousands—of accelerators behave as a single system.
Helios is AMD’s answer to that systems problem. Rather than merely shipping the MI455X as a component for server makers to integrate into conventional designs, AMD has created a rackscale reference architecture around 72 GPUs. The design includes EPYC “Venice” CPUs, Pensando Vulcano AI NICs, UALink connectivity, Ethernet-based scale-out networking, liquid-cooling provisions, and the company’s ROCm software environment. AMD’s product page says the architecture is built around open standards including Open Compute Project Open Rack Wide, UALink, and the Ultra Ethernet Consortium approach.
That is a consequential repositioning. Nvidia’s leadership has never been only about high-performance GPUs; it has been reinforced by the integration of GPUs, CPUs, networking, interconnects, management tools, and an enormous CUDA-centered software ecosystem. Nvidia’s own Vera Rubin NVL72 combines 72 Rubin GPUs, 36 Vera CPUs, NVLink 6, ConnectX-9 SuperNICs, and BlueField-4 DPUs as one tightly integrated platform. Nvidia’s Rubin announcement makes clear that the company is selling a complete AI factory building block rather than a loose collection of chips.
AMD’s Helios launch therefore represents more than another Instinct GPU release. It is an effort to give hyperscalers, cloud providers, model developers, and enterprise AI operators a credible second source for the kind of infrastructure previously associated most closely with Nvidia’s NVL systems.
The word coherent deserves attention here. In large language model workloads, model weights, context data, intermediate activations, and communication traffic must move rapidly among accelerators. A rack containing 72 isolated GPUs is not the same thing as a rack in which those GPUs can cooperate efficiently as a large logical compute resource.
AMD’s design attempts to make those 72 accelerators operate as a practical unit for training, fine-tuning, and inference. Independent reporting on the technical unveiling described Helios as AMD’s first rackscale platform that joins the GPU memory into a single coherent domain, a particularly important feature for inference tasks where data locality and memory bandwidth can determine whether raw accelerator performance translates into real-world output. Tom’s Hardware’s analysis notes that the design is AMD’s first true system-scale response to Nvidia’s NVL72 architecture.
The capacity comparison is strategically important. AMD claims the MI455X offers 50% more memory capacity than Nvidia Vera Rubin at the GPU level—432GB versus 288GB in AMD’s comparison—and slightly higher peak memory bandwidth. AMD’s comparison table lists 432GB and 23.3TB/s for MI455X, against 288GB and 22.0TB/s for Vera Rubin.
Aggregated across 72 accelerators, that works out to roughly 31TB of HBM4 for Helios. Reporting based on the launch materials places Nvidia Vera Rubin NVL72 at about 20.7TB of aggregate HBM, which makes AMD’s 50% capacity claim straightforward in arithmetic terms. Tom’s Hardware identifies memory capacity as one of Helios’s clearest potential advantages.
For frontier models, long-context reasoning, retrieval-heavy applications, and multi-agent workloads, HBM capacity is not a footnote. More local memory can reduce sharding complexity, lessen data movement, permit larger model partitions, and help operators handle longer context windows without turning memory constraints into the principal bottleneck.
The company’s flagship messaging is aggressive: Helios is designed to prove that the CPU, GPU, networking stack, and software stack should be evaluated together. That is a sound approach. An AI rack can be limited by host-side preparation, networking congestion, storage access, scheduling overhead, or memory movement even when its GPU peak-FLOPS figure looks exceptional.
That does not invalidate AMD’s 15% comparison. It means the headline cannot be treated as a universal declaration that Helios wins every workload by 15%. OCP MXFP4 and Nvidia’s NVFP4 are not labels that can be casually collapsed into a single apples-to-apples metric. Data types, sparsity assumptions, model architecture, context length, batch size, networking behavior, power envelope, and software maturity can all move the results.
AMD’s claim of as much as 30% more tokens per dollar is similarly significant but should be treated as an AMD-supplied performance-and-economics projection rather than an independent industry benchmark. Analysis of the launch materials notes that AMD’s internal tests cited 10% to 15% higher tokens per second per GPU on Kimi K2 Thinking and up to 30% more tokens per dollar, while correctly emphasizing that the comparison used AMD’s own figures against Nvidia’s published metrics. StorageReview’s launch analysis makes that distinction explicit.
