Zyphra has launched Zyphra Cloud, a full-stack open AI platform powered by a 15MW AMD Instinct MI355X GPU cluster, marking one of the clearest attempts yet to build large-scale AI infrastructure outside NVIDIA’s ecosystem.
The San Francisco-based company announced the platform on May 4, 2026, in partnership with AMD and cloud infrastructure provider TensorWave. On May 11, Zyphra followed up by disclosing 15MW of MI355X capacity now available through Zyphra Cloud, expanding the platform to bare-metal deployments capable of handling pretraining, large-scale reinforcement learning, post-training, and agentic inference.
What Is Zyphra Cloud?
Zyphra Cloud brings together model serving, agent tooling, and scalable GPU compute under one roof. It is designed for developers, enterprises, and AI hyperscalers who want to run demanding workloads without assembling separate services from multiple providers.
The platform launched with Zyphra Inference, a serverless inference service that supports leading open-weight models like GLM 5.1, DeepSeek V3.2, and Kimi K2.6. According to Zyphra, the service relies on custom kernels, long-context algorithms, and advanced parallelism to achieve high throughput and low latency for production use cases, including autonomous coding agents, deep research pipelines, and extended workflow automation.
Krithik Puthalath, Founder and CEO of Zyphra, said the platform reflects years of developing and stress-testing AI systems on AMD hardware. He described Zyphra Cloud as going beyond a typical AMD-focused cloud, noting that the company has proven its capabilities across the entire AI lifecycle and is now packaging that expertise as production infrastructure.
Why The 15MW MI355X Cluster Matters?
A 15MW GPU cluster represents a significant power footprint for AI compute. While household comparisons vary by region and usage, the figure signals that Zyphra is targeting infrastructure well beyond a standard cloud deployment, large enough to support demanding AI workloads across the full training and inference pipeline.
The physical infrastructure comes from TensorWave, an AI-focused cloud provider backed by AMD Ventures and Magnetar. TensorWave was one of the earliest cloud operators to offer MI355X GPUs and had previously announced plans to assemble one of the largest AMD accelerator clusters in the world, spanning MI300X, MI325X, and MI350X hardware.
Zyphra Cloud offers two deployment options: on-demand bare-metal GPU clusters for flexible workloads and custom hyperscale AMD configurations for organizations running training or inference at scale.
Why Zyphra Is More Than A Neocloud?
What separates Zyphra from other AMD-oriented cloud providers is the depth of its research. In the week leading up to the 15MW announcement, Zyphra Research published a series of models and tools demonstrating full-lifecycle AI capability on AMD silicon.
ZAYA1-8B is a Mixture-of-Experts model activating fewer than 1 billion parameters, yet trained end-to-end on AMD Instinct MI300X through pretraining, midtraining, and fine-tuning. Zyphra reports that it performs competitively with models many times its size on math, coding, and reasoning benchmarks, including Mistral-Small-4-119B on select evaluations.
- ZAYA1-74B-Preview scales up to 4 billion active and 74 billion total parameters, serving as a pre-RL reasoning checkpoint that demonstrates large-scale MoE pretraining can run entirely on AMD.
- ZAYA1-VL-8B extends the architecture into vision-language territory, handling visual understanding, grounding, and OCR tasks.
- TSP (Tensor and Sequence Parallelism) is a new inference strategy built for scalable long-context and agentic workloads on AMD GPUs.
- All ZAYA1 models ship under the Apache 2.0 license, consistent with Zyphra’s stated mission as an open superintelligence research company.
How AMD And IBM Fit Into Zyphra’s AI Stack?
The Zyphra and AMD relationship predates Zyphra Cloud. In November 2025, the two companies teamed up with IBM to engineer a large-scale training cluster running AMD Instinct MI300X GPUs with Pensando Pollara networking on IBM Cloud. That system spanned 128 nodes, each housing 8 MI300X GPUs, and delivered over 750 PFLOPs of real-world training throughput. Zyphra used this cluster to train the original ZAYA1-base model.
The effort was widely viewed as a proof of concept showing that AMD’s hardware, networking, and ROCm software stack could handle large-scale MoE pretraining without any NVIDIA components.
Negin Oliver, Corporate Vice President of Business Development for AI at AMD, noted that the collaboration illustrates how purpose-built AI software paired with AMD’s accelerator design can deliver competitive inference performance for the most resource-intensive open-weight models available today.
Roadmap: MI450 And Beyond
Zyphra Cloud’s development plan extends past inference. Upcoming features include distributed RL and fine-tuning services, sandboxed environments for agent development running on AMD EPYC CPUs, and dedicated bare-metal GPU access for customers who need isolated compute.
The company also signaled plans to support next-generation AMD platforms, including the MI450 series, as part of a longer-term commitment to scaling its software-driven infrastructure across future AMD hardware.
What This Means For AI Infrastructure
Zyphra Cloud’s launch with 15MW of MI355X capacity arrives as the AI infrastructure market gradually moves toward a more multi-vendor landscape. AMD has been gaining ground through a series of large partnerships, and Zyphra’s demonstrated ability to train competitive MoE models on AMD and then serve production inference on MI355X reinforces the argument that credible alternatives to a single-vendor GPU supply chain are taking shape.
For developers and enterprises weighing their infrastructure options, Zyphra Cloud presents a combination that remains relatively rare: high-performance AMD compute, a research team with published results on AMD silicon, fully open-source model releases, and a platform that spans from training through production deployment.