---
title: "AI250"
url: "https://www.qualcomm.com/data-center/products/qualcomm-dragonfly-ai250"
description: "Qualcomm Dragonfly AI250 rack-scale AI inference platform delivering 768GB memory, 133TB/s bandwidth for low-latency, cost-efficient real-time AI."
---

# AI250

Qualcomm Dragonfly™

Active

[Contact Sales](https://www.qualcomm.com/support/contact/forms/data-center)

![chip image](https://s7d1.scene7.com/is/image/dmqualcommprod/Dragonfly_AI250_Raster_RGB_Color?fmt=webp-alpha&qlt=85)

On this page

[Overview](https://www.qualcomm.com/data-center/products/qualcomm-dragonfly-ai250/#overview)
[Benefits](https://www.qualcomm.com/data-center/products/qualcomm-dragonfly-ai250/#benefits)
[Features / Specs](https://www.qualcomm.com/data-center/products/qualcomm-dragonfly-ai250/#features)

Rack-scale AI inference with breakthrough memory bandwidth — engineered for the real-time agentic AI and low TCO.

Qualcomm Dragonfly™ AI250, our second-generation, rack-scale AI inference platform introduces the groundbreaking [Qualcomm® High Bandwidth Compute (HBC) Gen 1](https://www.qualcomm.com/data-center/expertise/ai-accelerators) — enabling 133 TB/s of effective memory bandwidth per card (18x more than AI200). Combining massive memory bandwidth and capacity enables fast, low-latency inference on SOTA models up to 10T parameters and context lengths to 1M tokens, all within an ORv3 compliant air- and direct liquid-cooled rack.

Built for hyperscale agentic AI, it enables disaggregated inferencing at superior token-per-watt and token-per-dollar.

[Product license agreement]()

## Benefits

- **Real-Time Token Generation**
With industry-leading 133 TB/s effective memory bandwidth per card¹ addressing memory-bound AI inference decode, HBC enables fast token generation.

- **Low Latency per Token**
Massive effective memory bandwidth keeps decode pipelines fed, reducing latency per token for more responsive real-time agentic AI experiences.

- **Large Context Window**
With 768 GB per card, agentic workloads with long reasoning chains are designed to minimize context eviction.

- **Total Cost Advantage at Scale**
AI250 is engineered to deliver superior total cost of ownership (TCO) versus competition at iso tokens-per-second-per-user — making real-time inference economically viable at scale.

- **Superior Tokens per Watt**
4x-8x better performance per watt compared to contemporary GPU-based architectures on memory bandwidth per watt per card.²

- **Deploy AI Faster**
Qualcomm AI Inference Suite streamlines deployment across bare metal, cloud VMs, and inference-as-a-service, helping teams move from model to production faster.

¹Compared to competing published product specifications normalized at card- and rack-level

²Qualcomm estimates compared to contemporary GPU-based architectures on memory bandwidth per watt per card

## Features

- Industry-leading effective memory bandwidth¹ to deliver compelling cost per token on real-time inferencing on SOTA models up to 10T and context lengths up to 1M.
- Designed for memory-bound and real-time inference at hyperscale economics: LLM decode, reasoning, agentic AI, and multimodal generation
- HBC architecture for industry-leading performance per watt²
- 43 TB memory capacity and 7.4 PB/s effective bandwidth with HBC per rack
- Over 6 TB of HBC memory per server, capable of supporting over 10T parameter model, reducing networking dependency.
- PCIe Gen6 scale-up; Ethernet-with-RoCE scale-out
- Qualcomm AI Inference Suite for bare-metal, VM, or IaaS deployment
- Rack-scale solution with liquid cooling, storage, network switches, and NICs
- Air and direct liquid cooling
- OCP ORv3-compliant rack with cableless backplane

### Specifications

#### Rack

| Specification | Value |
|---|---|
| Form Factor | Single Wide, Open Rack v3 (ORv3) compliant |
| Number of Cards | 56 |

#### Memory

| Specification | Value |
|---|---|
| Capacity | 43 TB |
| Effective Bandwidth | 7.455 PB/s [1](https://www.qualcomm.com/data-center/products/qualcomm-dragonfly-ai250/#footnote1) |

#### Scale-Up

| Specification | Value |
|---|---|
| Interface Type | PCIe 6.0 |

#### Scale-Out

| Specification | Value |
|---|---|
| Interface Type | Ethernet w/RoCE |

#### Thermal Management

| Specification | Value |
|---|---|
| Cooling | Direct Liquid Cooling (DLC), Air Cooling |
| Thermal Design Power | 140 kW |

1. With HBC Gen 1 (18x compared to AI200)

QUALCOMM DRAGONFLY AI INFRASTRUCTURE MANAGEMENT SUITE

Orchestrate and
scale efficiency

The suite provides provisioning, monitoring, orchestration and fault handling across rack‑scale deployments. Together, hardware, connectivity and software form the foundation of a cohesive data center platform approach — one designed to scale with customers as AI workloads evolve. 

[Learn more](https://www.qualcomm.com/data-center/solutions/ai-infrastructure-management-suite)