AMD unveils Helios, its AI rack to challenge Nvidia

AMD wants to shift competition from the chip to the rack

AMD no longer wants to be seen only as Nvidia’s challenger in the artificial intelligence accelerator market. With Helios, presented as a rack-scale AI computing system, the U.S. group is explicitly expanding its playing field: the issue is no longer simply offering a high-performance GPU, but delivering a complete infrastructure designed to run very large models and to be deployed in the data centers of cloud providers, AI laboratories and large companies.

According to TechCrunch, in its article entitled AMD takes on Nvidia with its Helios AI rack-scale system, the first Helios systems are expected to ship to customers later this year. The timing matters because the generative AI market has gradually become organized around massive deployments: thousands, or even tens of thousands, of accelerators must be connected, powered, cooled and orchestrated as a single computing system. In this configuration, the GPU remains indispensable, but it is no longer the only determining product.

The Helios name therefore points to a shift in how AMD presents itself to infrastructure buyers. For a long time, the company sold components: EPYC processors for servers, Instinct GPUs for accelerated computing, network cards and, since its acquisitions, technologies from Xilinx and Pensando. The rack-scale system aims to bring these building blocks together in a coherent offering. The goal is to reduce the gap between delivering an accelerator and the actual operation of a production AI cluster.

This change in the offering reflects an industrial reality. Buying GPUs is not enough to create a competitive AI platform. It also requires a network topology suited to very frequent communications between accelerators, host processors capable of feeding the computations, high-bandwidth memory, software tools, rack management mechanisms and thermal and electrical integration compatible with rising power levels. Value is therefore shifting toward the entire system.

AMD is, in effect, taking on the field on which Nvidia has built a major part of its lead. Jensen Huang’s company does not merely sell GPUs: for years, it has offered DGX systems, followed by integrated rack architectures intended for very large deployments. Nvidia combines its accelerators with interconnects, network cards, storage solutions and, above all, CUDA, its software environment, which has become a reference for a large part of the accelerated computing ecosystem.

Helios therefore does not mean that AMD is abandoning the component logic. On the contrary, the rack is precisely a way to enhance several product families from the group at the same time. But the announcement reflects a different ambition: to offer customers an alternative that can be evaluated against a complete Nvidia platform, rather than only against an individual GPU card in a benchmark.

Helios, infrastructure designed for massive AI workloads

The elements disclosed by AMD and reported by TechCrunch describe Helios as “rack-scale” AI computing infrastructure, meaning it is designed at the level of the entire rack. This expression is not merely marketing language. In a large computing environment, achieved performance depends on the whole: accelerator speed, memory capacity, the speed at which chips exchange data, network latency, power consumption, heat dissipation and the software that distributes tasks.

Generative AI models explain this growing complexity. Training them requires carrying out a very large number of parallel operations and continuously sharing parameters among graphics processors. Inference, which consists of using an already trained model, also creates specific constraints, particularly when a large number of users must be served with low latency. In both cases, the quality of the network between chips and between servers can become as important as raw computing power.

Helios specifically targets this level of integration. The system combines the computing power required for AI workloads, networking components and the software layer that makes it possible to use them together. AMD is therefore not simply telling its customers how to assemble a cluster from separate products: the group intends to provide an integrated reference architecture intended to accelerate the deployment of large-scale computing capacity.

The distinction is crucial for operators. Building infrastructure from standard servers and separate GPUs provides greater freedom in choosing suppliers, networks and software. But this approach can lengthen technical validation, complicate support and require considerable optimization work. By contrast, a rack-scale platform promises a more predictable configuration: components are selected to work together, and responsibility for integration is borne to a greater extent by the architecture provider.

This direction does not make technical issues less important. It simply places them at a higher level. For a buyer, evaluating Helios will not be limited to the number of GPUs installed in the rack. It will be necessary to examine the bandwidth available between nodes, how networks are organized, power requirements, cooling requirements, compatibility with existing tools and the ability to run the models actually used by the organization.

The promise of an integrated system also has an operational dimension. In a data center, a modern AI rack can represent a considerable concentration of electrical power and heat. Architectural choices therefore have direct consequences for the building, cooling systems and the pace at which new capacity can be brought online. The market is thus moving closer to the logic of supercomputers, where hardware integration and operation of the facility cannot be separated.

