Data Center Network Solutions

No current Forrester Wave covers this category. Forrester last scored data center networking in 2013 and evaluated the adjacent software-defined networking hardware market in 2018. A new evaluation is planned. What follows is what the historical research established, what changed while the category went unscored, and what a returning evaluation would have to account for.

Forrester evaluated eight data center networking vendors against eighty five criteria in 2013 and named no Leaders at all.

Not a narrow field of one. Zero. Strong Performers and Contenders, and nobody who met the bar for the top tier. Forrester's own commentary at the time acknowledged how dramatic that sounded and stood by it, telling buyers that anyone investing in the infrastructure their digital business would run on was going to have to make trade-offs.

Thirteen years later that verdict has aged better than most analyst research, and the category has not been scored since 2018.

The evaluation that found nobody

The Q1 2013 Wave, authored by Andre Kindness, covered Alcatel-Lucent, Arista Networks, Avaya, Brocade Communications Systems, Cisco Systems, Extreme Networks, HP, and Juniper Networks.

Its methodology is worth describing because it explains the result. Rather than comparing features, Forrester took an outside-in approach, constructing eighty five questions around a composite company assembled from client inquiries, network assessments, and survey data. In effect it wrote an RFP on behalf of a realistic buyer and scored vendors against it.

That composite buyer was under budget pressure, trying to move its operations team from technology silos toward a service organisation, and attempting to convert consolidated data centre infrastructure into a private cloud. What it needed was a system that reconfigured elements automatically and monitored the results against business policy, which Forrester had defined in 2011 as virtual network infrastructure.

The five requirements under that heading were: leverage both virtual and physical infrastructure, act as a vertically integrated module from Layer 2 through Layer 7, create a fabric of horizontally interwoven components, automate and orchestrate to deliver the right services per user, and allow management by lines of business.

Read that list against what the vendors were actually shipping in 2013 and the absence of Leaders becomes unsurprising. The hardware was capable. The automation, orchestration, and policy layers were early, and the idea that a line of business would manage network services was aspirational to the point of fantasy.

Forrester's framing of the market at the time called it the awkward teenage years. The assessment was that the ambition had been articulated and nobody had built it yet.

What data centre networking has to do

The physical job is connecting servers, storage, and services inside a facility, and connecting that facility to others and to the outside world.

The architecture that dominates is leaf-spine. Every access switch connects to every spine switch, so any server is the same number of hops from any other. That matters because traffic patterns inside a modern data centre are overwhelmingly server to server rather than server to user, and a design optimised for traffic leaving the building performs badly when the traffic stays inside it.

On top of the physical topology sit overlays, network virtualisation, segmentation policy, and the management and orchestration layer that decides what connects to what. This is the part Forrester was testing in 2013 and the part the vendors were weakest at.

Then the coverage narrowed

The Q1 2018 evaluation was titled Hardware Platforms For Software-Defined Networking, scoring nine vendors against forty criteria: Alcatel-Lucent Enterprise, Arista Networks, Extreme Networks, Cisco Systems, Cumulus Networks, Dell, Huawei, Juniper Networks, and Pluribus Networks.

Two things changed. The criteria count fell from eighty five to forty, and the scope narrowed from data centre networking generally to hardware platforms for a specific architectural approach.

The vendor list also records consolidation already underway. Extreme Networks appears having acquired Brocade's data centre networking business, and Avaya and HP have dropped out under those names.

After that, Forrester's scored coverage of this market stops.

Where the attention went instead

The analyst who ran both evaluations remains active in networking coverage, and what he covers now is instructive.

Recent Forrester networking research under his name addresses secure access service edge, with a Q3 2025 Wave that he described as marking a dramatic shift from the 2023 Zero Trust edge evaluation, alongside networking platforms as a category, the consolidation following HPE's acquisition of Juniper Networks, and vendor moves at the enterprise edge.

That is a coherent set of priorities and it is all outside the data centre. The enterprise networking conversation moved to the branch, the remote worker, and the security perimeter, because that is where the buying problem was for most organisations after 2020.

Meanwhile the data centre network became one of the most consequential purchases in enterprise infrastructure, for a reason nobody was forecasting in 2018.

