NVIDIA CPO Switches Enter Production as AI Data Centers Embrace Silicon Photonics

NVIDIA says its first Spectrum-X Ethernet Photonics switches are now in production, giving substance to a widely circulated Google News headline about mass production. The announcement moves co-packaged optics, or CPO, from a product roadmap toward commercial AI infrastructure. However, NVIDIA has not disclosed shipment volumes, customer deployments, manufacturing yields, or generally available system configurations.

That distinction matters. Production status confirms that NVIDIA and its manufacturing partners have crossed an important engineering threshold. It does not establish that CPO has already displaced conventional pluggable optical modules across data centers.

The real contest is therefore not NVIDIA against a single chipmaker. It is CPO against the established pluggable optics architecture that operators already understand, source, replace, and maintain. NVIDIA is betting that the power and reliability demands of larger GPU clusters will outweigh the operational advantages of removable transceivers.

Broadcom has pursued the same architectural shift for several years. Its earlier CPO systems also show why NVIDIA’s announcement is not the beginning of silicon photonics. What changed is NVIDIA’s decision to connect CPO directly with the Rubin computing platform, Spectrum-X networking, and its broader AI factory strategy.

What the Google News Headline Gets Right

NVIDIA has moved Spectrum-X Ethernet Photonics into production, but “full-scale mass production” remains a broader claim than its public evidence supports.

NVIDIA announced in 2025 that it would bring silicon photonics to both its Ethernet and InfiniBand networking families. Spectrum-X Ethernet Photonics was scheduled for the second half of 2026, while Quantum-X Photonics targeted an earlier introduction.

The company’s newer Rubin production update now describes Spectrum-X Ethernet Photonics as “in production.” NVIDIA also calls it the first CPO switch with 200 gigabit-per-second SerDes, the high-speed electrical interfaces inside the switch.

That is the strongest available confirmation behind the headline. The statement comes from NVIDIA rather than an anonymous supply-chain source or a financial discussion board.

The wording still needs careful interpretation. “In production” can cover several stages, including initial manufacturing, customer qualification, controlled system integration, or a wider commercial ramp. It does not automatically mean high-volume shipments across multiple customers.

NVIDIA’s release says system builders are in full-scale production of Vera Rubin equipment. That partner list includes Dell Technologies, HPE, Lenovo, Supermicro, Foxconn, Quanta, Wistron, and Wiwynn. The statement does not say every partner is building CPO-equipped switches at the same volume.

The company has also retained an explicit availability window. Its silicon photonics page says Spectrum-X Ethernet Photonics will be available in the second half of 2026. Production can begin before customers receive broadly available systems.

This is why the Google News wording should not be treated as a shipment report. No public disclosure identifies units produced, ports deployed, revenue recorded, or operational clusters using the finished switches.

Industry research points to an early ramp rather than universal availability. TrendForce reported in late July that NVIDIA and Broadcom had begun limited CPO shipments. It also identified optical-engine yield and advanced packaging capacity as constraints.

According to the TrendForce assessment, NVIDIA developed the Spectrum-X switch with TSMC and expects capacity to expand during the second half of 2026. The report describes a 400-terabit-per-second class system, consistent with NVIDIA’s largest announced configuration.

“Limited shipments” and “full-scale mass production” are not interchangeable descriptions. Together, the available sources support a narrower conclusion: production has started, commercial expansion is planned, and the scale of that expansion remains undisclosed.

NVIDIA’s announcement still marks a real transition. CPO is no longer presented only through engineering samples, conference slides, or a future product schedule. It has entered the manufacturing phase of NVIDIA’s next computing platform.

That creates the article’s central tension. NVIDIA has cleared enough technical barriers to call the product in production, yet customers still lack public evidence about deployment economics and serviceability at scale.

Why NVIDIA Is Moving Optical Engines Next to the Switch

CPO addresses the growing electrical distance between switch silicon and optical conversion, where faster links consume more power and become harder to maintain.

A conventional optical transceiver plugs into the front of a network switch. Electrical signals travel from the switch application-specific integrated circuit, or ASIC, across the circuit board before reaching that module. The transceiver then converts those signals into light.

That design has served data centers well because technicians can replace a failed module without replacing the entire switch. Operators can also source compatible components from several vendors and introduce newer optics without redesigning the switch package.

The electrical journey becomes more difficult as each lane carries more data. Signal loss rises with frequency, board traces require additional processing, and connectors add more loss. Retimers and digital signal processors compensate for those weaknesses, but they consume electricity and produce heat.

