TensorWave raises $350M Series B to build AMD-exclusive AI supercomputing infrastructure

TensorWave raises $350M Series B to build AMD-exclusive AI supercomputing infrastructure

TensorWave is betting that enterprises fed up with Nvidia pricing and supply constraints will route their AI workloads through AMD hardware instead. The company builds and operates supercomputing data centers exclusively using AMD GPUs, with facilities already running in Arizona, Florida, and Pittsburgh, and more going up across North America.

The AMD relationship

TensorWave's connection to AMD predates the GPU shortage. The company's prior business deployed FPGAs at scale, primarily Xilinx chips. When AMD acquired Xilinx, TensorWave was already working closely with the company, receiving early silicon and helping debug deployments. When AMD announced its GPU lineup in 2023 and Nvidia supply dried up, TensorWave positioned itself as one of the first to bring AMD chips to market at scale.

“We raised 350,000,000 Series B. Our focus is on building out AI infrastructure exclusively with AMD. We were able to show that you could switch from NVIDIA and switch to AMD and it works out of the box in the early days. CUDA is not the moat of NVIDIA — their ecosystem is the moat, and that's what AMD is now building out.”

The CUDA question

The persistent objection to AMD has been the CUDA ecosystem — not the chip itself, but the decade-plus of libraries, developer tooling, and institutional familiarity Nvidia has built around it. Tatarchuk's view is that CUDA was never the moat; the ecosystem surrounding it is. AMD's ROCm software stack had real gaps early on, and Tatarchuk credits Lisa Su with going into what he calls founder mode to close them, directly engaging the developer community and pushing her team to accelerate. AMD's CVP of AI, Anush Elangovan, argues that AI coding tools now let AMD move faster than ever to close those gaps.

Databricks demonstrated as early as 2023 that workloads could migrate from Nvidia to AMD without major rework on common open-source architectures. TensorWave's own team includes Greg Diamos, one of the original engineers who built CUDA at Nvidia, who has argued the infrastructure was always designed to extend beyond any single chip.

Where AMD has an edge

AMD's MI-series chips ship with more HBM memory than comparable Nvidia hardware. That matters for large model inference, where higher memory capacity means running bigger models on fewer GPUs. Luma AI, which does video and image generation, is a TensorWave customer — the kind of memory-intensive creative workload where AMD's HBM advantage is most tangible.

The round

TensorWave has raised $350M in a Series B. The capital is going toward expanding the North American data center footprint.

The company's near-term proposition is straightforward: customers who want an alternative to Nvidia-only supply chains, and who don't want to hand all their margin to Jensen Huang, have a credible place to land. How fast AMD closes the remaining software gaps will determine whether TensorWave is a bridge or a destination.