Broadcom is using the 2026 Open Compute Project summit to present an end-to-end networking stack for AI clusters: scale-up connections inside tightly coupled systems, scale-out fabrics across racks and scale-across links between facilities. The featured components include Tomahawk 6 and Tomahawk Ultra switches, Jericho4 routers, 800-gigabit network interfaces and co-packaged optics.[1]
The emphasis reflects a change in where expensive AI capacity can be lost. An accelerator waiting for parameters, activations or collective operations is powered but not producing useful model work. As clusters grow, network congestion, failures, link power and communication patterns can determine how much of the theoretical compute customers actually receive.
The networking layers Broadcom is targeting
| Measure | Value | Context |
|---|---|---|
| Scale up | Connections within a rack or tightly coupled accelerator domain | Prioritises very low latency and high bandwidth for closely coordinated compute.Source 1 |
| Scale out | Fabric across racks inside a cluster | Tomahawk 6 supports 102.4 Tbps of switching capacity in one chip.Source 2 |
| Endpoint connection | Accelerator or host to the fabric | Thor Ultra is an 800G Ethernet network interface designed for AI workloads.Source 3 |
| Scale across | Connections between data centers | Jericho4 targets lossless transport over regional distances and clusters beyond one facility.Source 4 |
| Optical integration | Co-packaged optics beside the switch silicon | The design aims to reduce electrical reach and power challenges at high bandwidth.Source 5 |
Bandwidth numbers need a unit conversion—and a warning
A 102.4-terabit-per-second switch capacity is 12.8 terabytes per second when divided by eight. An 800-gigabit-per-second interface is 100 gigabytes per second by the same arithmetic. These are theoretical line-rate conversions before protocol overhead, topology, oversubscription, traffic imbalance or application behaviour. They make the scale tangible without claiming a training job will move useful tensors at those rates.
The gap between line rate and useful work is where congestion control and routing matter. Large training jobs send synchronized bursts and collective communications. A few overloaded paths can leave other links underused while accelerators wait. Telemetry, adaptive routing and rapid failure detection aim to keep traffic balanced and shorten the time a job spends recovering from a degraded link.
One facility is becoming a power constraint
Broadcom positions Jericho4 for distributed clusters spanning more than one data center. The reason is not networking ambition alone. A single building may not have enough power, cooling or physical space for the desired accelerator count. Connecting facilities creates access to more megawatts, but it also introduces longer distances, more failure domains, encryption requirements and stricter timing constraints.[4]
Why more network capacity does not automatically mean faster AI
| Constraint | How performance is lost | What to inspect |
|---|---|---|
| Topology | Traffic funnels through oversubscribed links | Path diversity and expected collective patterns.Source 2 |
| Congestion | Bursts queue and accelerators idle | Telemetry, load balancing and tail latency.Source 2 |
| Optics and power | High-speed links consume power and generate heat | Energy per bit, link reach and cooling.Source 5 |
| Failure recovery | A link or switch fault stalls a large job | Reroute time, checkpointing and job-level recovery.Source 1 |
| Software | Communication libraries do not exploit the fabric | Measured scaling efficiency on the real model and framework.Source 1 |
The network is part of cost per token
A faster or more efficient network can reduce idle accelerator time, shorten job completion and make a fixed cluster serve more work. Co-packaged optics may lower link power and improve signal integrity, but it also changes serviceability and the way optical components are integrated. The economically relevant measure is not the cheapest switch; it is the infrastructure cost required to deliver an accepted model result within a time and reliability target.
Accelerators remain the visible centre of AI infrastructure, but a cluster is a distributed machine. Its memory, switches, interfaces, optics, power and software determine whether costly chips operate together or wait on one another. The next performance frontier may be won not by a single faster processor, but by making the spaces between processors behave like part of the computer.
Sources & further reading
- Broadcom Redefines AI Infrastructure with Industry-Leading Networking Innovations at 2026 OCP Global Summit Broadcom · 8 Oct 2026 · Accessed 8 Oct 2026
- Broadcom Now Shipping World’s First 102.4 Tbps Switch in Production Volume Broadcom · 12 Mar 2026 · Accessed 8 Oct 2026
- Broadcom Introduces Industry’s First 800G AI Ethernet NIC Broadcom · 14 Oct 2025 · Accessed 8 Oct 2026
- Broadcom Ships Jericho4, Enabling Distributed AI Computing Across Data Centers Broadcom · 4 Aug 2025 · Accessed 8 Oct 2026
- Broadcom Announces Tomahawk 6 – Davisson, the Industry’s First 102.4-Tbps Ethernet Switch with Co-Packaged Optics Broadcom · 8 Oct 2025 · Accessed 8 Oct 2026