The most expensive component of an AI data center is not the most famous one. The GPUs get the headlines, but the network between them decides what the GPUs are actually worth. Gartner's Market Guide for AI Network Fabrics, published May 20, 2026, makes that argument in infrastructure terms: network performance directly affects GPU utilization, job completion time, and total training cost.
The bill is a network bill
An AI training cluster is a coordination problem. Thousands of GPUs splitting a model update across themselves, synchronized constantly. Every idle GPU waiting on the network is an invoice, and the invoice is paid in training time.
The fabric is what Gartner calls the dedicated hardware and software built for exactly that problem: bandwidth, low latency, near-zero packet loss, stable congestion control, fine-grained telemetry, dynamic traffic scheduling, and power efficiency. These are not switch features. They are the difference between a cluster that trains and a room full of expensive accelerators.
Every idle GPU hour is a network invoice. The category exists to make the invoice smaller.
Three scenarios, one fabric
The guide organizes the market into three deployment scenarios, and the names matter.
Scale-up is the interconnect inside a single AI system. Scale-out is the big one: large-scale GPU and xPU cluster interconnect across a data center, where most training money lives. Scale-across is connectivity between data centers, the scenario that turns a cluster into an estate.
The competition has moved from the switch to the fabric. Gartner's own framing says future data center network competition will shift from single-switch performance to end-to-end fabric capabilities optimized for AI workloads. A fast switch in a slow fabric is a component, not a product.
What the May 2026 Market Guide for AI Network Fabrics actually defines
The guide published May 20, 2026, with Gartner Distinguished VP Analyst Andrew Lerner identified among its authors. It is Gartner's formalization of a market that previously lived inside data center switching coverage, and the separation is the finding: AI networking now has its own buyer, its own scenarios, and its own economics.
A Market Guide names representative vendors and does not rank them. The confirmed representative vendor in the public record is Ruijie Networks, for its AI Fabric data center network solution, a name that matters because it is a Chinese vendor entering Gartner's observation scope for this market alongside the Western switching incumbents. The fabric market is globalizing at the same moment it is forming.
The spending split Gartner's own numbers expose
Lerner's commentary attached to this research is harsher than the guide. Spending on AI network fabrics will surpass general-purpose data center networks in 2026 and more than double through 2029.
The split underneath is the story. Roughly 200 organizations operate high-end AI data centers, against about 100,000 organizations running traditional data centers, and the 200 outspend the 100,000 by a factor of three. Switch vendors are reallocating their best engineering talent to AI build-outs, which means the general-purpose market gets the leftovers: price hikes, long lead times, and declining support quality, by Gartner's own forecast.
Two markets, one switch supply. Lerner's advice for the ordinary data center is uncomfortable reading: extend equipment life, improve port utilization, consider certified refurbished equipment, and order early if new purchases are unavoidable.
The new faces on the field
The fabric market's vendor field is forming from three directions: the switching incumbents rearchitecting for AI traffic, the GPU-adjacent specialists building fabric management into AI stacks, and newcomers like Ruijie entering through the high-end AI data center door.
The guide's inclusion of Ruijie is the quiet signal. A vendor previously known for mainstream data center switching now holds a representative spot in the AI fabric guide, which is exactly the competition shift Gartner describes: whoever controls the end-to-end fabric controls the AI data center's economics, and the field has not settled who that is.
What the guide cannot promise
A Market Guide is a map, not a ranking. There is no Leader to shortcut to, and the market is too young for one. The scenarios are defined, the vendor field is forming, and the buyer carries the evaluation.
The other limit is the supply reality the guide documents. The fabric is a scarce, engineer-constrained market, which means lead times and prices are moving targets. A fabric strategy written today reads differently in two quarters.
Four questions for the fabric buyer
Which scenario are you actually buying for? Scale-out within a data center and scale-across between them are different products with different failure modes. Name yours before the vendors name it for you.
What does the telemetry show? Fine-grained, real-time visibility into congestion and packet loss is where the fabric's value lives. Ask for a live view, not a spec sheet.
Who owns the end-to-end path? The switch is one hop. The fabric is the whole journey. Ask which vendor is accountable for the entire fabric, not one layer of it.
Where does the supply come from? In a market with price hikes and long lead times on record, the vendor's component supply chain is part of your deployment date. Ask directly.
Analyst Source
Gartner Market Guide
Category definition, vendor inclusion, and market scenarios in this article draw on Gartner's Market Guide for AI Network Fabrics, published May 20, 2026, with Gartner Distinguished VP Analyst Andrew Lerner identified among its authors. The guide defines AI network fabrics as dedicated hardware and software for the connectivity requirements of AI workloads across scale-up, scale-out, and scale-across scenarios, and names Ruijie Networks among its representative vendors. The guide does not rank vendors or name Leaders. Related market commentary on AI fabric spending, supply constraints, and buyer advice comes from Gartner's published commentary by Andrew Lerner.
Source research
Gartner does not endorse any vendor, product or service depicted in its research publications, and does not advise technology users to select only those vendors with the highest ratings or other designation. Gartner research publications consist of the opinions of Gartner's research organization and should not be construed as statements of fact. Gartner disclaims all warranties, expressed or implied, with respect to this research, including any warranties of merchantability or fitness for a particular purpose.
The same shift shows up next door, in Cloud-Native Application Platforms. Gartner's 2026 criteria now treat AI agent governance and cost control as fundamental to a Leader placement, not containers alone, which is why five incumbents all had to add new capability just to keep their seats.
The related category on this site is Data Center Network Solutions. Forrester scored eight vendors in 2013 and named zero Leaders. It has not scored the category since 2018, right before AI training turned data center networking into one of the most consequential purchases in enterprise infrastructure.