Why AMD Helios Is Unlikely to Overtake NVIDIA – Even If MI455X Is Competitive

Executive thesis

AMD Helios is probably the most important attempt yet to create a credible open alternative to NVIDIA's rack-scale AI architecture. The platform is technically serious. It combines 72 AMD Instinct MI455X accelerators, EPYC Venice CPUs, Pensando networking, UALink-over-Ethernet and the ROCm software ecosystem in a single rack-scale architecture.

This should allow AMD to gain meaningful share of the AI accelerator market. It does not, however, automatically mean that AMD will overtake NVIDIA.

Foxorox thesis: The competitive battle is no longer GPU versus GPU. It is AI factory versus AI factory. A modern hyperscale customer is buying compute, memory, scale-up networking, scale-out networking, software, libraries, orchestration, support and a deployment roadmap. NVIDIA currently controls substantially more of this stack. Therefore even an AMD accelerator that is competitive, or superior in selected specifications, does not by itself eliminate NVIDIA's structural advantage.

1. AMD Helios is a real rack-scale competitor

AMD Helios should not be dismissed as simply another GPU server. AMD designed the platform as a complete rack-scale architecture for frontier AI training and inference.

The official AMD Helios configuration integrates 72 AMD Instinct MI455X GPUs into a single scale-up domain. The platform provides up to 260 TB/s of aggregate scale-up bandwidth and up to 43 TB/s of scale-out bandwidth. So the problem maybe is the price of Helios and recend Blackwell chips launched by Nvidia. Both companies have almost same price for computing unit but of course AMD looks a little but cheaper then Nvidia and it must be cheaper beacouse only marketing get biral information that it could be better then NVIDIA. So in conlusion we live only in NVIDIA world at all. It could be also Cerebras that would come to compete with those two and of course chineese GPU developers but still race is only beetwen this two GPU producers.

Each MI455X accelerator provides up to 432 GB of HBM4 memory and up to 19.6 TB/s of peak memory bandwidth. At rack scale AMD specifies up to approximately 2.9 exaFLOPS of OCP MXFP4 performance and approximately 1.4 exaFLOPS of OCP MXFP8 performance.

The architecture is aligned with Meta's Open Rack Wide concept and Open Compute Project standards. AMD also uses liquid cooling and open networking technologies instead of attempting to reproduce NVIDIA's proprietary NVLink architecture.

Source: AMD Helios and AMD Helios Rackscale Solution brochure .

2. Engineering case study – 125 MW AI Data Center

The physical implications of adopting Helios become particularly interesting when the architecture is examined at data-center scale.

The following Foxorox engineering case study considers a 125 MW IT facility.

Rack count

At first sight this appears to produce a substantial real-estate advantage. Approximately 893 Helios racks can theoretically provide the same IT power that requires 1,250 conventional 100 kW racks.

But rack count is only one dimension of data-center density.

The double-width problem

This produces an important engineering conclusion.

Fewer racks do not automatically mean a smaller data center. In this conceptual comparison the Helios rack count falls by approximately 29%, yet the physical floor area occupied by the racks themselves increases because the Open Rack Wide architecture is materially wider.

Consequently, replacing a conventional rack architecture with Helios does not justify automatically reducing the external dimensions of a previously designed AI data center.

The additional space may actually be valuable for service aisles, CDU infrastructure, maintenance access, future expansion and hydraulic distribution.

3. Liquid cooling is an advantage – but not an NVIDIA killer

Helios provides a very high degree of direct liquid cooling. The architecture supplies liquid cooling to the compute infrastructure and is designed as a rack-scale liquid-cooled system.

For a conceptual 125 MW installation, a design previously based on an 85/15 split between direct-to-chip cooling and residual air cooling could potentially move toward approximately 90/10, subject to confirmation from the final equipment supplier.

The reduction in air-side cooling infrastructure can be material. But NVIDIA is moving in exactly the same direction.

Vera Rubin NVL72 is also designed around direct liquid cooling, rack-scale power management and increasingly sophisticated energy-buffering architecture. NVIDIA is therefore not standing still while AMD catches up.

4. The real problem for AMD: NVIDIA is not selling a GPU anymore

The largest misconception in comparing AMD and NVIDIA is to look only at theoretical accelerator performance.

NVIDIA now sells an integrated AI factory architecture.

Vera Rubin NVL72 combines:

• 72 Rubin GPUs,
• 36 Vera CPUs,
• NVLink 6 scale-up networking,
• ConnectX-9 SuperNICs,
• BlueField-4 DPUs,
• Spectrum-X Ethernet and Quantum InfiniBand scale-out networking,
• CUDA, CUDA-X and NVIDIA AI Enterprise software,
• rack management and orchestration,
• and a large OEM and infrastructure partner ecosystem.

This makes the relevant comparison much larger than MI455X versus Rubin GPU.