What happened while nobody was scoring

Training and serving large models is a distributed computing problem, and distributed computing at that scale is a networking problem.

The numbers make the point better than the argument. Industry analysis puts networking at around ten percent of AI infrastructure cost while accounting for the substantial majority of deployment difficulty. GPU clusters routinely run at well under full efficiency, with the losses attributed to inter-GPU communication, availability, and stragglers rather than to the accelerators themselves.

An idle GPU is the most expensive idle asset in enterprise computing. A network that keeps thousands of them fed is worth far more than its share of the bill suggests, which inverts the traditional position of networking as the cheap plumbing under the expensive boxes.

The market response has been rapid. Data centre switch revenue forecasts now run to tens of billions cumulatively over five years, driven largely by AI infrastructure. AI backend networking was a small fraction of the data centre switch market in 2023 and is projected to approach half of it by 2027.

The fabric argument

The specific technical contest is between InfiniBand and Ethernet as the fabric for AI back-end networks, and it has turned decisively within about two years.

InfiniBand held roughly eighty percent of AI back-end deployments in 2023. It was purpose-built for high-performance computing, with low latency and lossless behaviour that suited tightly coupled training workloads.

By mid-2025 Ethernet was taking the majority of new builds. The reasons are unglamorous and familiar from every previous standards contest: cost, a broad multi-vendor ecosystem, existing operational skills, and the absence of single-supplier dependency. Loss-tolerant multipath Ethernet fabrics closed enough of the performance gap that the economics decided it.

Standards work followed the demand. The Ultra Ethernet Consortium published its first specification in 2025, and further work on Ethernet for scale-up networking has been proceeding through the Open Compute Project. Both are attempts to give Ethernet the properties that made InfiniBand attractive without the ecosystem constraints.

InfiniBand has not disappeared and continues to advance through successive generations. Its trajectory reversed rather than ended.

For a buyer, the practical consequence is that fabric choice is no longer a technical decision made in isolation. It carries supplier concentration implications, operational skill implications, and a bet on which standards ecosystem matures fastest.

What a returning evaluation would have to measure

If Forrester scores this category again, the criteria will need to look substantially different from either previous edition, for three reasons.

The workload changed the requirement. A network designed for virtualised enterprise applications optimises for flexibility, segmentation, and mixed traffic. A network for AI training optimises for predictable low latency across synchronised collective operations, where a single slow path stalls an entire job. Those are different engineering problems, and a vendor excellent at one is not automatically competent at the other.

Power and cooling entered the conversation. Rack densities for accelerated computing have risen to levels that constrain what can physically be deployed, and networking equipment now competes for the same power budget. Optical interconnect approaches including co-packaged optics are being pursued specifically because the power cost of conventional pluggable optics has become material at scale.

And the supply position shifted. The company supplying the accelerators also supplies both InfiniBand and Ethernet fabrics tuned for its own hardware, which is a degree of vertical integration the 2013 and 2018 vendor lists did not have to account for. Any evaluation that treats fabric selection as independent of compute selection will misdescribe how these decisions are actually made.

Why no Leaders may still be right

Return to the 2013 result. Forrester assessed eight capable vendors, found none of them met the standard it had defined on behalf of a realistic buyer, and said so.

That kind of conclusion is uncomfortable for everyone involved. Vendors cannot use it. Buyers do not get a shortlist. The analyst firm publishes a document whose headline is that nobody wins.

It was also, on the evidence, correct. The vision Forrester tested against, automated policy-driven infrastructure manageable by the business rather than by network engineers, took another decade to become even partially real, and the parts of it that did arrive came through cloud abstraction rather than through better switches.

There is a version of the same judgement available now. AI networking is being sold with confidence, on standards that are still maturing, into architectures whose requirements shift with each model generation, by vendors whose positions changed completely inside two years.

A buyer approaching this category should probably assume they are making trade-offs rather than selecting a winner, which is exactly what Forrester told them in 2013 and has not had occasion to revise.

Forrester does not endorse any vendor named here, and tier placement should not be read as a recommendation to buy.