Other90 minute exam; the recommended DCAIE and DCAIAOT courses add several days of training·$300 exam or Cisco Learning Credits; instructor-led five-day training from partners runs around $4,400

Cisco Certified Specialist: Data Center AI Infrastructure (DCAI 300-640)

3.8

The most useful AI certification Cisco has, because it tests something specific and hard: building Ethernet fabrics that can carry GPU training traffic. It is heavily tied to Cisco gear, but the networking fundamentals underneath it are genuinely valuable.

What We Liked

Covers RoCEv2, PFC, ECN and congestion control, the real skills behind AI networking
Counts toward CCNP Data Center, so it slots into an existing career path
Published blueprint with clear weights (20/30/30/20)
Includes power, cooling and efficiency, which most AI infrastructure courses skip

What Could Be Better

Large chunks are Cisco product configuration (UCS policies, APIC, Nexus Dashboard, Intersight)
Instructor-led prep is expensive, with partner courses around $4,400
Does not complete CCNP on its own; you still need the 350-601 core exam
Hard to practise without lab access to Cisco data centre hardware

Detailed review

DCAI 300-640, Implementing Cisco Data Center AI Infrastructure, first opened for testing at Cisco Live Amsterdam on 9 February 2026. It is a 90 minute exam priced at $300 or Cisco Learning Credits. Passing it earns the Cisco Certified Specialist in Data Center AI Infrastructure and also satisfies the concentration requirement for CCNP Data Center, alongside the 350-601 DCCOR core exam. Cisco recommends two courses for preparation: AI Solutions on Cisco Infrastructure Essentials (DCAIE) and Operate and Troubleshoot AI Solutions on Cisco Infrastructure (DCAIAOT). The official blueprint has four domains: AI fundamentals and applications (20%), AI infrastructure components and architecture (30%), AI infrastructure deployment and data management (30%), and AI infrastructure operations and troubleshooting (20%).

The fundamentals section covers workload types such as RAG, training and inference, plus cloud, hybrid, on-premises and edge deployment models. The deployment section is where the substance is: congestion control with PFC, ECN and ETS, RDMA over Converged Ethernet, QoS and load distribution, then UCS domain, power and storage policies, and fabric deployment with Nexus Dashboard, APIC, Hyperfabric and Intersight. I rate this higher than Cisco's AITECH exam for a simple reason. It tests something hard and specific. Making Ethernet behave losslessly enough for distributed GPU training is a real skill that relatively few network engineers have, and the blueprint covers it properly.

I was also pleased to see power usage effectiveness and cooling in the architecture domain. Power is the actual constraint on most AI build-outs now, and almost nobody teaches it. The downside is how much of the exam is Cisco product configuration. UCS policy objects, APIC and Intersight are only useful if you run Cisco, and the exam leans on them. Partner-led five-day training is listed around $4,400, which puts the real cost well beyond the $300 exam fee unless your employer covers it. My take: if you are a data centre network engineer and your company is buying Cisco AI PODs or Hyperfabric, this is one of the most job-relevant AI credentials you can get.

If your AI clusters are InfiniBand or fully NVIDIA networking, look at NVIDIA's networking and infrastructure exams instead.

[ final ]

The verdict.

A strong pick for data centre network engineers who expect to build AI fabrics on Cisco kit. The lossless Ethernet material is worth learning even if you never sit the exam. Skip it if your organisation runs InfiniBand or NVIDIA Spectrum-X end to end.