Nvidia's grip on the AI chip market weakens as Microsoft switches to AMD

Microsoft will deploy AMD's new Helios platform in its Azure data centres, according to multiple sources. The platform is scheduled for the second half of 2026 and targets the segment that Nvidia has dominated to date with its H100 and H200 GPUs. At the same time, indications are emerging that Anthropic is also testing AMD hardware, increasing the pressure on Nvidia's market position.

The shift is notable because Microsoft and Anthropic are among Nvidia's largest customers. If they move part of their computing workloads to AMD, it will affect not only Nvidia's revenue but also its negotiating position with other major clients.

What AMD's Helios platform entails

AMD CEO Lisa Su introduced the Helios platform as an integrated AI training solution consisting of multiple Instinct MI350 accelerators connected via AMD's proprietary interconnect technology. It is designed for large-scale training and inference tasks, the same type of workloads for which customers currently turn almost exclusively to Nvidia.

Microsoft has maintained a long-standing collaboration with AMD on AI solutions. The choice for Helios fits into a broader strategy in which the company aims to reduce its dependence on a single supplier. CEO Satya Nadella has previously spoken publicly about the need to diversify the supply chain for AI compute, although Microsoft has not yet commented in detail on its specific Helios plans.

In addition to Microsoft, AMD has secured agreements with Meta, OpenAI and Oracle as major compute customers. The fact that OpenAI is a customer of both Nvidia and AMD illustrates how players in the AI ecosystem deliberately maintain multiple suppliers.

Indications that Anthropic is testing AMD

A public GitHub profile linked to Anthropic contains references to AMD hardware. These are test environments, not production infrastructure, but the signal is clear enough to attract attention in the industry. Anthropic has not yet commented on the matter.

Anthropic currently relies primarily on two platforms: Alphabet's Tensor Processing Units (TPUs) and Nvidia's GPUs. On 23 October 2025, the company signed a multi-billion-dollar agreement with Google for access to up to one million TPUs, which had already reduced its dependence on Nvidia. If AMD hardware is also deployed on a structural basis, Anthropic would be distributing its compute across three suppliers.

Anush Elangovan, Vice President of AI Software at AMD, has publicly stated that the software layer around AMD's hardware has improved significantly in recent years. That was historically the main objection to AMD as an alternative to Nvidia: developers regarded the ROCm software stack as less mature than Nvidia's CUDA ecosystem.

Nvidia's position remains strong, but margins are under pressure

Nvidia retains a dominant position in the AI accelerator market. Large language models are still trained predominantly on Nvidia hardware, and the CUDA platform has built an ecosystem of libraries, frameworks and specialists that is difficult to match.

Nevertheless, the signals that major customers are seriously considering alternatives are relevant to Nvidia's pricing. As long as Microsoft, Meta and OpenAI had virtually no alternative, Nvidia was able to maintain high margins. A credible AMD alternative changes that negotiating dynamic, even if customers ultimately opt for a mixed infrastructure rather than switching entirely.

Julien Simon, Chief Evangelist Officer at Hugging Face, previously noted that the availability of more competitive hardware options also benefits the open-source AI community, as it puts downward pressure on compute access costs. Intel and Samsung are also attempting to establish a position in the AI chip segment, although their products have not yet reached a level comparable to Nvidia's or AMD's emerging offerings.

What this means for the European AI scene

For European AI companies, cloud providers and policymakers, the shift in the chip market is significant. The high cost of Nvidia compute is widely perceived across Europe as a barrier to scaling AI models. If AMD's Helios platform genuinely becomes competitive in the second half of 2026, it will increase the number of suppliers available to European players and put pricing pressure on the entire segment.

At the same time, software compatibility remains a practical concern: many European AI teams have built their pipelines on CUDA, and migrating to ROCm requires time and technical capacity. The next two years will reveal whether AMD's catch-up on the software side is sufficient to remove that barrier.