NVIDIA Put CPO Switches Into Production, but the Hard Part Starts Now

NVIDIA says its first CPO Ethernet switches are now in production, moving a long-promised optical networking design closer to commercial deployment. The milestone matters because co-packaged optics places optical engines beside the switch chip, instead of relying on removable transceivers at the front panel.

However, NVIDIA did not make this announcement in August. The company disclosed the production status on May 31, 2026, during GTC Taipei. Its current materials still describe Spectrum-X Ethernet Photonics as available during the second half of 2026.

That distinction frames the real story. NVIDIA has crossed from product development into manufacturing, but it has not disclosed shipment volume, customer deployment scale, field reliability, or operating costs. Broadcom and suppliers of conventional pluggable optics still have room to contest how quickly data centers adopt the new architecture.

What NVIDIA Actually Put Into Production

The confirmed milestone is a production start for Spectrum-X Ethernet Photonics, not proof of broad deployment across hyperscale data centers.

NVIDIA tied the announcement to the production ramp for its Vera Rubin computing platform. In its dated production announcement, the company called Spectrum-X Ethernet Photonics the first CPO switch family using 200Gb/s SerDes.

A SerDes converts data between parallel chip interfaces and high-speed serial links. In this system, those links connect the Spectrum-6 switch silicon to optical engines inside the same package.

The wording requires care. NVIDIA said Spectrum-X Ethernet Photonics was “now in production,” while its product page listed availability in the second half of 2026. Neither statement identifies units produced, accepted customer orders, or the number of live production clusters.

Production can cover several stages in a hardware ramp. A company might be building qualified systems for early customers while suppliers increase component output. That is different from sustained, high-volume deployment across many data centers.

NVIDIA has disclosed two main Spectrum-X configurations. The SN6810 uses one Spectrum-6 switch ASIC and provides 102.4 terabits per second of total bandwidth. It exposes 128 ports operating at 800 gigabits per second.

The larger SN6800 combines four switch ASICs with an integrated fiber shuffle. NVIDIA specifies 409.6 terabits per second across 512 ports at 800 gigabits per second. The design can also support 2,048 connections at 200 gigabits per second.

These are scale-out switches, meaning they connect many server racks and GPU systems into one computing fabric. They are separate from the optical NVLink technology associated with tighter scale-up connections among processors.

Confusing those two markets creates a misleading picture of NVIDIA’s progress. Spectrum-X CPO is entering production for Ethernet scale-out networks. It does not mean every connection inside a Rubin system has moved to co-packaged optics.

The company identified CoreWeave, Lambda, and Oracle Cloud Infrastructure as early ecosystem partners and adopters. NVIDIA has not published cluster sizes or deployment dates for those partners. It also has not separated evaluation systems from customer-facing production capacity.

The production announcement nevertheless closes an important verification gap. NVIDIA introduced its photonics switch strategy in March 2025, with commercial availability planned for 2026. The May 2026 update confirms that the Ethernet product moved beyond a conference prototype.

NVIDIA’s updated silicon photonics page also describes Spectrum-X Ethernet Photonics as reaching production. Its list of manufacturing and component partners includes TSMC, Foxconn, Coherent, Corning, Fabrinet, Lumentum, Senko, SPIL, and Sumitomo Electric.

This partner list shows why the transition took more than a switch-chip redesign. CPO requires photonic dies, external lasers, fiber connections, packaging, testing, cooling, and final system assembly to work as one manufacturing flow.

The milestone is therefore credible and dated. The stronger claim that NVIDIA has reached unrestricted, full-volume CPO production remains unsupported by public shipment data.

Why CPO Has Become an Urgent NVIDIA Priority

CPO is arriving because electrical loss and optical-module power are becoming harder to tolerate as AI networks add faster links and more endpoints.

A conventional optical switch places removable transceivers along its front panel. Electrical signals leave the switch ASIC, cross the package and circuit board, enter a module, and then become optical signals.