For buyers, that means the right evaluation criteria are more demanding than peak performance:
That approach can appeal to major cloud providers and hyperscalers for several reasons:
That is why software remains central. AMD’s ROCm platform has improved its relevance in AI, and AMD positions ROCm.AI as the development layer for Helios-class deployments. AMD’s Advancing AI event page places ROCm.AI alongside Helios and MI455X as a core part of the company’s platform strategy. Yet the ecosystem breadth of CUDA, the availability of tuned frameworks, the depth of existing enterprise expertise, and Nvidia’s accumulated library of optimized AI infrastructure tools remain formidable competitive advantages.
This is a meaningful acknowledgment that AI inference is not one uniform workload. The “prefill” phase processes the user’s prompt and context, often favoring high compute capacity and large memory pools. The “decode” phase then generates tokens one at a time, creating different latency and memory-bandwidth pressures.
AMD and Cerebras say the combined architecture could provide up to 5x higher tokens per second per watt by directing different stages of inference to the hardware best suited for each one. That figure remains a vendor projection, and the companies have not disclosed all interconnect and end-to-end performance details. Tom’s Hardware reports that the first combined service is expected through Cerebras Cloud in the second half of 2026.
The partnership’s value is conceptual as much as technical. AMD is not positioning Helios merely as a substitute GPU rack. It is positioning Helios as a flexible infrastructure layer that can participate in heterogeneous AI factories, including deployments involving specialized inference silicon.
Reporting on that deal says Microsoft intends to deploy Helios at scale on Azure, though the companies have not publicly specified the exact number of racks, total power commitment, or financial scale. Tom’s Hardware notes that the size of the deployment was not disclosed.
That absence of a precise capacity number is not unusual at this stage, but it does make a difference. The AI accelerator market rewards volume execution. AMD must demonstrate not only that Helios performs in controlled comparisons, but that it can be manufactured in quantity, integrated reliably by partners, delivered to cloud providers, and operated at scale over years.
AMD’s announcement of a strategic partnership to deploy up to 2 gigawatts of MI450-series GPU capacity with Anthropic adds to the evidence that the company is gaining strategic AI customers. AMD’s announcement identifies the deployment target and underscores the scale of demand AMD is seeking to serve. Helios, with its MI455X and EPYC Venice components, is well positioned to become the physical system architecture through which some of that demand is fulfilled.
For enterprise IT leaders, a credible AMD rackscale AI platform creates several practical possibilities:
But Helios changes AMD’s competitive standing in a meaningful way. It gives AMD a large, technically coherent, high-memory AI rack with a credible systems architecture, customer momentum, an openness story, and performance claims that are strong enough to demand serious testing rather than dismissal.
The decisive question is no longer whether AMD can build an accelerator capable of challenging Nvidia. With Helios, the question is whether AMD can turn that hardware into a repeatable platform ecosystem: one that delivers measured performance, dependable software, rapid deployment, and compelling economics across real AI factories. If it can, the market for AI infrastructure will become less of a one-company contest—and enterprise buyers will be the beneficiaries.
Overview: AMD Moves From Accelerator Supplier to Rackscale System Architect
For years, the AI hardware race has centered on individual accelerators: the GPU, its memory capacity, its compute throughput, and the software needed to use it. That framing is increasingly incomplete. The largest AI training jobs and the most demanding inference deployments depend on how dozens—or thousands—of accelerators behave as a single system.Helios is AMD’s answer to that systems problem. Rather than merely shipping the MI455X as a component for server makers to integrate into conventional designs, AMD has created a rackscale reference architecture around 72 GPUs. The design includes EPYC “Venice” CPUs, Pensando Vulcano AI NICs, UALink connectivity, Ethernet-based scale-out networking, liquid-cooling provisions, and the company’s ROCm software environment. AMD’s product page says the architecture is built around open standards including Open Compute Project Open Rack Wide, UALink, and the Ultra Ethernet Consortium approach.
That is a consequential repositioning. Nvidia’s leadership has never been only about high-performance GPUs; it has been reinforced by the integration of GPUs, CPUs, networking, interconnects, management tools, and an enormous CUDA-centered software ecosystem. Nvidia’s own Vera Rubin NVL72 combines 72 Rubin GPUs, 36 Vera CPUs, NVLink 6, ConnectX-9 SuperNICs, and BlueField-4 DPUs as one tightly integrated platform. Nvidia’s Rubin announcement makes clear that the company is selling a complete AI factory building block rather than a loose collection of chips.
AMD’s Helios launch therefore represents more than another Instinct GPU release. It is an effort to give hyperscalers, cloud providers, model developers, and enterprise AI operators a credible second source for the kind of infrastructure previously associated most closely with Nvidia’s NVL systems.