For AMD, Helios is an opportunity to present a more complete response to the needs of very large customers. Cloud providers and groups that train their own models are not necessarily looking for a simple component: they are seeking capacity that is available, repeatable and supported over time. A rack architecture is a way to address this expectation directly.

The fact that the first shipments are announced for later this year now sets the real test for the initiative. A technology presentation is not enough in this segment. Credibility will depend on the concrete availability of the systems, their integration into customer environments and measurable results on real AI workloads. Rack announcements have become frequent in the industry; the difficulty lies in moving from announcement to large-scale production.

The Nvidia precedent and the evolution of AMD’s strategy

The current battle is part of a longer history. AMD was long primarily associated with its x86 processors, which compete with Intel’s Xeon processors in PCs and servers. With EPYC processors, the group strengthened its presence in data centers. It then accelerated its ambitions in heterogeneous computing, where CPUs, GPUs, specialized accelerators and networking components are combined to handle the most demanding workloads.

In the world of accelerated computing, Nvidia has built a structural lead. This lead is due to the performance of its chips, but also to the time spent building an ecosystem around CUDA. Developers, researchers, software vendors and cloud providers have built numerous tools and workflows around this platform. The result is a network effect: a technology is all the more attractive when it is already supported by the libraries, models and skills available on the market.

AMD has responded by developing its Instinct range and its ROCm software stack. The Instinct family targets compute-intensive and AI workloads in data centers, while ROCm aims to provide the software layers needed to program and run accelerated workloads on AMD hardware. The challenge is considerable: it is not merely about making a model capable of starting on another architecture, but about ensuring that it runs efficiently, with deployment, monitoring and optimization tools usable in production.

The group nonetheless has several relevant assets for a platform strategy. EPYC processors give AMD a strong presence on the server CPU side. The acquisition of Xilinx, completed in 2022, brought capabilities in FPGAs and adaptive systems. That same year, AMD completed the acquisition of Pensando, a specialist in data processing units and networking technologies. These operations are not enough on their own to create an AI platform, but they strengthen the group’s ability to assemble building blocks beyond the GPU.

The MI300 range illustrated this integration strategy. AMD notably highlighted the MI300X, an AI-focused accelerator equipped with 192 GB of HBM3 memory. In the eight-GPU configurations presented by the group, total memory reached 1.5 TB. This memory capacity was an important argument for large language models, which require storing a very large number of parameters as well as temporary data during execution.

But the next step is different. A platform such as Helios seeks to go beyond the traditional comparison between two accelerators. The debate is no longer only about the memory of a GPU or a theoretical computing figure. It is about how a complete rack processes a model, the utilization rate of accelerators, the cost of managing infrastructure and the speed at which a customer can expand its fleet.

Nvidia has played a major role in establishing this new standard. Its DGX systems showed that it was possible to sell a complete infrastructure product, rather than only cards intended to be integrated by third parties. NVL systems, associated with more recent GPU generations, extended this trajectory toward very large systems. Nvidia presented Blackwell in 2024 as an architecture designed for the needs of large-scale AI, with increased attention to interconnection and deployment in data centers.

Helios must be read in this context. AMD is not simply seeking to respond to a technical benchmark; the company is responding to a commercial and operational benchmark. Faced with customers that can buy a highly integrated solution from Nvidia, AMD must demonstrate that a stack based on its own components can be ordered, installed, used and maintained with a comparable level of simplicity.

This dynamic also explains why AI can no longer be analyzed solely as a semiconductor market. It is becoming an infrastructure market. Potential winners are not only those that design the best transistors, but those that master system assembly, software, supply, support and relationships with major data center operators.

Competition playing out through networking, software and execution

The Helios announcement comes in a market that is much more contested than at the beginning of the current generative AI cycle. Nvidia remains the central player in AI accelerators, but AMD is seeking to capture part of the demand driven by cloud providers and companies that want to diversify their computing sources. Intel, for its part, remains present in data centers and is continuing its efforts in accelerators and processors for AI. Cloud providers are also developing their own chips, notably to reduce their dependence on general-purpose accelerators and to optimize certain internal workloads.