That modular design has clear benefits. Operators can replace a failed transceiver without discarding the switch. They can also source standards-based modules from multiple vendors and select optics for different distances.

The electrical path becomes more demanding as lane speeds rise. Higher-frequency signals lose more energy while crossing traces, connectors, and module interfaces. Signal conditioning can recover data, but it adds power, latency, components, and heat.

CPO shortens that electrical journey. Optical engines sit beside the switch ASIC, and fiber carries data away from the package. This arrangement removes much of the long, high-speed electrical path between the switch and front-panel optics.

NVIDIA says its design reduces channel loss from roughly 22 decibels to about 4 decibels. The company also claims per-port power can fall from about 30 watts to 9 watts in the compared configuration.

Those figures come from NVIDIA rather than independent customer testing. They still illustrate the physical problem driving the architecture. A few saved watts per connection become significant when one switch supports hundreds of high-speed ports.

The network also competes with GPUs, CPUs, storage, and cooling systems for a data center’s limited power budget. Any reduction in networking power can leave more capacity available for computing.

NVIDIA’s January 2026 engineering details claim five times lower power per 1.6-terabit port than a pluggable design. The company also claims five times longer operation between link flaps and ten times greater network resiliency.

A link flap occurs when a network connection repeatedly drops and recovers. In distributed AI training, one unstable link can delay synchronization across many accelerators. Network reliability can therefore affect expensive computing resources far beyond the failed connection.

NVIDIA has designed the switch for Mixture of Experts models as well. These models activate selected groups of parameters for each input, creating frequent communication among GPUs hosting different experts.

When traffic arrives unevenly or suffers jitter, some processors wait while others finish their work. NVIDIA argues that its combined Ethernet hardware and software can maintain more predictable collective communication.

CPO does not create that software behavior by itself. Spectrum-X combines switch silicon, ConnectX networking hardware, congestion control, telemetry, and NVIDIA Collective Communications Library integration. The optical package removes one physical constraint within a larger system.

The company also says CPO eliminates digital signal processing retimers from these links. Removing retimers can reduce latency and component count, but the optical package must deliver comparable signal quality without the same modular safety net.

Manufacturing explains why the production milestone matters. NVIDIA developed a detachable fiber connector that allows fiber attachment late in assembly. Its optical engines are compatible with solder reflow and can undergo screening before integration with the expensive switch package.

NVIDIA says this process allows manufacturers to use known-good optical engines. That phrase describes component screening before final assembly, not a guarantee that every finished switch will operate without defects.

Automated fiber attachment also addresses a longstanding obstacle. Aligning many fibers with small photonic structures at commercial yield is harder than placing ordinary electronic components. Production volume depends on repeating that alignment accurately and quickly.

The architecture therefore responds to real physical constraints. Whether it becomes the default depends on economics and operations, not bandwidth alone.

CPO Versus Pluggable Optics Is Now a Commercial Contest

NVIDIA is pressuring pluggable optics at the highest network speeds, but it has not made the modular alternative obsolete.

Pluggable transceivers have decades of operational familiarity behind them. Data center technicians understand how to stock, diagnose, replace, and mix these modules. Open interfaces also support a large supplier base.

CPO trades some of that modularity for shorter electrical connections and lower power. The optical engines become part of a more integrated switch assembly, making failures and upgrades more closely tied to the system vendor.

This contest resembles earlier transitions in computing hardware. Integration can improve efficiency and density, while modular components preserve repairability and purchasing flexibility. The winning balance changes as performance requirements rise.

NVIDIA’s strongest case applies to very large AI clusters. These environments contain dense collections of expensive accelerators, run synchronized workloads, and place a high value on every watt saved.

A conventional enterprise network does not face the same pressure. Smaller clusters can continue using pluggable optics when deployment flexibility matters more than maximum bandwidth density.

Near-packaged optics, or NPO, creates another path. NPO moves optical engines closer to the switch chip but retains more physical separation than full co-packaging. It seeks a middle ground between electrical efficiency and modular service.

