TensorWave is a deep-tech company that delivers scalable, memory-optimised AI and HPC cloud services powered by AMD Instinct GPUs, which include the world’s first deployments of MI300X accelerators. Established in 2023, TensorWave will enable AI innovation through accessible High Performance Computing (HPC) infrastructure designed to meet the needs of demanding AI workloads. TensorWave is currently developing its HQ base in Las Vegas, Nevada.
TensorWave Inc. was founded in November 2023 by Piotr Tomasik to provide scalable AI/HPC Cloud with AMD GPUs.
Raised $100 million in Series A funding led by Lux Capital to build AI-specific Cloud Infrastructure (announced May 2024).
In May 2025 announced its Las Vegas Headquarters expansion with Nevada economic incentives and MI300X GPU deployment.
Appointed CEO Darrick Horton and other executive team members to lead the development of TensorWave’s AI Compute Platform.
TensorWave offers AMD Instinct MI300X, MI325X, and MI355X GPUs each with up to 288GB HBM3E memory per GPU, and are used to train and infer large language models with industry leading memory bandwidth.
Dedicated MI355X prices start at $2.95/GPU hour, MI325X at $2.25/GPU hour, and MI300X at $1.71/GPU hour. Custom pricing is available for enterprise clusters and storage. Three-year reservations are also possible.
TensorWave has exclusively utilised AMD Instinct GPUs to avoid NVIDIA supply chain constraints and price premiums. The superior memory capacity offered by AMD GPUs (up to 288GB HBM3E), are particularly well-suited to memory intensive LLM applications. Additionally, TensorWave’s use of direct liquid cooling enables 51% energy cost savings.
Yes, TensorWave is both SOC 2 Type II certified and HIPAA compliant. TensorWave prioritises security and complies with enterprise grade infrastructure located within North American Data Centres.
Yes, TensorWave has available both bare metal GPU nodes and fully managed Kubernetes clusters so that you can run your containerized AI applications with either of them providing you with full root access to dedicated hardware.
TensorWave will provide you with expert support through their Book A Call system or they will assign a team from their enterprise group to assist you with how to optimize your GPU clusters for your specific AI application workload.
There was no mention of a free trial for the typical sense of the word (free) although you can deploy your GPU access in under 10 seconds using their dashboard with a pay per use model which starts at $1.71/GPU hr for the MI300X.
TensorWave is an AMD GPU exclusive company. As such, they will need to be able to optimize your software for the ROCm stack if you plan to use any of their GPUs. If you are planning to run a memory intensive workload, this is likely going to be a good option for you but if you have a lot of CUDA optimized applications, you should check the compatibility of those applications with the ROCm stack before purchasing. Also, enterprise pricing will need to be customized for your needs.
TensorWave is delivering a competitive AMD based AI solution platform for organizations that are looking for a way to achieve higher memory capacity and lower energy consumption relative to other solutions such as NVIDIA, making it a viable alternative to NVIDIA for many memory intensive workloads. Additionally, the ability to rapidly scale bare metal environments is attractive to many organizations that do not want to worry about getting caught up in GPU shortages; however, the development of the AMD stack is lagging behind the development of the CUDA stack.
TensorWave is best suited for organizations that are running memory intensive AI training workloads and are interested in running these workloads on high capacity AMD GPUs without the price and availability concerns associated with NVIDIA GPUs.
TensorWave provides a service that is specifically focused on cloud-based systems using the AMD Instinct GPU (MI300X/MI325X/MI355X) with as much as 288 GB of HBM3E memory and liquid-cooling, which can reduce your energy expenses by 51%. For customers who want to run their own instance (bare-metal) pricing begins at $1.71/GPU-hour with SOC 2 Type II and HIPAA compliance. The company has a significant enterprise focus due to over $166 million in funding and has experienced very rapid growth.