Chips developed by hyperscalers do not, however, address exactly the same need as Helios. In-house accelerators are often designed primarily for the operator’s own infrastructure or cloud services. AMD, meanwhile, is promoting an architecture intended to be purchased by different customers. This position may interest organizations that do not want to depend on a single cloud platform, or that want to install capacity in their own data centers and colocation environments.

The network is one of the most important battlefields. In a distributed AI system, accelerators must exchange data at an extremely high rate. If these exchanges become too slow, GPUs spend part of their time waiting rather than computing. The effective performance of the cluster then diverges sharply from the sum of the theoretical performance of the installed chips.

This constraint favors suppliers capable of offering a coherent approach to interconnection. Nvidia strengthened its networking offering with the acquisition of Mellanox, completed in 2020. AMD, for its part, has the assets from Pensando and its own experience in data center platforms. Helios is a manifestation of this reality: a rack-scale offering cannot be credible without a clear story about how nodes communicate.

Software is the other decisive factor. To win a lasting place, AMD must convince research teams, platform engineers and publishers that their workloads can be ported and optimized without excessive cost. ROCm plays a strategic role here. The availability of an open software environment is regularly emphasized by AMD, but openness does not automatically solve adoption difficulties. Companies need stable libraries, compatibility with their frameworks, diagnostic tools and trained staff.

The migration issue is central. An organization that has already industrialized its models on an Nvidia stack does not change suppliers simply because a competing accelerator offers an interesting feature. It must assess porting costs, regression risks, engineer availability, third-party software support and how performance varies from one model to another. Competition therefore rests as much on reducing this friction as on hardware characteristics.

AMD may nevertheless benefit from a favorable context. The concentration of demand around a dominant supplier naturally pushes some major buyers to examine alternatives. Diversification is not only a matter of commercial negotiation. It also concerns supply security, the ability to avoid bottlenecks and the possibility of selecting the most appropriate hardware depending on model types. Helios gives AMD a more legible product for engaging with these buyers.

It is nevertheless necessary to distinguish the strategic interest of an alternative from its ability to establish itself. In AI systems, the total power of a rack does not tell the whole story. Power consumption, density, cooling requirements, delivery times, support availability and integration with existing storage systems play a direct role in the decision. The market now values industrial execution as much as architectural innovation.

The battle between AMD and Nvidia is therefore not limited to which player has the fastest GPU. It concerns the ability to make a collection of thousands of components usable as a single platform. This is the ambition summarized by Helios: turning a portfolio of chips, processors and software into a complete infrastructure product.

What Helios could change for French and European companies

In France and Europe, the announcement deserves particular attention because AI infrastructure needs are growing within a framework marked by sovereignty, data control and energy constraints. European organizations do not constitute a homogeneous market: cloud providers, telecommunications operators, public administrations, research laboratories, banks, industrial companies and startups do not have the same workloads or purchasing criteria. But all face the question of access to suitable computing power.

An AMD rack-scale offering could theoretically expand the options available to operators building or expanding AI clusters. For customers, the existence of an integrated alternative is important, even when it does not immediately result in a change of supplier. It can improve the ability to compare, encourage multi-platform testing and support a less concentrated procurement strategy.

This issue of technological plurality is particularly sensitive in Europe. Debates about trusted cloud, hosting sensitive data and dependence on major U.S. platforms do not directly concern the GPU manufacturer, but hardware choices influence the operational autonomy of infrastructure. A European operator seeking to offer AI services must be able to secure its supplies, train its teams and ensure the sustainable operation of its clusters.

Helios does not eliminate dependencies: AMD is also a U.S. group and the semiconductor value chain remains global. However, the arrival of an additional system offering can help prevent purchasing decisions from being limited to a single architecture. For integrators, hosting providers and companies operating their own infrastructure, this opens the possibility of comparing platforms rather than choosing only among different card models.

Physical constraints will be decisive in the European context. Deploying high-density AI racks requires available electrical power, suitable cooling solutions and data centers capable of supporting a significant concentration of power. These issues are taking on growing importance in France, where data center projects are viewed through the lens of electricity consumption, access to land, water when it is used for cooling, and the recovery of waste heat.

