Advanced Micro Devices (AMD) has unveiled the AMD Instinct MI350P PCIe GPU cards, a new line of enterprise-grade AI accelerators engineered to slot into standard air-cooled server racks without requiring infrastructure upgrades. The launch targets a fast-growing segment of organizations that need serious AI computing power but aren't ready or able to rebuild their data centers around dedicated GPU platforms.

The MI350P cards are dual-slot, PCIe form factor accelerators designed to work within existing power, cooling, and rack infrastructure. For IT decision-makers weighing the costs of cloud AI versus on-premises deployment, the new cards offer a middle path: high-throughput AI inference on hardware organizations already own.

The cards are available in air-cooled systems supporting up to eight accelerator cards per server, making them suitable for a wide range of deployment scales from small business AI pilots to large enterprise inference pipelines.

AMD is positioning the MI350P as the highest-performing enterprise PCIe card currently on the market, citing an estimated 2,299 TFLOPS of AI compute with peak performance reaching 4,600 TFLOPS at MXFP4 precision. The cards are also equipped with 144GB of HBM3E memory running at up to 4TB/s bandwidth — figures that place them at the top end of the PCIe accelerator market.

Native support for lower-precision MXFP6 and MXFP4 formats for maximum throughput

Sparsity acceleration for mainstream 8-bit and 16-bit precision workloads

Support for FP8, MXFP8, INT8, and BF16 — covering the full range of enterprise AI model precision requirements

The breadth of supported precision formats is central to AMD's pitch: the MI350P can handle today's most demanding inference workloads without requiring liquid cooling or specialized power infrastructure.

Beyond the hardware, AMD is emphasizing the economics of its software stack. The company provides an open-source AMD enterprise AI reference stack to partners at no licensing cost, with the goal of lowering operating expenses compared to cloud-based AI services that charge per token.

The software ecosystem includes native support for PyTorch, a Kubernetes GPU Operator for full lifecycle management, and AMD Inference Microservices, all designed to enable organizations to migrate existing inference workloads with minimal code changes. AMD frames the stack as interoperable with a wide range of existing AI tools and pipelines.

Enterprises adopting AI at scale face a genuine infrastructure dilemma. Cloud AI services offer flexibility but can introduce data privacy concerns and unpredictable costs that scale with usage. Purpose-built GPU accelerator platforms offer raw performance but often demand expensive data center redesigns to accommodate new power and cooling requirements.

The MI350P is AMD's answer to this gap, positioning PCIe accelerator cards as a practical on-ramp for organizations that need to move beyond CPU-only inference but aren't ready to commit to a full GPU-native infrastructure buildout.

The cards are designed to support retrieval-augmented generation (RAG) pipelines and small-to-large model inference, two workload categories that have become central to enterprise AI deployments in 2025 and 2026.

The MI350P enters a market where NVIDIA has long dominated with products like the L40S and the recently released Blackwell-based PCIe cards. AMD's pitch centers on the openness of its software stack and the total cost of ownership advantages of avoiding proprietary licensing fees, a strategy the company has used in its broader campaign to gain ground in the AI accelerator market.

With the MI350P, AMD is signaling that enterprise AI doesn't have to mean starting from scratch. For the many organizations sitting on racks of standard air-cooled servers and weighing their next AI infrastructure move, that message may land at exactly the right moment.