BuildNot yet confirmed elsewhere1 publisher2 min readPublished
Arista's Etherlink SU-144 reference designs stretch an Ethernet scale-up domain to 1,024 accelerators
Arista paired its 7060EX7 switches with three Ethernet rack designs, the largest reaching 1,024 accelerators in one scale-up domain across racks. Teams on non-Nvidia chips such as Microsoft's Maia, the market Arista expects to grow, get a second scale-up fabric to evaluate.
The Engineer · Build desk
What happened
- Arista is rolling out the 7060EX7 with fiber patch panels, liquid-cooling manifolds and leak detection, designed to run software from AMD, Arm, Broadcom, Meta, Microsoft and Qualcomm.
- A second design uses a cabled backplane and modular rack so technicians can replace or upgrade parts with little disruption to the rest of the system.
- Arista is active in the Open Compute Project's Ethernet for Scale-Up Networking group, which aims to make standard Ethernet reliable under accelerated AI loads.
Why it matters
- decision Buying the rack as the unit puts cooling, power and network validation into one purchase, so network and facilities teams have to sign off on the same order.
- cost The write-up includes no prices, so an infra team cannot yet use the 7060EX7 as leverage against a proprietary fabric, and the case that closed cabling costs more stays unquantified.
- precedent If the Ethernet for Scale-Up Networking work yields a spec other switch makers build to, the scale-up switch becomes a part a buyer can bid out between vendors.
The bandwidth figures fail a unit check. A dev.to write-up of the launch puts per-rack bandwidth at about 1.6 petabytes today and 6.5 petabytes with next-generation optics [14]. That is roughly a fourfold increase [17]. A petabyte with no time attached is an amount of data, so as printed the figure is not a rate. I would not plan floor space around the 50 percent footprint claim until a 7060EX7 datasheet states the unit.
Power is the next check. Spread the orthogonal design's 100 to 400 kilowatts across 144 accelerators and each gets about 0.7 to 2.8 kW [12]. That is a ceiling, because the switch blades sit in the same envelope. For the 144 density to transfer to a given data hall, the accelerator a team plans to buy has to fit under it at full population. The cross-rack design is the bigger bet. Its 1,024-accelerator scale-up domain extends past a single rack [6] and is about seven times the orthogonal chassis [11].
Leak detectors are an unusual item in a switch launch [2], and I think including them is correct. Each of the three designs, grouped as the Etherlink SU-144 family, folds compute, networking and liquid cooling into one unit [9]. The write-up says the rack is becoming the basic unit of deployment. It says that takes integration across power distribution, thermal management and network validation [15].
Arista frames standard Ethernet as an open alternative to proprietary interconnects [1]. Its stated bet is that Ethernet demand grows with the non-Nvidia accelerator market [13]. The examples given are Microsoft Maia, Meta's training accelerators and Google's TPUs [7]. For teams building on those chips, this is a second scale-up fabric with reference designs behind it. The cost case is thinner. According to the write-up, closed systems need specific, often more expensive cabling and switching components [16].
Compatibility rests on the standards work. The write-up credits Arista's adherence to those standards for keeping its products compatible with a wide range of hardware [8]. Arista's pitch also names switching silicon, alongside accelerators and rack designs, as something operators should be free to vary [10].
What to watch
- A 7060EX7 datasheet that gives port speeds, latency and the per-rack bandwidth figure as a rate with a time unit.
- A published Ethernet for Scale-Up Networking specification from the Open Compute Project, and a second switch vendor shipping against it.
- A named deployment of the cross-rack design at or near its 1,024-accelerator ceiling.
Clarity's read
What the record supports and how the coverage leans. The claims behind it follow.
Reality
- Evidence35
- Adoption
- Insufficient
- Hype gap+30
- Incentives70
- Confidence40
Claim ledger
Ranked by verification strength, evidence, and original report placement.
- [1]
Arista Networks is positioning Ethernet as the networking fabric for AI scaling as an open alternative to proprietary systems, and introduced new reference architectures and integration partnerships.
- [2]
Arista is rolling out its 7060EX7 Series switches with fiber patch panels, liquid-cooling manifolds and leak detection systems, designed to work with software from partners including AMD, Arm, Broadcom, Meta, Microsoft and Qualcomm.
- [3]
Arista is heavily involved in the Ethernet for Scale-Up Networking initiative formed by the Open Compute Project, whose goal is to ensure standard networking can handle the demands of accelerated AI infrastructure while maintaining reliability.
- [4]
The first reference design uses an orthogonal chassis with direct connectivity between accelerator and switch blades, is efficient for liquid cooling, supports up to 144 processing units in a single envelope, and is built for power consumption of 100 to 400 kilowatts.
- [5]
The second reference design uses a cabled backplane with a modular rack structure, emphasizing serviceability so technicians can replace or upgrade components with minimal disruption.
- [6]
The third, cross-rack design extends the scale-up domain beyond a single rack and supports up to 1,024 accelerators while maintaining low latency.
- [7]
The AI accelerator market is diversifying, with organizations developing chips such as Microsoft Maia, Meta's training accelerators and Google's Tensor Processing Units.
- [8]
By adhering to these standards, Arista keeps its products compatible with a wide range of hardware.
- [9]
Arista defined three reference designs in the Etherlink SU-144 family that integrate compute, networking and liquid cooling into a single functional unit.
- [10]
Arista aims to offer network operators greater flexibility across accelerators, switching silicon and physical rack designs without locking users into a single vendor ecosystem.
- [11]
The cross-rack design's 1,024-accelerator domain is about seven times the 144-accelerator orthogonal chassis.
- [12]
The orthogonal design's 100 to 400 kW envelope works out to about 0.7 to 2.8 kW per accelerator at 144 accelerators, before subtracting switch and cooling load.
- [13]
Arista believes that as the non-Nvidia accelerator market grows, demand for standard Ethernet will follow.
- [14]
Current configurations provide about 1.6 petabytes of bandwidth per rack; Arista anticipates next-generation optics will push this to 6.5 petabytes, which could lead to a 50 percent reduction in data center physical footprint.
- [15]
The entire rack is becoming the basic unit of deployment instead of individual switches, requiring integration between power distribution, thermal management and network validation.
- [16]
Closed systems require specific, often more expensive, cabling and switching components.
- [17]
The projected per-rack bandwidth increase from 1.6 to 6.5 is roughly fourfold.
Sources
1 independent publisher whose own reporting we read for this story.
- dev.toArista Networks Challenges Nvidia with AI Ethernet Fabric
1 article · October 8, 2026
Topics and entities
Follow any of these and your For You feed starts watching them — no settings page required.
Topics
- Custom AI AcceleratorsFollow
- AI scale-up networkingFollow
- Ethernet for AIFollow
- Data center liquid coolingFollow
Entities
- Arista NetworksFollow
- Etherlink SU-144Follow
- Arista 7060EX7 SeriesFollow
- Network Diagnostics InfrastructureFollow
- Open Compute ProjectFollow
- Ethernet for Scale-Up NetworkingFollow
- Ultra Ethernet ConsortiumFollow
- NvidiaFollow
- Microsoft MaiaFollow
- Google TPUFollow