CPO shortens that electrical path. The architecture places silicon photonic engines beside the switch ASIC inside the same package. Data reaches the optical engine through a much shorter electrical connection, then travels through fiber.

NVIDIA says this approach lowers networking power and removes many active components found in pluggable designs. Its published claims vary by product generation and comparison method, which makes the baseline important.

For its production Spectrum-X platform, NVIDIA claims five times better power efficiency than networks using traditional transceivers. It also claims five times longer sustained application runtime and 1.3 times faster deployment.

Those figures remain company claims. NVIDIA has not released complete field data, independent benchmarks, workload definitions, or failure records that would let outside engineers reproduce each comparison.

An earlier CPO engineering analysis presented a different generation and baseline. It described up to 3.5 times greater power efficiency and ten times greater network resiliency than pluggable transceivers.

The changing figures do not necessarily conflict. NVIDIA has introduced newer switches and faster SerDes since the first announcement. However, they show why readers should connect each performance claim to a specific platform.

The target scale explains NVIDIA’s urgency. The company is designing networks intended to connect extremely large GPU installations. At that size, networking does more than move data between independent servers.

Distributed AI training requires thousands of accelerators to exchange partial results repeatedly. If links pause or fail, accelerators can sit idle while the job waits. A failure can also force software to restart work from a checkpoint.

This means optical power and link stability affect expensive computing resources beyond the switch itself. Saving energy in the network can release electrical and cooling capacity for additional accelerators. Avoiding a link interruption can protect hours of coordinated computation.

Spectrum-X Ethernet Photonics is built around NVIDIA’s Spectrum-6 switch ASIC. NVIDIA has announced a 102.4-terabit-per-second SN6810 configuration with 128 ports running at 800 gigabits per second.

The larger SN6800 design reaches 409.6 terabits per second through 512 ports at the same speed. NVIDIA positions that system for the scale-out network connecting many computing racks.

These port counts create the physical problem CPO is meant to solve. Hundreds of high-speed pluggable modules can impose a substantial power and cooling burden around the switch faceplate. They also create hundreds of removable connections that require monitoring and maintenance.

NVIDIA has paired the switches with external laser sources. Separating the laser from the hottest switch components can improve thermal management and allow technicians to service the light source independently.

That choice does not make the entire optical path modular. The photonic engines remain closely integrated with the switch package. A failure inside that assembly can therefore carry different repair consequences than a failed front-panel transceiver.

CPO’s mechanism is persuasive at very high bandwidth. Its operational value is less automatic. Operators must weigh lower link power and fewer active components against more concentrated packaging and maintenance risks.

Pluggable Optics Now Faces a Full-Stack Competitor

NVIDIA is pressuring the pluggable model by tying optical switching to its GPUs, network adapters, software, and complete Rubin system roadmap.

A component vendor can offer a more efficient optical engine without changing how data centers buy networks. NVIDIA has a different route. It can specify the GPUs, server architecture, network adapters, switches, management software, and communication libraries together.

That control gives NVIDIA several advantages. It can tune collective communication around known switch behavior. It can coordinate system qualification across its partner network. It can also introduce CPO as part of a complete Rubin deployment instead of asking customers to adopt an isolated component.

The approach narrows the number of unresolved interfaces during an early technology ramp. NVIDIA can validate the network against its own platform and software stack before supporting every possible server design.

This same integration can increase customer dependence on NVIDIA’s architecture. Operators adopting Spectrum-X already accept NVIDIA’s congestion control, telemetry, network adapters, and software. CPO adds packaging and optical supply relationships to that stack.

Pluggable optics follows a more modular model. Switch suppliers, transceiver manufacturers, cable vendors, and cloud operators can qualify components separately. A customer can replace an optical module without replacing the switch ASIC package.

That modularity supports competition and field repair. It also introduces more electrical interfaces and active devices, especially as port speed rises. The established model is therefore under pressure for technical reasons, not only because NVIDIA wants a larger market.

Broadcom demonstrates that CPO is an industry direction rather than an NVIDIA-only invention. The company showed an early 25.6-terabit-per-second Humboldt platform before developing its 51.2-terabit Tomahawk 5 Bailly system.

Broadcom says it launched Bailly in 2024 with eight 6.4-terabit-per-second optical engines. Its architecture integrates 512 optical channels with the Tomahawk 5 switch.

