AI Data Center Networking in 2026: How Ultra Ethernet and UALink Are Challenging InfiniBand, and Why the Wires Now Decide Who Wins
- Internet Pros Team
- August 18, 2026
- Networking & Security
For most of computing history the interesting question about a data center was what chips were inside it. In 2026 it is how those chips talk to each other. Training a frontier AI model means tens of thousands of accelerators working on one calculation, pausing several times a second to exchange results, and waiting for the slowest link before anyone continues. When that exchange is slow, extremely expensive silicon sits idle. The network stopped being plumbing and became the thing that decides whether a cluster delivers what it cost.
Why the Bottleneck Moved to the Wires
A single accelerator can no longer hold a large model, so the model is split across many. Every training step ends with a synchronization: each chip shares its partial results, everyone combines them, and only then does the next step begin. That pattern, a collective operation, differs from ordinary internet traffic in two ways: it is enormous and bursty rather than steady, and it finishes only when the last packet arrives. A network that is fast on average but occasionally slow is worse than a slower one that is consistently predictable.
This is why utilization matters more than headline bandwidth. On identical hardware, the gap between a well-tuned fabric and a poorly tuned one can be tens of percent, which on a cluster costing hundreds of millions of dollars funds an entire additional facility. Networking became a capital efficiency question.
In an AI cluster the slowest link sets the pace for every chip in the building. Average speed is a marketing number; tail latency is the number that costs money.
Two Networks, Two Different Problems
The confusing part of this topic is that a modern AI system contains two distinct networks solving different problems with different technologies.
Scale-Up Versus Scale-Out
- Scale-up connects the accelerators inside one rack so tightly that software can treat them as a single very large processor with shared memory. Distances are short, latency budgets are measured in hundreds of nanoseconds, and the traffic is memory access rather than messaging.
- Scale-out connects racks to each other and to storage across the building, spanning hundreds of thousands of ports. Here the concerns are routing, congestion, failure recovery, and cabling economics, at a scale where one design mistake multiplies by ten thousand.
For years the scale-out layer belonged largely to InfiniBand and the scale-up layer to proprietary links designed by the accelerator vendor. Both worked well. Both also meant the network came from the same company as the chips, which is a comfortable position for exactly one supplier and an uncomfortable one for every buyer.
What Ultra Ethernet Actually Changes
Ethernet has always been the cheap, universal option with one flaw for this workload: it was designed to tolerate loss and reorder packets, which is fine for web traffic and poisonous for a synchronized collective. Earlier fixes bolted remote memory access onto Ethernet and then demanded a lossless network underneath, achieved through pause signals and tuning that famously fell apart at scale.
The Ultra Ethernet effort inverts that: keep standard Ethernet wires, optics, and switches, and rebuild the transport layer above them for this traffic. The important pieces are practical rather than glamorous.
The Core Ideas
- Spray packets across every available path. Instead of pinning a conversation to one route and hoping it is not congested, traffic is spread over all of them and reassembled at the destination. This alone removes the single worst source of unpredictable slowdowns in large fabrics.
- Tolerate loss instead of forbidding it. Selective retransmission means a dropped packet costs one packet, not a stall across the whole network. The fabric no longer has to be perfectly lossless to perform well.
- Move congestion control to the endpoints. Senders react to measured conditions rather than relying on switches to issue blunt pause commands that ripple outward and freeze unrelated traffic.
- Keep the ecosystem. Standard cables, standard optics, standard switch silicon from many vendors, and the same operational tooling network teams already use.
Meanwhile a parallel effort, UALink, is prying open the scale-up layer with an open memory-semantic link between accelerators in a rack. Buyers want to mix accelerators from different vendors in one tightly coupled domain, and they want the specification to outlive any single supplier roadmap. Whether the alliance ships silicon fast enough to matter is the open question, and the incumbent is not standing still.
| Technology | Layer | Strength | Trade-off |
|---|---|---|---|
| InfiniBand | Scale-out | Mature, low latency, proven at very large scale | Single dominant supplier; separate skills and tooling |
| Ultra Ethernet | Scale-out | Multi-vendor, huge supply chain, familiar operations | Young software stack; results depend on implementation quality |
| Proprietary scale-up links | Scale-up | Highest bandwidth and lowest latency available today | Locks the rack to one vendor generation |
| UALink | Scale-up | Open specification, potential for mixed accelerators | Early; ecosystem and silicon availability still forming |
Why Buyers Pushed for Open Standards
None of this is really an engineering argument; it is a procurement one. When the network, the accelerator, and the systems software all come from one company, that company sets the price, the roadmap, and the lead time. Cloud providers spent three years discovering how uncomfortable that is when supply is constrained, and responded the way large buyers always do: by funding a standard.
History favors that bet. Ethernet has repeatedly absorbed specialized networking technologies that were superior in their niche, because volume manufacturing and a deep pool of trained engineers eventually outweigh a latency advantage. The likely outcome through 2028 is a split: proprietary links keeping the scale-up crown a while longer, and Ethernet steadily taking the scale-out layer where cost per port and vendor choice dominate.
What This Means If You Are Not a Hyperscaler
Very few businesses will ever buy a switch that speaks these protocols, but the effects arrive in three ways.
First, prices: competition at the interconnect layer lowers the cost of the AI compute you rent, and inference pricing has been falling partly because clusters are being used more efficiently. Second, features: the congestion management, telemetry, and multi-path techniques being standardized now will reach ordinary campus and branch switching within a few product cycles. Third, skills: network engineering is a specialization worth paying for again, after a decade in which the interesting work moved to software.
The Honest Caveats
A specification is not a product. Performance depends on the network cards, switch buffers, and collective libraries that implement it, and early deployments show wide variation between vendors claiming the same compliance. The incumbent also has a decade of software tuning behind it. Treat any single benchmark chart with suspicion.
The Practical Takeaway
The habit worth borrowing is measuring predictability, not peak speed: a database cluster or a multi-site file sync feels slow because of its worst behavior under load, not its best, yet most office networks are still specified by bandwidth alone. More broadly, infrastructure bottlenecks move, and rarely where the marketing is pointing. Attention went to chips for five years while the constraint quietly relocated to the connections between them, then to power and cooling, and it will move again. Organizations that periodically re-measure where their own systems actually wait spend less and get more.
If you would like a clear-eyed look at where your own network is the bottleneck, whether that is switching, wiring, wireless design, or the link to your cloud provider, Internet Pros designs and tunes business networks for the traffic they actually carry.
