Nvidia has started shipping its next-generation Spectrum-X Ethernet switches that use co-packaged optics to selected partners, while Broadcom continues limited shipments of its 51.2Tbps Bailly platform. Both moves mark the beginning of a critical manufacturing transition for networking inside AI data centers, where traditional pluggable transceivers struggle to keep pace with skyrocketing bandwidth demands.

The shift from pluggable modules to onboard optics

Co-packaged optics, or CPO, mounts the optical engines directly next to the switching silicon on the same package, eliminating the need for removable transceivers at the front of the switch. That change shortens the electrical paths that signals must travel, which reduces power consumption and signal degradation—problems that worsen as networks push toward 200Gb/s-per-lane signaling and total switch bandwidth climbs beyond 50 terabits per second.

Nvidia’s new Spectrum-X Ethernet Photonics switches are built to be part of its Vera Rubin AI platform. The company claims the switches are now in production and that the photonics approach improves power efficiency and deployment speed versus conventional transceiver-based networks. According to market intelligence firm TrendForce, which first reported the production ramp, Nvidia’s design was developed with TSMC and uses the foundry’s COUPE advanced-packaging technology. The reported switching capacity of 400 terabits per second puts the product squarely in the realm of fabrics needed for GPU clusters that can scale to a million accelerators or more, not in typical enterprise wiring closets.

Broadcom’s Bailly is a 51.2Tbps Ethernet switch built around its Tomahawk 5 ASIC and eight 6.4Tbps silicon-photonics optical engines. Broadcom had previously described the platform as production-ready, but TrendForce’s latest assessment indicates shipments remain limited. The distinction is important: being production-capable on paper is different from having a scalable supply chain. CPO brings together the optical engine, switch silicon, package substrate, fiber attachment, thermal solution, and test process into a single integrated product. A yield hiccup in any one component can delay the whole assembly.

Why this matters for Windows users and administrators

For everyday Windows users, the arrival of CPO switches inside hyperscale data centers will be invisible but meaningful over time. Cloud services—including AI features inside Windows, Microsoft 365 Copilot, and any Azure-based application—rely on massive backend clusters whose performance is often gated by networking. Faster, more power-efficient interconnects between GPUs mean those services can handle more simultaneous users, deliver responses with lower latency, and possibly run more sophisticated models without breaking the cloud provider’s energy budget.

Enterprise IT professionals who manage Windows Server environments or hybrid Azure deployments should pay closer attention. Many organizations now test or host AI-adjacent workloads on Windows Server, using services like Azure Stack HCI or on-premises GPU clusters managed by Windows Admin Center. If your roadmap includes building a private AI infrastructure, the networking layer will soon become a harder constraint than the number of GPUs you can buy. CPO switches won’t appear in your procurement catalog next quarter, but they signal the direction of travel for the high-end Ethernet fabrics that Microsoft Azure and other cloud providers use internally. That, in turn, affects the service level agreements and capabilities those providers can offer you.

For developers working on AI applications, the immediate takeaway is that the tools you use to orchestrate distributed training or inference—whether that’s Azure Machine Learning, PyTorch on Windows Subsystem for Linux, or third-party MLOps platforms—will increasingly be shaped by the underlying network topology. Nvidia’s Spectrum-X is not just faster hardware; it comes with adaptive routing and congestion-control software that can dynamically steer traffic away from congested links. As these capabilities trickle into SDKs and cloud APIs, they’ll influence how you design parallel workloads, perhaps reducing the manual tuning needed to keep GPUs fed with data.

How the AI buildout forced a networking rethink

The push for co-packaged optics didn’t come from a desire to innovate for innovation’s sake. It’s the result of power and signal-integrity walls that traditional networking hits when connecting tens of thousands of accelerators in a single fabric. A state-of-the-art GPU like Nvidia’s Blackwell can spit out data at 1.8 terabytes per second; linking thousands of them with conventional copper traces and pluggable optical modules becomes wildly expensive in both electricity and heat.

Pluggable transceivers—the small modules you snap into a switch’s front panel—have served data centers well for decades. But as per-lane speeds climb, the electrical trace from the ASIC to the front of the switch acts like an increasingly lossy and power-hungry antenna. CPO solves that by moving the conversion from electrical to optical inside the package, mere millimeters from the chip. The trade-off is manufacturability: you can’t just unplug a faulty optical engine and replace it; the entire switch assembly becomes more monolithic.

Nvidia acquired Mellanox for $6.9 billion in 2019, giving it control over high-speed Ethernet and InfiniBand. Since then, it has married its own GPU roadmap with a networking one, culminating in the Spectrum-X brand. Broadcom, already dominant in merchant switch silicon with Tomahawk and Jericho families, took a silicon-photonics route through its own internal development. Both paths now converge on the same conclusion: the next generation of AI clusters will need optical engines integrated at the switch level.

What you should do now

For most organizations, the correct posture today is “watch and plan,” not “buy.” CPO switches are not yet available through general distribution, and even if they were, they’d be grossly overpowered—and overpriced—for anything short of hyperscale GPU fabrics. TrendForce estimates that CPO penetration in AI data-center optical communications will still be only around 0.5% in 2026. This is a manufacturing transition, not an overnight replacement of pluggable optics.

If you are a network architect or IT director who might one day spec out an AI cluster,
- Track the component supply chain, not just the switch specs. Yields on optical engines and advanced packaging capacity at TSMC are the hidden levers that will determine when CPO becomes cost-effective for enterprises, not just hyperscalers.
- Evaluate your future AI workloads honestly. If your training jobs can fit on a few dozen GPUs inside a single rack, you likely won’t need CPO in the foreseeable future. If you’re planning a multi-rack, hundred-plus GPU setup, ask your server and networking vendors for their CPO roadmaps now.
- Keep an eye on software. Nvidia’s Spectrum-X differentiates itself as much through adaptive routing and telemetry as through raw bandwidth. Even if you’re not ready for new hardware, adopting similar software-defined networking principles on your existing Ethernet fabrics can reduce tail latency and boost GPU utilization.

Microsoft Azure customers should pay attention to any announcements around upgraded networking tiers. Azure’s AI infrastructure already uses Nvidia GPUs and Mellanox-derived networking. When Microsoft adopts CPO internally, it will likely market new instance families with improved inter-GPU bandwidth. Being an early evaluator of those instances could give your AI projects a head start.

What comes next

The next 12 to 18 months will be about yields, not about who ships the fastest switch. Both Nvidia and Broadcom have demonstrated working silicon, but scaling production requires solving the simultaneous equations of silicon photonics capacity, advanced packaging availability, and optical-engine test times. If those bottlenecks ease—and history suggests they will, though not on a predictable schedule—CPO could follow a trajectory similar to 400G pluggable optics: niche at first, then quickly table stakes.

For the Windows ecosystem, the real inflection point will arrive when CPO-derived networking appears in the service-level guarantees of cloud AI services. That’s when the technology stops being an infrastructure curiosity and starts changing how everyday applications behave. Until then, consider this the moment the plumbing got a major upgrade, even if you can’t see it behind the walls.