In a later CPO manufacturing update, Broadcom described automated fiber attachment, manufacturing-test development, and the practical work required to move beyond prototypes. Those details underscore the challenge facing every supplier.

Broadcom’s experience also complicates any claim that NVIDIA has single-handedly started the silicon photonics era. CPO systems, prototypes, and customer programs existed before Spectrum-X Photonics entered production.

NVIDIA’s contribution is different. It is attaching the technology to a widely anticipated GPU platform and an expanding Ethernet product. That can push CPO into purchasing decisions for complete AI clusters.

The competitive field extends beyond finished switches. TSMC’s packaging capability is central because CPO brings electronics and photonics into a tightly integrated assembly. Optical-engine suppliers, fiber specialists, laser vendors, and test-equipment companies must also meet production tolerances.

Traditional transceiver vendors do not simply disappear. CPO systems still require lasers, fiber connections, connectors, and optical components. Data centers will also continue using pluggable optics in network layers where CPO’s density advantage does not justify its integration cost.

The transition is likely to be selective. The largest AI scale-out fabrics face the strongest power and bandwidth pressure. Enterprise networks, storage systems, and smaller clusters have different economics and maintenance requirements.

NVIDIA’s own product demonstrations reflect this mixed environment. It has shown CPO switches connecting systems that also use OSFP pluggable modules elsewhere in the topology. CPO can enter one network layer without replacing every transceiver.

This hybrid path reduces the immediate threat to the installed optical supply chain. It also gives operators a way to test CPO where its benefits are strongest while preserving pluggable components in familiar roles.

The central contest is not a clean winner-takes-all battle. It concerns which network links become sufficiently dense, hot, and failure-sensitive to justify tighter optical integration.

NVIDIA’s full-stack position gives it leverage in that decision. Customers buying Rubin systems may evaluate CPO as part of total cluster performance rather than as an independent networking purchase.

Production Does Not Settle the Reliability Question

The decisive test is whether CPO improves cluster uptime without creating unacceptable manufacturing, repair, and supplier-concentration costs.

NVIDIA emphasizes that CPO removes components that can fail. Shorter electrical paths require less signal conditioning, while integrated optical engines reduce the number of connectors between the ASIC and optical conversion.

That argument addresses one form of reliability. Fewer active devices and shorter signal paths can reduce the number of link interruptions. NVIDIA describes this benefit through “link flap-free” runtime, meaning longer operation without a link repeatedly disconnecting and reconnecting.

However, reliability also concerns the cost and scope of a repair. A technician can remove a failed pluggable transceiver from the front of a switch. A faulty co-packaged optical engine sits much closer to expensive switch silicon.

Designers can introduce replaceable fiber connectors and external laser sources, but those choices do not make every failure field-serviceable. Operators need evidence about which components fail, how failures are isolated, and what must be replaced.

Manufacturing yield presents a related concern. Combining advanced switch silicon, photonic engines, packaging, fiber attachment, and thermal systems creates more points where a defect can reduce usable output.

Automated assembly can improve repeatability. NVIDIA says its manufacturing process attaches fibers during a final production stage using precision machinery. Yet the company has not published production yields or rejection rates.

TrendForce specifically identified optical-engine yield and advanced packaging capacity as constraints during the early shipment period. That warning matters because CPO economics depend on producing complex assemblies consistently.

A low-yield optical engine does more than raise its own cost. If integration occurs early, one defective component can affect the value of other parts in the package. Manufacturers can limit that exposure through testing and assembly sequencing, but execution determines the result.

Thermal behavior also needs field validation. Switch ASICs generate substantial heat, while lasers and photonic components can react to temperature changes. External laser sources remove one major heat-sensitive element from the immediate package.

Liquid cooling adds another dependency. NVIDIA’s largest Spectrum-X Photonics configurations target dense AI systems where liquid-cooled networking becomes practical. Customers must integrate those switches with facility cooling and service procedures.

The question is not whether engineers can make the system work. NVIDIA’s production declaration indicates that it has a manufacturable design. The question is whether large deployments preserve the promised benefits through installation, maintenance, and several years of operation.

Independent workload comparisons will also matter. NVIDIA’s fivefold efficiency claim is not the same as a fivefold reduction in total data-center power. Networking represents only part of a cluster’s electrical load.

A per-port saving can still be valuable across hundreds or thousands of links. Readers should not convert that figure into a facility-wide saving without knowing the network topology, utilization, cooling overhead, and comparison platform.