AMD has to compete simultaneously against NVIDIA's GPU, NVLink, networking, CUDA software, libraries, cluster orchestration, OEM ecosystem and deployment experience.

Source: NVIDIA Vera Rubin NVL72 .

5. CUDA remains NVIDIA's strongest moat

Hardware performance changes every generation. Software ecosystems change far more slowly.

CUDA has been developed for nearly two decades and has become embedded in the workflows of AI researchers, universities, hyperscalers, engineering organizations and software companies.

NVIDIA's CUDA-X ecosystem now includes hundreds of GPU-accelerated libraries spanning AI, data science, mathematics, scientific computing, simulation, communications and other specialized workloads.

NVIDIA has previously disclosed an ecosystem containing more than six million developers and thousands of GPU-accelerated applications.

ROCm has improved substantially. AMD now supports PyTorch, TensorFlow, JAX, ONNX Runtime, vLLM, SGLang and other important frameworks, and the gap is significantly smaller than several years ago.

But compatibility is not identical to ecosystem dominance.

Large AI operators care about:

• debugging tools,
• optimized kernels,
• distributed training libraries,
• inference runtimes,
• profilers,
• cluster management,
• documentation,
• developer experience,
• technical support,
• and the availability of engineers who already know the stack.

Migration therefore has a cost even when the alternative hardware offers an attractive price/performance ratio.

Sources: NVIDIA CUDA-X and AMD ROCm / Helios .

6. The revenue gap illustrates the installed-base problem

The strongest evidence of NVIDIA's advantage is not a benchmark. It is revenue.

The periods and segment definitions are not perfectly comparable, but the order-of-magnitude difference remains strategically important.

NVIDIA's Data Center revenue is currently more than ten times AMD's entire Data Center segment revenue on this simple quarterly comparison.

That scale creates several reinforcing advantages:

• more production volume,
• larger engineering budgets,
• greater software investment,
• more customer deployments,
• more operational data,
• stronger supplier relationships,
• and faster feedback between customers and product development.

This is a classic flywheel. The largest installed base attracts developers, developers improve the ecosystem, the ecosystem attracts customers, and customers justify further hardware investment.

7. Networking may be more important than the GPU benchmark

At small scale, individual accelerator performance matters greatly. At tens of thousands of GPUs, network architecture becomes equally important.

A trillion-parameter training workload must continuously exchange large quantities of data between accelerators. The performance of the complete cluster therefore depends on communication latency, bandwidth, congestion control, collective operations and software orchestration.

NVIDIA owns much of this path through NVLink, NVSwitch, ConnectX, Spectrum-X, Quantum InfiniBand and BlueField.

AMD is pursuing a different strategy: open standards. Helios uses UALink over Ethernet for scale-up and standards-based Ethernet for scale-out, supported by Pensando networking.

This is strategically attractive because hyperscalers often dislike vendor lock-in.

But openness also creates an integration challenge. A vertically integrated system can often be optimized more quickly than a multivendor architecture.

AMD's open architecture may ultimately be one of its strongest advantages, but in the short term NVIDIA's vertical integration reduces deployment risk for customers building multi-billion-dollar AI clusters.

8. NVIDIA controls the upgrade path

Another important advantage is product cadence.

NVIDIA customers that deployed Hopper, Grace Hopper, Blackwell and Blackwell Ultra can move toward Vera Rubin within an ecosystem that preserves many elements of the software, networking and operational model.

Vera Rubin NVL72 uses the third generation of NVIDIA's MGX NVL72 rack architecture. NVIDIA states that more than 80 MGX ecosystem partners support the platform.

This matters because hyperscalers do not rebuild their entire infrastructure strategy every year.

They prefer a roadmap.

The winning platform must therefore demonstrate not only today's accelerator, but the next several generations of:

• compute,
• memory,
• networking,
• power delivery,
• cooling,
• software,
• and rack architecture.

9. The power problem is becoming a platform problem

At hundreds of megawatts, AI infrastructure design increasingly begins with electricity rather than processors.

A 125 MW IT facility cannot simply increase accelerator density indefinitely.

The limits become:

• utility connection capacity,
• transformer capacity,
• busbar current,
• UPS and energy-storage architecture,
• liquid distribution,
• heat rejection,
• and transient power behavior.

NVIDIA is already developing an 800 VDC architecture for future AI factories and megawatt-class rack concepts.

Vera Rubin also introduces local energy buffering and power smoothing designed to reduce rapid rack-level power fluctuations.

This is important because AI workloads can produce extremely rapid changes in electrical demand. At gigawatt scale these transients become a grid and data-center electrical-design problem.

The competitive advantage therefore shifts again: from GPU efficiency to complete electrical-system integration.

10. Helios can gain market share without overtaking NVIDIA

None of this means AMD will fail. In fact, the opposite may be true.

AMD is building the strongest AI accelerator position in its history.

The company has publicly announced major commitments from some of the world's most important AI organizations. OpenAI expects AMD Helios deployments through multiple partners, while AMD and Anthropic announced plans involving up to 2 GW of AMD Instinct infrastructure.