Industry research firm LightCounting reported in July that NVIDIA networking executive Gilad Shainer confirmed CPO manufacturing for scale-out systems. Its industry conference summary also said CPO and NPO would coexist.

That coexistence weakens any simple claim that one architecture will replace every other optical link. Operators can deploy CPO where power density justifies deeper integration and use pluggable or near-packaged designs elsewhere.

Broadcom represents the clearest competitive pressure. It has worked on CPO for several switch generations and supplies Ethernet switching silicon used across major cloud environments.

Broadcom announced Tomahawk 6 Davisson in 2025, combining a 102.4-terabit Ethernet switch with optical engines based on TSMC’s COUPE platform. Its Davisson design also operates at 200 gigabits per second per channel.

The Broadcom comparison matters because NVIDIA is not competing only on photonics. It is asking customers to adopt a vertically coordinated network containing NVIDIA switches, adapters, software, and computing systems.

That integration can simplify performance tuning for buyers already standardizing on NVIDIA infrastructure. It can also increase their dependence on one vendor’s roadmap and management environment.

Broadcom supports a more disaggregated market of switch manufacturers, cloud operators, and optical suppliers. Some hyperscalers prefer that model because they develop their own network software and negotiate across several hardware sources.

The competition will therefore turn on more than switch bandwidth. Buyers will compare power consumption, failure rates, maintenance procedures, delivery schedules, software integration, and supplier diversity.

Pluggable optics suppliers also have responses. Linear pluggable optics can reduce or remove some digital signal processing inside modules. Faster module standards can preserve replaceability while narrowing part of CPO’s power advantage.

CPO still offers the shortest electrical path. However, the best laboratory power result does not automatically deliver the best total operating result. Cooling equipment, spare systems, external lasers, fiber management, and replacement procedures all affect actual efficiency.

The likely near-term market is mixed. The largest AI clusters can absorb a tightly integrated design when each efficiency gain unlocks more computing. General cloud and enterprise networks can wait for stronger field evidence.

NVIDIA’s production start forces that evaluation to become concrete. Customers can now compare orderable systems and deployment plans instead of architecture slides.

Production Success Is Not Deployment Proof

The unresolved question is whether tightly integrated optics can preserve its efficiency advantage through manufacturing, maintenance, and years of field operation.

CPO places heat-sensitive optical components close to a high-power switch ASIC. That proximity improves electrical performance but creates a more difficult thermal environment.

NVIDIA uses liquid cooling for its photonics switches. Liquid cooling can remove dense heat more effectively than conventional air systems, but it adds facility requirements and operational procedures.

Thermal behavior matters because optical modulators and lasers can change performance as temperature moves. NVIDIA keeps laser sources outside the main switch package, which separates a sensitive and failure-prone component from the hottest silicon.

External lasers also improve serviceability because operators can replace a laser module without replacing the entire switch package. Yet failures inside a co-packaged optical engine remain harder to repair than failures in front-panel transceivers.

An IEEE Electronics Packaging Society review identifies signal integrity, packaging, thermals, reliability, and field replacement as central CPO challenges. Its CPO technical review notes that most optical-engine components inside a switch are not field-replaceable.

That does not make CPO unreliable. It raises the standard required before operators trust the integration. Optical-engine failure rates must approach the reliability expected from the switch ASIC and surrounding system.

NVIDIA says Spectrum-X uses fewer lasers and active components than a comparable pluggable network. Fewer parts can reduce the number of failure points, while redundant external laser sources can maintain service after individual failures.

The company also claims ten times greater resiliency and five times longer link-flap-free operation. Public materials do not disclose the test duration, workload mix, comparison system, or number of deployed switches behind those ratios.

Independent field data would carry more weight. Buyers need failure rates across months of continuous training and inference, not only component screening or internal qualification.