What AMD Helios Brings to the Data Center
A 72-GPU coherent AI system
The basic Helios configuration is formidable. AMD specifies 72 Instinct MI455X GPUs in a full rackscale deployment, with four-GPU compute trays connected through a UALink-over-Ethernet fabric. The company rates the system at up to 2.9 exaFLOPS FP4 and 1.4 exaFLOPS FP8, alongside 31TB of HBM4 capacity and 43TB/s of scale-out bandwidth. AMD’s Helios architecture page lists those figures and describes 260TB/s of aggregate scale-up bandwidth inside the rack.The word coherent deserves attention here. In large language model workloads, model weights, context data, intermediate activations, and communication traffic must move rapidly among accelerators. A rack containing 72 isolated GPUs is not the same thing as a rack in which those GPUs can cooperate efficiently as a large logical compute resource.
AMD’s design attempts to make those 72 accelerators operate as a practical unit for training, fine-tuning, and inference. Independent reporting on the technical unveiling described Helios as AMD’s first rackscale platform that joins the GPU memory into a single coherent domain, a particularly important feature for inference tasks where data locality and memory bandwidth can determine whether raw accelerator performance translates into real-world output. Tom’s Hardware’s analysis notes that the design is AMD’s first true system-scale response to Nvidia’s NVL72 architecture.
MI455X: memory capacity is the headline specification
At the component level, the AMD Instinct MI455X is designed around AMD’s fifth-generation CDNA architecture. AMD lists up to 432GB of HBM4 memory per GPU and peak theoretical memory bandwidth of up to 23.3TB/s. It also rates the accelerator at up to 40 PFLOPS at FP4 and 20 PFLOPS at FP8. AMD’s MI400 Series specifications place the MI455X specifically within the Helios rackscale design rather than treating it as a conventional standalone accelerator.The capacity comparison is strategically important. AMD claims the MI455X offers 50% more memory capacity than Nvidia Vera Rubin at the GPU level—432GB versus 288GB in AMD’s comparison—and slightly higher peak memory bandwidth. AMD’s comparison table lists 432GB and 23.3TB/s for MI455X, against 288GB and 22.0TB/s for Vera Rubin.
Aggregated across 72 accelerators, that works out to roughly 31TB of HBM4 for Helios. Reporting based on the launch materials places Nvidia Vera Rubin NVL72 at about 20.7TB of aggregate HBM, which makes AMD’s 50% capacity claim straightforward in arithmetic terms. Tom’s Hardware identifies memory capacity as one of Helios’s clearest potential advantages.
For frontier models, long-context reasoning, retrieval-heavy applications, and multi-agent workloads, HBM capacity is not a footnote. More local memory can reduce sharding complexity, lessen data movement, permit larger model partitions, and help operators handle longer context windows without turning memory constraints into the principal bottleneck.
The CPU still matters in AI
AMD also used the Helios launch to emphasize the EPYC 9006 Series, code-named “Venice,” which is based on “Zen 6” and “Zen 6c” CPU cores. AMD positions the CPUs as a foundation for AI host nodes and agentic workloads, where large numbers of CPU threads may be needed to orchestrate requests, manage system state, run tools, process data, coordinate services, and support the parts of an AI pipeline that do not belong on a GPU. AMD’s EPYC 9006 overview describes the family as balancing higher-frequency cores with high core density for modern data-center and agentic AI workloads.The company’s flagship messaging is aggressive: Helios is designed to prove that the CPU, GPU, networking stack, and software stack should be evaluated together. That is a sound approach. An AI rack can be limited by host-side preparation, networking congestion, storage access, scheduling overhead, or memory movement even when its GPU peak-FLOPS figure looks exceptional.