AMD also states that Oracle Cloud Infrastructure plans to deploy 50,000 MI455X GPUs beginning in 2026.

These are not experimental installations. They indicate that AMD is becoming a legitimate second platform for hyperscale AI.

The most realistic outcome may therefore not be "AMD defeats NVIDIA." It may be the creation of a two-platform accelerator market in which AMD captures a meaningful share while NVIDIA remains the dominant ecosystem.

11. Why hyperscalers actually want AMD to succeed

There is another reason AMD can grow rapidly. NVIDIA's customers do not necessarily want NVIDIA to maintain near-total control over AI infrastructure.

Hyperscalers have several reasons to encourage an alternative:

• lower accelerator prices,
• greater negotiating leverage,
• diversified supply chains,
• reduced dependence on proprietary interconnects,
• greater control over networking architecture,
• and reduced vendor lock-in.

AMD's open strategy directly addresses these concerns.

Therefore AMD does not need to replicate NVIDIA's ecosystem perfectly to become extremely successful.

It only needs to be sufficiently competitive that large customers can economically operate a second platform.

12. Helios versus Vera Rubin – strategic comparison

13. The 125 MW data-center conclusion

From an engineering perspective, the analysis of a 125 MW installation produces an important result.

The accelerator architecture does not fundamentally change the electrical capacity required by the site.

Whether the 125 MW is delivered by approximately 1,250 conventional 100 kW racks or roughly 893 conceptual 140 kW Helios racks, the facility still has to provide approximately 125 MW of IT power and reject approximately the corresponding amount of heat.

Therefore the following major systems do not disappear:

• high-voltage grid connection,
• HV/MV transformers,
• medium-voltage distribution,
• large pumping systems,
• heat exchangers,
• dry coolers or cooling towers,
• emergency power systems,
• and substantial hydraulic infrastructure.

Higher compute density changes the internal topology of the data halls much more than it changes the fundamental physics of a 125 MW site.

14. Why AMD is unlikely to overtake NVIDIA in the current cycle

15. Bull case for AMD

The strongest AMD scenario is not based on beating NVIDIA on every benchmark.

It is based on customers deciding that diversification is strategically necessary.

If ROCm continues improving, UALink becomes a widely adopted industry standard, and large deployments by OpenAI, Anthropic, Oracle and other operators prove reliable, the perceived risk of choosing AMD can fall sharply.

Under that scenario AMD can potentially capture a materially larger portion of future accelerator spending without becoming the number-one supplier.

16. Bear case for NVIDIA

NVIDIA's greatest strength can also become a vulnerability.

Its highly integrated architecture generates strong customer lock-in, but very large customers increasingly have an economic incentive to avoid dependence on a single supplier.

An open ecosystem built around AMD, UALink and Ethernet could gradually reduce NVIDIA's control over the infrastructure stack.

Custom accelerators from Google, Amazon and other hyperscalers increase this pressure further.

The most important long-term threat to NVIDIA may therefore not be a single competing GPU. It may be the gradual commoditization of the entire AI accelerator stack.

17. Foxorox conclusion

AMD Helios changes the competitive landscape.

MI455X provides enough memory, bandwidth and rack-scale compute to make AMD a serious candidate for frontier AI deployments. ROCm has improved substantially, AMD has credible networking technology, and major AI companies are committing to large Instinct deployments.

But this is not yet enough to conclude that AMD will overtake NVIDIA.

Foxorox final thesis: NVIDIA's competitive advantage is no longer located only inside the GPU. It exists across the complete AI factory: GPU + CPU + HBM + NVLink + Ethernet/InfiniBand + DPU + CUDA + libraries + orchestration + OEM ecosystem + installed base + deployment experience. AMD Helios attacks almost every layer of this architecture, which makes AMD a much more dangerous competitor than before. But closing a hardware-performance gap can happen in one generation. Closing an ecosystem gap takes considerably longer.

For this reason, the most probable scenario is not that AMD replaces NVIDIA as the dominant AI infrastructure platform during the current generation.

A more realistic scenario is that AMD becomes a powerful number-two supplier, captures a meaningful portion of hyperscale deployments, forces NVIDIA to compete more aggressively on price and openness, and gradually weakens NVIDIA's near-monopoly economics.

That outcome could still be extremely valuable for AMD shareholders. AMD does not have to overtake NVIDIA to become one of the largest beneficiaries of global AI infrastructure investment. What about Cerebras? We think that this company can suprise with some extraordinary gains as it start to use less power and less machinery( optics, memory, etc) to compute on the same level as Nvidia and AMD GPU chips. We receon that If hyperscalers will adopt new langages and LLM models to Cerebras wafers ot couldd be killer for Nvidia pricing for new products as this is new player on the market with extraordianry technology.

Written by Pawel Demczuk, MSc
Foxorox AI Analyzer