Good - comprehensive specs from official website and AMD partnership announcements. Pricing transparent for dedicated instances, enterprise details require sales contact. Limited public info on customer metrics and API capabilities.
TensorWave’s entire infrastructure is powered by an exclusive partnership with AMD Instinct GPUs. This partnership is backed by AMD Ventures, Magnetar ($166M+) in funding. The company also collaborates with Modular (the MAX inference engine) and SuperMicro for hardware.
TensorWave operates one of the world’s largest all-AMD GPU clouds, featuring an Arizona cluster with 8,192 GPUs and >1 GW of computing power. They have North American datacenter locations with UEC-ready Ethernet networking.
Felafax (YC S24) optimized a 405 billion parameter model to train on a single 8-GPU MI300X node. Felafax used JAX on bare metal and demonstrated stable operation throughout multiple iterations of training.
SOC 2 Type II certified and HIPAA compliant. Liquid-cooled systems provide up to 51% reduction in energy costs versus air cooled systems.
Pricing for the three types of systems are: MI355X ($2.95/GPU hr with 288GB of HBM3E), MI325X ($2.25/GPU hr with 256GB), and MI300X ($1.71/GPU hr with 192GB). Three year reservations are available for large enterprise organizations.
| Architecture Type | Primary Optimization | Framework Support | Scalability Model | Training vs Inference | Use Case Fit |
|---|---|---|---|---|---|
| AMD Instinct MI325X (TensorWave) | Chiplet design with massive HBM3e | ROCm, PyTorch, JAX, TensorFlow | Cluster scaling via Ethernet (8,192 GPUs) | Both (large model training/inference) | LLM fine-tuning, generative AI, cost-sensitive scale |
| NVIDIA Blackwell (B200) | General-purpose parallel compute | CUDA, PyTorch, TensorFlow, JAX | NVLink/InfiniBand multi-rack | Both (training dominant) | Mixed workloads, diverse AI models |
| Google TPU (Ironwood) | Systolic array matrix operations | TensorFlow, JAX (XLA) | Synchronized pod architecture | Both (inference-optimized) | Large-scale training, cost-sensitive inference |
| GPU Generation | Primary Precisions Supported | Quantization Support | TOPS Improvement | Sparsity Features | TensorWave Workloads |
|---|---|---|---|---|---|
| MI300X (CDNA 3) | FP8, FP16, BF16 | INT8 quantization | 5x vs MI250 | Structured sparsity | 405B parameter fine-tuning (8-GPU node) |
| MI325X (CDNA 3.5) | FP8 (2.6 PFLOPS), BF16, FP16 | Advanced INT8/FP8 | Aggressive vs NVIDIA H200 | Enhanced sparsity | 8,192 GPU training cluster, inference |
| MI355X (CDNA 4 - Upcoming) | FP4, FP6, FP8, BF16 | Microscaling formats | Next-gen throughput | Advanced structured/unstructured | Planned integration for extreme scale |
| Workload Type | Optimal Hardware | Memory Requirement | Bandwidth Priority | Scaling Method | Key Metrics |
|---|---|---|---|---|---|
| LLM Training (Billions-Trillions Parameters) | TensorWave MI325X 8,192 cluster | 256 GB HBM3e minimum | Critical (6 TB/s per GPU) | Ethernet cluster scaling | FLOPS utilization, model convergence |
| LLM Inference (Real-time) | MI355X / MI325X bare metal | 192-256 GB (model dependent) | Critical (memory-bound) | Physical GPU partitioning | Tokens/sec, latency p50/p95, 70% cost reduction |
| LLM Fine-tuning (405B Parameters) | MI300X 8-GPU nodes | 192 GB per GPU (1.5 TB total) | High bandwidth essential | Bare metal JAX/ROCm | Training stability, token throughput |
| Generative AI / Multi-Model Serving | MI325X with partitioning | Scalable to 8 concurrent models | High (caching optimizations) | Managed endpoints autoscaling | 2x throughput, low latency |