The Direct Comparison With Nvidia Vera Rubin NVL72
AMD has not hidden the target. Helios is explicitly aimed at Nvidia’s Vera Rubin NVL72, and the two systems share the attention-grabbing 72-GPU configuration. Nvidia states that its NVL72 connects 72 GPUs in a non-blocking all-to-all NVLink topology, delivering 260TB/s of total GPU bandwidth and up to 3.6 exaFLOPS of AI compute. Nvidia’s NVLink documentation describes this fabric as a way to make the 72 GPUs function as a single high-performance accelerator.Where AMD’s claims are strongest
AMD’s hardware comparison has three particularly strong points:- HBM capacity: Helios’s 31TB total HBM4 capacity is the most tangible differentiator. It directly addresses workloads that become difficult or costly to partition across smaller memory pools. AMD’s Helios specifications
- Scale-out networking: AMD rates Helios at 43TB/s of scale-out bandwidth using Pensando networking and Ethernet-based technologies. That figure matters for clusters that must communicate beyond one rack, although actual deployment results will depend heavily on topology and software tuning. AMD’s Helios platform page
- Open infrastructure positioning: By basing Helios on Open Rack Wide and open interconnect standards, AMD is offering customers more architectural choice than a wholly proprietary stack may provide. AMD’s Helios overview
Where the benchmark language needs careful reading
The most important caution is that performance claims in this category are not automatically interchangeable. Nvidia rates Vera Rubin NVL72 at 3,600 PFLOPS of NVFP4 inference, while AMD rates Helios at 2.9 exaFLOPS of OCP MXFP4 compute. Nvidia’s Vera Rubin specifications and AMD’s Helios page use different low-precision formats and performance conventions.That does not invalidate AMD’s 15% comparison. It means the headline cannot be treated as a universal declaration that Helios wins every workload by 15%. OCP MXFP4 and Nvidia’s NVFP4 are not labels that can be casually collapsed into a single apples-to-apples metric. Data types, sparsity assumptions, model architecture, context length, batch size, networking behavior, power envelope, and software maturity can all move the results.
AMD’s claim of as much as 30% more tokens per dollar is similarly significant but should be treated as an AMD-supplied performance-and-economics projection rather than an independent industry benchmark. Analysis of the launch materials notes that AMD’s internal tests cited 10% to 15% higher tokens per second per GPU on Kimi K2 Thinking and up to 30% more tokens per dollar, while correctly emphasizing that the comparison used AMD’s own figures against Nvidia’s published metrics. StorageReview’s launch analysis makes that distinction explicit.
For buyers, that means the right evaluation criteria are more demanding than peak performance:
- What model is being run?
- What is the input and output context length?
- Is the workload training, prefill, decode, fine-tuning, or a mixed pipeline?
- How many racks are in the cluster?
- What is the delivered throughput at a fixed latency and power budget?
- How much engineering effort is required to reach that result?
Openness Is Helios’s Most Strategic Argument
Helios’s hardware specifications are impressive, but AMD’s broader pitch is about reducing dependence on a single proprietary AI platform. The company calls Helios a reference design, not a branded turnkey appliance sold directly as a finished AMD product. OEM and ODM partners can build their own systems around the architecture. AMD’s Helios FAQ says the design is a blueprint for partner-branded offerings built on Open Rack Wide standards.That approach can appeal to major cloud providers and hyperscalers for several reasons:
- They can customize power, cooling, networking, storage, and physical-rack configurations.
- They may avoid becoming fully locked into one vendor’s interconnect and software ecosystem.
- They can negotiate with a broader group of integrators and component suppliers.
- They can design clusters around Ethernet-based networking strategies already used elsewhere in their data centers.
- They gain more leverage in a market where AI infrastructure capacity has become strategically scarce.
That is why software remains central. AMD’s ROCm platform has improved its relevance in AI, and AMD positions ROCm.AI as the development layer for Helios-class deployments. AMD’s Advancing AI event page places ROCm.AI alongside Helios and MI455X as a core part of the company’s platform strategy. Yet the ecosystem breadth of CUDA, the availability of tuned frameworks, the depth of existing enterprise expertise, and Nvidia’s accumulated library of optimized AI infrastructure tools remain formidable competitive advantages.
The Cerebras Partnership Shows AMD Is Thinking Beyond the GPU
AMD also announced a partnership with Cerebras that could become as strategically interesting as the Helios rack itself. The plan combines AMD EPYC processors and Helios infrastructure with Cerebras’s Wafer-Scale Engine systems for a disaggregated inference platform. Tom’s Hardware’s report says the proposed design assigns prompt processing and large-context work to the AMD platform, while Cerebras hardware focuses on bandwidth-intensive, latency-sensitive token generation.This is a meaningful acknowledgment that AI inference is not one uniform workload. The “prefill” phase processes the user’s prompt and context, often favoring high compute capacity and large memory pools. The “decode” phase then generates tokens one at a time, creating different latency and memory-bandwidth pressures.
AMD and Cerebras say the combined architecture could provide up to 5x higher tokens per second per watt by directing different stages of inference to the hardware best suited for each one. That figure remains a vendor projection, and the companies have not disclosed all interconnect and end-to-end performance details. Tom’s Hardware reports that the first combined service is expected through Cerebras Cloud in the second half of 2026.