For a French buyer, the relevance of Helios will therefore depend on very concrete data that the general announcement is not enough to settle: rack power, cooling method, availability through partners, support arrangements, compatibility with internal software and performance on the relevant models. A laboratory training scientific models, a bank running risk-analysis models and an AI service provider will not necessarily have the same criteria.

Software availability will be equally decisive for the French-speaking ecosystem. Engineering teams do not work in the abstract: they rely on frameworks, containers, libraries and deployment pipelines already in place. Adopting new infrastructure requires that common AI tools be supported and that performance be reproducible. AMD will therefore have to convince not only IT departments, but also developers and MLOps platform managers.

Public research and scientific computing players may also follow this development closely. Europe has a strong tradition of supercomputing and shared research infrastructure. Generative AI is gradually bringing certain industrial needs closer to the historical needs of high-performance computing: massive parallelism, fast interconnects, energy optimization and planning of scarce resources. A complete rack offering may be examined from this perspective, provided it meets requirements for code portability and long-term maintenance.

It would nevertheless be premature to interpret Helios as an automatic shift in the European market. Nvidia has a deep ecosystem, broad presence among cloud providers and strong recognition among developers. AMD’s success will depend on the ability of the first customers to put systems into production and obtain convincing results on representative workloads. Initial deployments, more than announcements, will serve as references for subsequent buyers.

The next phase will depend on AMD’s ability to turn the announcement into a deployment standard

With Helios, AMD implicitly recognizes that the next stage of AI will be played out less in the isolated demonstration of a chip than in the industrialization of complete systems. This development has consequences for the entire chain: semiconductor manufacturers, networking equipment suppliers, server assemblers, data center operators, cloud providers and software vendors will have to work in more closely coordinated ways.

The move to rack scale may also change the relationship between AMD and its partners. When a supplier mainly sells components, server manufacturers and integrators retain a large share of responsibility for the final architecture. When it offers a more integrated platform, it defines the reference product more extensively. This can simplify implementation for the end customer, but it also requires a greater ability to provide support, validation and availability for all elements.

The prospect of first shipments later this year will therefore be examined from several angles. Customers will want to know whether systems can be installed in existing infrastructure without disproportionate modifications. Developers will look at the maturity of the software environment. Operators will measure energy efficiency and cooling constraints. Finance departments will compare the total cost, which includes much more than the purchase price of accelerators.

The market could also become more segmented. Some companies will continue to favor the most established platforms to reduce risks. Others will seek to diversify their suppliers. Cloud providers may offer several architectures depending on customer needs. In this framework, AMD does not necessarily need to reproduce Nvidia’s model exactly to succeed: it must build an offering robust enough to become a credible choice in a substantial portion of deployments.

This credibility will rest on continuity. AI rack buyers do not make a decision for just a few months. They anticipate renewal cycles, model developments and growing power needs. They therefore want to understand how an architecture will be supported, how it will evolve and whether software investments made today will remain useful with the next generations of hardware.

For AMD, Helios is therefore a test of industrial strategy. The group has already demonstrated its ability to compete with major server players with EPYC and to offer data center AI accelerators with Instinct. The new step is to make these products the elements of a solution whose customers will not have to rebuild all the coherence themselves. In an industry where deployment times and capacity availability have become competitive advantages, this integration can carry as much weight as chip specifications.

The response from Nvidia, cloud providers and other manufacturers will continue to accelerate this dynamic. AI is pushing computing toward systems that are ever denser, more interconnected and more expensive to install. For French and European players, the challenge will be to benefit from this increased competition without underestimating the conditions for success: software skills, access to energy, supply chains, hosting and operational management.

Helios does not instantly reshape the market hierarchy. But the announcement, as reported by TechCrunch, confirms that the center of gravity is shifting. AMD is no longer presenting itself only as an accelerator manufacturer facing Nvidia: it wants to become a provider of foundations for large-scale AI. The ability to deliver the first racks, achieve their adoption and demonstrate their effectiveness in real-world environments will determine whether this ambition can translate, in the coming years, into a lasting alternative in data centers.