PostHeaderSection

Symbol

Global NavBar

PostContentSection

Global Footer

Blog Detail Page

Beirit Harvey

TensorWave co-founder and president Piotr Tomasik sat down with DCD's Kat Sullivan to talk through the decisions behind the choice for AMD.

How AI Workloads Are Shaping the Next Data Center Buildout

AI changes fast, and every shift in how people use it ripples straight into how we design the data centers that run it. Take a recent example: some TensorWave customers are now asking for more CPUs in their next AI deployment. This is a shift that's already reshaping how we think about data center design. For a cloud whose earliest customers came to run inference on AMD GPUs, that's a real signal.

At Yotta in Las Vegas, TensorWave Co-founder and President Piotr Tomasik sat down with DCD's Kat Sullivan to talk through the decisions behind that shift, from why we chose AMD to how customer commitments shape the capacity we build.


Why GPU Memory Drew Us to AMD

TensorWave's relationship with AMD grew out of an earlier venture where Tomasik's co-founders deployed Xilinx FPGAs in the cloud. When it came time to build TensorWave, AMD's GPU memory roadmap made the case for an AMD-exclusive cloud. For inference teams, more memory per GPU means room for model weights and the KV cache that grows with longer prompts and higher concurrency.

It was also important to the three co-founders to give developers an alternative to NVIDIA's closed-source walled garden. Tomasik traced that decision from the AMD Instinct MI300 through the MI325X and MI355X to AMD Helios, which TensorWave is preparing to deploy across North America. Today the relationship extends to how we build: our engineers work closely with AMD on network architecture and data center design to keep our facilities optimized for performance and resource use.


Training-Ready Clusters Gave Inference Teams Options

Inference drove early demand, but we designed our clusters to handle training from the start. The same high-speed network that connects GPUs for training can support inference spread across multiple nodes, and teams can still serve models that fit on a single node. That gave us room to take on training workloads as more customers began asking to train on AMD.

AI agents are starting to change the hardware mix again. An agent might call a model, run code or use a tool, then pass the result back to the model. CPUs can handle some of that work between model calls, which is exactly why customers are now asking us to dial up CPU capacity relative to GPU count in their next deployments, and why we're revisiting the CPU:GPU ratio we design around.AMD’s November announcement of its 6th Gen EPYC server processors gives a concrete example of what that shift can look like. Its Helion Liquid design puts CPU capacity for running agents in racks that can sit alongside Helios racks serving the models those agents call. As customers run more agents or give them more work to do between model calls, they can size CPU capacity separately from GPU capacity. That changes what teams need to plan for at the facility level, from rack space to power and cooling.


Customer Demand and Power Shape What Gets Built

When Sullivan asked whether the industry was building the right kind of capacity, Tomasik pointed to the unused fiber left over from the dot-com buildout. At TensorWave, we sign a customer contract before financing a deployment. That commitment gives us a specific workload to design around, rather than building speculative capacity and hoping demand arrives. 

This is yet another example of how TensorWave is dedicated to responsible expansion and one we think the industry should treat as a standard rather than an exception: growth that's tied to real demand, financed against real commitments, and built with an eye on the power and memory constraints that will decide who can actually deliver.

How much capacity the industry can bring online is a separate question. "I think power is the ultimate constraint in the industry," Tomasik said. He put memory supply next on the list.


From Large Campuses to Smaller Sites

Tomasik used the gradual rollout of 5G to explain how he sees AI infrastructure developing. For now, he favors building capacity at large campuses and getting better at operating it there. 

Longer term, he expects compute to spread across smaller sites with a wider mix of hardware. Running thousands of facilities would be a different operating challenge, and he expects that model to make more sense as data center GPUs mature and demand for geographically distributed compute grows.

Hear Piotr Tomasik explain why power is the biggest constraint on AI infrastructure and how customer demand is shaping what gets built next.

Watch the full DCD Studio interview.


TensorWave Cloud: Built for AI Velocity

TensorWave is an AMD-exclusive cloud provider delivering memory-optimized AI infrastructure for training, fine-tuning and inference. Our platform pairs AMD's latest GPU architectures with liquid cooling and ultra-high-speed networking to run demanding AI workloads with the performance, scalability and cost-efficiency modern AI initiatives require.

Ready to get started? Connect with sales. 

How AI Workloads Are Shaping the Next Data Center Buildout

Product

Solutions

Resources

Company

Product

Resources

Solutions

Company