The partnership’s value is conceptual as much as technical. AMD is not positioning Helios merely as a substitute GPU rack. It is positioning Helios as a flexible infrastructure layer that can participate in heterogeneous AI factories, including deployments involving specialized inference silicon.
Microsoft, Anthropic, and the Importance of Customer Validation
Big infrastructure claims carry more weight when customers commit to deployment. AMD’s Advancing AI materials state that Microsoft and AMD are expanding their AI infrastructure partnership across Helios deployments, new Azure instances based on EPYC Venice, and AMD networking technologies. AMD’s event page frames the Microsoft relationship as a full-stack expansion rather than a limited processor supply arrangement.Reporting on that deal says Microsoft intends to deploy Helios at scale on Azure, though the companies have not publicly specified the exact number of racks, total power commitment, or financial scale. Tom’s Hardware notes that the size of the deployment was not disclosed.
That absence of a precise capacity number is not unusual at this stage, but it does make a difference. The AI accelerator market rewards volume execution. AMD must demonstrate not only that Helios performs in controlled comparisons, but that it can be manufactured in quantity, integrated reliably by partners, delivered to cloud providers, and operated at scale over years.
AMD’s announcement of a strategic partnership to deploy up to 2 gigawatts of MI450-series GPU capacity with Anthropic adds to the evidence that the company is gaining strategic AI customers. AMD’s announcement identifies the deployment target and underscores the scale of demand AMD is seeking to serve. Helios, with its MI455X and EPYC Venice components, is well positioned to become the physical system architecture through which some of that demand is fulfilled.
What Helios Means for Windows and Enterprise IT
Helios is not a consumer product, and it will not change the buying decision for a Windows 11 desktop or gaming PC. Its influence will still reach Windows-focused organizations because AI infrastructure decisions increasingly shape the services, copilots, developer tools, business applications, and cloud platforms that Windows users depend on.For enterprise IT leaders, a credible AMD rackscale AI platform creates several practical possibilities:
- More cloud choice: Azure and other cloud services may be able to offer additional accelerator options for AI workloads.
- Potentially better economics: If AMD’s tokens-per-dollar claims survive independent validation, organizations could benefit from lower inference cost for selected models and workloads.
- Reduced vendor concentration risk: A second large-scale GPU platform can improve supply resilience and commercial negotiating leverage.
- More open hardware paths: Open-rack and Ethernet-oriented designs may fit organizations that prefer less vertically constrained infrastructure.
- Broader model deployment flexibility: High HBM capacity could be attractive for long-context models, retrieval-augmented generation, document intelligence, and agentic workflows.
Conclusion: Helios Makes the AI Hardware Race More Real
AMD Helios does not end Nvidia’s AI infrastructure advantage. Nvidia retains immense strengths in software, ecosystem maturity, networking integration, customer familiarity, and production-scale deployment. Its Vera Rubin NVL72 architecture is itself a powerful rackscale system with 72 GPUs, 36 CPUs, 260TB/s of NVLink bandwidth, and a tightly integrated platform strategy. Nvidia’s Vera Rubin NVL72 specifications show why the company remains the benchmark AMD must beat.But Helios changes AMD’s competitive standing in a meaningful way. It gives AMD a large, technically coherent, high-memory AI rack with a credible systems architecture, customer momentum, an openness story, and performance claims that are strong enough to demand serious testing rather than dismissal.
The decisive question is no longer whether AMD can build an accelerator capable of challenging Nvidia. With Helios, the question is whether AMD can turn that hardware into a repeatable platform ecosystem: one that delivers measured performance, dependable software, rapid deployment, and compelling economics across real AI factories. If it can, the market for AI infrastructure will become less of a one-company contest—and enterprise buyers will be the beneficiaries.
References
- Primary source: aol.com
Published: 2026-07-23T18:30:00+00:00
AMD launches Helios system in direct challenge to Nvidia's AI dominance - AOL
AMD launched its Helios rackscale systems to rival Nvidia's Vera Rubin-powered platform.www.aol.com - Related coverage: tomshardware.com
AMD takes the wraps off its Instinct MI455X AI accelerator — CDNA 5 and Helios rack-scale architecture combine to take the fight to Nvidia in the data center | Tom's Hardware
Massive performance gains, higher memory capacity, and a large scale-up domain make for AMD’s strongest AI chip yetwww.tomshardware.com - Related coverage: itpro.com
AMD adds Microsoft to its growing roster of Helios customers as firm launches three new VMs for Azure users | IT Pro
Microsoft has announced it will deploy AMD’s Helios rack-scale solution as it targets data processing and AI inference gains for Azure customers.www.itpro.com