AI companies are burning through GPUs and electricity at a pace that would have sounded absurd just a couple years ago. Now, startup Zyphra says it has secured 15 megawatts of AMD Instinct MI355X GPU capacity for its Zyphra Cloud platform, which is aimed squarely at the booming AI infrastructure market.
That is a massive amount of hardware. We are long past the days of companies bragging about a few servers in a rack somewhere. AI firms are now measuring compute deployments in megawatts like they are utility companies.
The San Francisco-based company says Zyphra Cloud is built for AI startups, enterprises, and hyperscalers looking to run workloads on AMD hardware instead of relying entirely on NVIDIA. The platform launched with serverless inference for open-source AI models running on MI355X GPUs, but Zyphra is now expanding into bare-metal deployments too. In other words, customers can get dedicated AMD GPU clusters for training models, reinforcement learning, inference, and whatever other AI buzzword is hot this week.
To be fair, AMD has been gaining momentum in AI. NVIDIA still dominates the space, but companies are clearly looking for alternatives as GPU prices stay sky high and supply constraints continue to frustrate customers. AMD sees an opening, and firms like Zyphra are trying to ride that wave.
Zyphra says it has already worked with IBM on custom AMD clusters using Instinct GPUs and Pensando Pollara networking. The company also claims its research division trained its ZAYA1 AI models entirely on AMD systems, which is exactly the sort of thing AMD likes hearing right now.
The bigger story here may actually be the scale of all this. Fifteen megawatts of GPUs means serious power consumption. AI companies keep talking about bigger models, longer context windows, and agentic systems, but all of that comes with a growing appetite for electricity. Folks worried about the environmental side of AI are probably not going to feel better after hearing announcements like this.
Zyphra also wants people to see it as more than just another GPU rental provider. The company is pitching a full stack approach that combines software, infrastructure, networking, and inference optimization into one AMD-focused ecosystem. Whether customers buy into that pitch remains to be seen, but AMD definitely benefits every time another company publicly commits to its AI hardware.
The MI355X capacity is available now, and Zyphra says future AMD platforms like the MI450 series are already part of its roadmap.
