TensorWave is expanding its AI cloud infrastructure by deploying the AMD Helios rackscale solution powered by AMD Instinct MI455X GPUs, positioning the company to support larger AI training, inference, and fine-tuning workloads. The announcement, made during AMD’s Advancing AI 2026 event, extends TensorWave’s strategy of building an AI cloud platform exclusively around AMD technologies while adding AMD’s latest rack-scale architecture for enterprise and frontier-scale AI deployments.
The AMD Helios rackscale solution integrates 72 AMD Instinct MI455X GPUs, 6th Generation AMD EPYC processors, AMD Pensando networking, and the ROCm software platform into a unified AI rack. AMD says the platform delivers up to 2.9 exaFLOPS of peak 4-bit performance and 1.4 exaFLOPS of peak 8-bit performance while providing 31 TB of HBM4 memory and 1.7 PB/s of aggregate memory bandwidth. Within the rack, AMD employs UALink over Ethernet (UALoE) for scale-up connectivity, while Ultra Ethernet Consortium-aligned Ethernet provides scale-out networking between racks. The platform also incorporates confidential computing, device attestation, encrypted multi-GPU communications, and support for frameworks including PyTorch, TensorFlow, JAX, ONNX Runtime, vLLM, and SGLang.
TensorWave said the new infrastructure will support customers developing frontier AI models and large-scale inference services. Among those customers is Featherless AI, whose CEO Eugene Cheah cited TensorWave’s expertise with AMD infrastructure as a factor in selecting the cloud provider. TensorWave, backed by investors including Magnetar, AMD Ventures, and Nexus Venture Partners, says it operates one of the world’s largest AI cloud environments built exclusively on AMD Instinct GPUs.
| AMD Helios Rackscale Configuration | |
| GPUs per Rack | 72 × AMD Instinct MI455X |
| Peak AI Performance | 2.9 exaFLOPS (4-bit) 1.4 exaFLOPS (8-bit) |
| HBM4 Memory | 31 TB per rack |
| Memory Bandwidth | 1.7 PB/s aggregate |
| Per-GPU Capability | 40 PFLOPs (4-bit), 20 PFLOPs (8-bit), 432 GB HBM4, 23.3 TB/s bandwidth |
| Compute Layout | 18 OCP ORW-aligned 4-GPU trays |
| Networking | AMD Pensando Vulcano 800 AI NIC, UALink over Ethernet within rack, UEC-aligned Ethernet between racks |
| Software | AMD ROCm with PyTorch, TensorFlow, JAX, ONNX Runtime, vLLM and SGLang |
| Security | Confidential Computing, device attestation, encrypted multi-GPU scaling |
At TensorWave, we believe the future of AI starts with better infrastructure. AMD Helios brings compute, memory, and networking together at rack scale, enabling our customers to build AI faster, more efficiently, and with better economics at scale,” said Jeff Tatarchuk, Chief Growth Officer and Co-Founder of TensorWave.
🌐 Analysis
AMD is building an ecosystem around its Helios rackscale architecture by aligning cloud providers, networking vendors, storage suppliers and software partners around a common AI infrastructure platform. TensorWave joins other early adopters positioning Helios as an alternative to proprietary AI systems for organizations seeking open Ethernet-based AI clusters.
For TensorWave, the announcement reinforces its strategy of differentiating as an all-AMD cloud provider rather than competing directly with general-purpose GPU clouds. The combination of MI455X GPUs, UALink-over-Ethernet, Pensando networking and ROCm software demonstrates AMD’s continued effort to offer an integrated AI infrastructure stack spanning compute, networking and software.