Service procedures present another question. A technician can replace a pluggable module while leaving the switch installed. A failed internal optical engine might require swapping a larger switch tray or returning the assembly for repair.

NVIDIA argues that its detachable fiber connector and reduced component count make systems easier to install and replace. That claim needs validation under real data center conditions, where fiber routing, contamination, maintenance windows, and spare inventory affect recovery time.

Manufacturing yield also deserves scrutiny. A switch package combining expensive logic and many photonic engines concentrates value in one assembly. One defective component can threaten more finished-product value than a defective standalone transceiver.

NVIDIA says it screens optical engines before attachment and uses automated pick-and-place assembly. These choices directly address yield risk. The company has not disclosed final package yield or how that yield changes as volume grows.

Supply capacity could become another constraint. The partner list spans foundries, packaging specialists, fiber companies, laser suppliers, connector makers, and system manufacturers. Weak output at one stage can slow the whole ramp.

The lack of public shipment data is therefore important. “In production” confirms that NVIDIA has a qualified manufacturing path. It does not disclose whether the company can deliver thousands of switches at stable yield.

There is also a standards question. Pluggable optics benefit from mature form factors and procurement practices. CPO interfaces and packaging arrangements remain less uniform across vendors.

A hyperscaler can accept vendor-specific integration when it delivers a clear operating advantage. Smaller buyers usually prefer interchangeable parts, broad support, and multiple qualified suppliers.

NVIDIA’s initial customers are therefore informative. CoreWeave, Lambda, and Oracle operate infrastructure where dense AI workloads can justify specialized networking. Their deployments can provide the first evidence about whether the efficiency claims survive outside NVIDIA’s labs.

The central risk is not that CPO fails technically. NVIDIA has shown enough engineering and manufacturing work to move beyond that binary question.

The risk is that CPO remains concentrated in a limited group of large clusters because serviceability, cost, or supply volume slows broader adoption. That outcome would still create a meaningful market without replacing pluggable optics.

NVIDIA Is Expanding Its Control of the AI System

The CPO launch matters strategically because NVIDIA is turning networking into another coordinated layer of its AI computing platform.

NVIDIA’s data center position began with accelerators and software, but modern AI performance depends on much more than GPU arithmetic. Memory bandwidth, processor links, storage traffic, and scale-out networking can all limit useful throughput.

The company has assembled products for each layer. NVLink connects processors within tightly coupled systems. InfiniBand and Spectrum-X address larger network domains. BlueField data processing units handle infrastructure and security tasks.

Spectrum-X Ethernet Photonics extends that approach into the optical layer. NVIDIA now coordinates the switch ASIC, optical package, network adapter, congestion controls, collective communication software, and Rubin computing platform.

This coordination can produce measurable benefits because engineering teams can optimize the entire path around AI workloads. A switch vendor serving many applications must support a broader range of systems and operating models.

NVIDIA can also test networking behavior against its own GPU communication libraries. That gives the company visibility into application-level effects that component suppliers may not receive.

The same structure creates buyer concerns. A customer adopting the complete stack becomes more exposed to NVIDIA’s release schedule, product qualification, software policies, and supply allocation.

Broadcom and merchant optical suppliers provide an alternative route. Customers can combine switch silicon, network operating software, modules, and server platforms from different sources.

The merchant model can encourage competition and customization. It can also require more integration work, while responsibility for performance problems is distributed among several vendors.

CPO sharpens this conflict because the optics are no longer a simple component at the edge of the switch. Package design, thermal management, fiber attachment, and manufacturing yield become part of the switch platform.

NVIDIA’s partnership network does not mean the company manufactures every component. TSMC and other suppliers still provide essential production capabilities. However, NVIDIA controls the product architecture and customer-facing system.

Optical-component vendors face a complicated transition. A move away from pluggable modules can reduce demand for some traditional products. At the same time, CPO requires lasers, fibers, connectors, photonic dies, and packaging services.

The value does not disappear. It moves to different components and stages of assembly.

