Key Takeaways
AMD Helios is shipping: AMD launched Helios on 23 July, its first rack-scale AI system, with 72 MI455X GPUs and 31 TB of HBM4 per rack.
MI450 deployment commitments are large: OpenAI has a 6 GW multigenerational agreement, Anthropic has committed to up to 2 GW, and Oracle plans 50,000 MI450 GPUs. Microsoft will deploy Helios at scale on Azure.
The AMD GPU cloud reaches neoclouds: Vultr placed a $1.2 billion Helios order. It brings a second-source GPU to a market that has run almost entirely on NVIDIA.
GPU vendor lock-in starts to loosen: A second supplier adds capacity and pricing pressure, but software and networking still decide how easily workloads move.
AMD Helios Ships Into the GPU Cloud Market
AMD launched Helios at its Advancing AI event on 23 July, and Fierce Network reported that the system is in production, with volume deployments ramping through the second half of 2026. Helios is AMD's first full rack-scale system. The rack targets the same buyers as NVIDIA's rack-scale systems.
The rack: Each Helios rack holds 72 MI455X GPUs in one scale-up domain, with 31 TB of HBM4 and 2.9 exaflops of peak FP4 compute.
The claims: AMD says Helios delivers up to 30% more tokens for the same spend than the NVIDIA Rubin NVL72 and 50% more memory.
MI450 Deployment Commitments Reach the Neoclouds
The demand side moved first. On 22 July, AMD and Anthropic announced a deal for up to 2 GW of MI450 GPUs in Helios racks, with the first gigawatt from the first half of 2027. OpenAI already had a 6 GW agreement.
On 1 October, HPE announced a $1.2 billion order from Vultr for Helios systems in US data centers, the first Helios order for HPE. That puts an AMD GPU cloud offer inside a neocloud, which is a step beyond the hyperscaler and lab deals.
Helios shipments started later in Q3, and are projected to ramp up in Q4 and continue through 2027. Oracle plans 50,000 MI450 GPUs starting in the second half of 2026.
A Second-Source GPU Changes the Supply Landscape
Supply has been the binding constraint all year. NVIDIA posted $89 billion of data center revenue in its July quarter, and Futurum notes demand remains supply-constrained, with procurement commitments up to $279 billion, mostly for memory. A second-source GPU adds a separate supply line at a time when every existing line is booked, which is a gap we covered in our analysis of the 2026 GPU shortage.
The idea of a second source predates the AI boom. When IBM picked an Intel processor for its first PC in 1981, it required Intel to line up a second manufacturer, because betting a flagship product on a single chipmaker looked too risky. Intel turned to a young AMD, and the two signed a technology exchange in 1982 that a federal appeals court later called in effect a reciprocal second-sourcing agreement.
That deal is how AMD first entered the x86 business. Four decades later, AMD is playing the same role again, this time for the chips that train and run AI models. Today’s largest AI buyers are reaching the same conclusion IBM did.
OpenAI has a 6 GW agreement with AMD, while NVIDIA has said it intends to invest up to $100 billion in OpenAI to support 10 GW of its hardware, according to PitchBook reporting, a figure Jensen Huang has since described as nonbinding. Anthropic’s new deal builds on its existing use of AMD chips. Spreading orders across suppliers protects a buyer if one supplier’s production slips and gives the buyer leverage in price negotiations.
GPU Vendor Lock-In Loosens as Supply Diversifies
Hardware is only half of GPU vendor lock-in. Software decides how easily a model moves between chips, and the Anthropic deal includes a multi-year effort in which Claude will help accelerate ROCm development.
Portability decides the outcome for neoclouds: A second supplier lowers prices only when workloads can move. That depends on GPU portability across frameworks, kernels, and serving stacks, which will determine how much real competition Helios brings to the GPU neocloud market.
Orchestration over ownership: Aethir has argued that orchestration is the product as GPU hours commoditize. Aethir’s decentralized GPU cloud aggregates capacity from third-party providers in more than 90 countries, and a more diverse hardware market gives that routing layer more supply to work with.
AMD Helios gives the GPU cloud market its first credible rack-scale alternative in production, backed by gigawatt-scale commitments and now a neocloud order. Extra supply is welcome in a market where every line is booked. The harder test is software, since GPU portability decides whether buyers can actually move.
A decentralized GPU cloud that routes work across many providers is built for a market with more than one supplier.
Frequently Asked Questions
What is AMD Helios?
AMD Helios is AMD's first rack-scale AI system. It launched on 23 July 2026, and each rack holds 72 MI455X GPUs with 31 TB of HBM4.
What is the AMD MI455X GPU?
The MI455X is the GPU in the Helios rack and part of the MI450 series. Seventy-two of them share one scale-up domain linked by UALink over Ethernet.
Which companies have committed to MI450 deployment?
OpenAI has a 6 GW agreement, and Anthropic has committed to up to 2 GW. Oracle plans 50,000 MI450 GPUs, Microsoft will deploy Helios on Azure, and Vultr placed a $1.2 billion order through HPE.
Does a second-source GPU reduce GPU vendor lock-in?
On the hardware side, yes. A second-source GPU adds supply and pricing pressure, loosening GPU vendor lock-in, but workloads still need to run well on the new chips, so software portability determines how far that goes.