Coherent and Lumentum appear on NVIDIA’s photonics partner list, showing that established optical suppliers can participate in the new architecture. Corning and Sumitomo Electric contribute fiber expertise, while Foxconn and other manufacturers support assembly.

For cloud providers, the practical question is whether one coordinated stack improves completed AI work enough to justify reduced modularity. That calculation will differ by workload and deployment scale.

Large training runs are sensitive to network delays because thousands of GPUs must exchange data repeatedly. An unstable or congested fabric can leave accelerators idle while consuming power and occupying costly capacity.

Inference creates different patterns, but large models can still require communication across many processors. Mixture of Experts systems make consistent network timing more valuable because requests can activate experts located on separate devices.

If Spectrum-X delivers lower power and fewer interruptions in those settings, NVIDIA can sell networking as part of the economics of each generated token. That is a stronger position than selling a faster switch in isolation.

The company still needs to prove that integration at purchase time does not become inflexibility later. Customers will watch whether systems support practical upgrades, manageable repairs, and competitive component sourcing.

Teams evaluating these systems will also generate large collections of qualification reports, architecture decisions, incident records, and vendor materials. A searchable knowledge base can help engineers retain that evidence across a long infrastructure review.

The strategic result is clear. CPO gives NVIDIA another way to define the preferred shape of an AI data center. The commercial result remains open because buyers still control how much of that design they adopt.

Three Signals Will Show Whether CPO Has Truly Arrived

The next stage should be judged by customer deployments, field reliability, and competitive product shipments, not another set of peak specifications.

The first signal is disclosed deployment scale. NVIDIA or its early partners need to identify live Spectrum-X Ethernet Photonics clusters, their approximate size, and the workloads running on them.

A named production cluster would strengthen NVIDIA’s claim that the switch has moved from manufacturing into operational use. Multiple customers running sustained workloads would provide stronger evidence than an isolated qualification system.

The most useful disclosure would connect infrastructure scale with measurable outcomes. Power per delivered bit, network availability, completed training time, and accelerator utilization would reveal more than raw switch bandwidth.

If customer deployments remain vague through the second half of 2026, the production announcement will look more like an early ramp than broad commercialization. That would not invalidate the technology, but it would weaken claims about immediate industry-wide change.

The second signal is field reliability. Operators should look for data on optical-engine failures, laser replacement, link flaps, switch-tray swaps, maintenance time, and spare-system requirements.

NVIDIA’s five-times uptime and ten-times resiliency claims establish a benchmark. Independent customer results need to show whether those advantages survive changing temperatures, continuous workloads, and repeated maintenance cycles.

A strong result would combine low failure rates with simple replacement procedures. If customers must hold expensive spare trays or schedule complex repairs, some energy savings could be offset by operational costs.

Published reliability evidence would strengthen the case for CPO beyond the largest specialist operators. Limited or mixed evidence would support continued use of pluggable and near-packaged alternatives.

The third signal is competitive delivery. Broadcom-based CPO systems, new NPO products, and improved pluggable optics will reveal how much of NVIDIA’s advantage comes from CPO itself.

If several vendors ship qualified CPO switches, the architecture will look like an industry transition rather than an NVIDIA-specific platform choice. Open interfaces and multiple suppliers would also reduce adoption risk for buyers.

If NPO gains faster customer acceptance, the market may favor a compromise between integration and serviceability. Continued progress in linear pluggable optics could preserve front-panel modules in more network tiers.

NVIDIA has already cleared an important boundary. Its CPO Ethernet switch is no longer just a roadmap item, and the official production date is May 31, 2026.

The silicon photonics era will not begin with one announcement, however. It will begin when customers publish evidence that optical integration improves useful AI work without creating unacceptable manufacturing or maintenance costs.

Infrastructure buyers should now ask vendors for deployment records, measured power, failure data, replacement procedures, and supply commitments. Those answers will determine whether CPO becomes the standard fabric for large AI clusters or one specialized option within a mixed optical market.