The uptime claim also needs context. Longer link stability can improve application runtime, but job completion depends on GPUs, memory, storage, software, power delivery, and cooling. Optical networking solves only one set of failure modes.

Standards and interoperability present another uncertainty. Hyperscale operators favor multiple sources when possible because supplier diversity protects availability and pricing. Tightly integrated CPO packages can make substitutions harder than replacing standardized pluggable modules.

NVIDIA can offset that concern through qualified manufacturing partners and predictable system support. Its extensive Rubin partner list helps. Still, the company has not disclosed whether customers can mix optical engines, laser sources, or replacement assemblies from multiple suppliers.

The broad CPO ecosystem continues to develop. Broadcom announced 200-gigabit-per-lane technology for its third CPO generation and described production lessons from Bailly. Its third-generation platform shows that competing roadmaps are advancing alongside NVIDIA’s.

That competition should produce valuable evidence. If several vendors achieve stable yields and interoperable supply chains, CPO becomes less dependent on one company’s manufacturing model.

If deployments remain confined to tightly controlled clusters, the technology can still succeed. Its addressable role would simply be narrower than the silicon photonics era suggested by the headline.

NVIDIA has therefore won an important engineering round, not the entire architectural contest. Production is the point where claims begin facing customer workloads, repair tickets, and supply constraints.

Three Signals Will Reveal Whether the CPO Era Has Arrived

Shipment scale, customer operating data, and competitive product ramps will determine whether NVIDIA’s production milestone becomes an industry transition.

The first signal is broad system availability during the second half of 2026. NVIDIA should identify shipping Spectrum-X Photonics configurations, system partners, and deployment schedules beyond the general Rubin production announcement.

A product catalog is not enough. The stronger evidence would include standard orderable configurations, documented lead times, supported network topologies, and customer acceptance milestones.

If those details appear across several system vendors, the mass-production description gains credibility. If availability remains limited to evaluation systems or selected customers, the production claim should remain narrowly framed.

Shipment or revenue information would provide an even clearer measure. NVIDIA does not currently separate CPO switch units or sales in its public disclosures. Partner commentary may therefore offer the first useful indication of volume.

The second signal is field evidence from an operating AI cluster. Buyers need data covering network power, link interruption rates, application completion time, cooling requirements, and repair procedures.

The most useful comparison would place CPO and pluggable networks under similar workloads and deployment conditions. It should define port speeds, cable distances, utilization, and the components included in each power figure.

Independent testing would strengthen NVIDIA’s efficiency and uptime claims. Customer testimony can also help when it includes specific operating conditions instead of general support for the platform.

Service records will be equally important. A CPO switch can use less power and still create operational friction if optical failures require replacing expensive assemblies or taking larger network segments offline.

Evidence of rapid fault isolation, replaceable external lasers, predictable repair times, and stable optical-engine performance would support NVIDIA’s central argument. Poor yields or difficult field repairs would preserve a major advantage for pluggable optics.

The third signal is the response from Broadcom and other networking suppliers. Broadcom already has production experience and a roadmap extending CPO to faster per-lane signaling.

A wider set of orderable systems would show that customers see CPO as an architectural category rather than one feature of Rubin. Multiple platforms could also accelerate standards, manufacturing tools, and supplier diversity.

The competitive response does not need to copy NVIDIA’s full-stack strategy. Broadcom can serve hyperscalers and switch makers that want more control over system design. Other silicon and optical suppliers can target specialized layers of the network.

A strong competitive ramp would validate the mechanism even if it reduced NVIDIA’s differentiation. Weak adoption outside NVIDIA would suggest that CPO works best when one vendor coordinates computing, networking, and software.

Pluggable optics vendors will provide another important response. Lower-power linear pluggable optics and improved transceivers can extend the existing architecture by reducing signal-processing overhead without moving optical engines into the ASIC package.

Those alternatives preserve field replacement and established sourcing patterns. If they narrow CPO’s power advantage, customers may reserve co-packaged designs for only the densest switches.

This is why calling 2026 the start of a silicon photonics era requires qualification. Silicon photonics already supports many pluggable transceivers, and CPO has a history that predates NVIDIA’s production announcement.

The meaningful change is architectural. NVIDIA is moving optical conversion beside its switch silicon and offering that design as part of a production AI platform.

Readers should treat the Google News headline as a prompt to examine the transition, not as proof that the transition is complete. NVIDIA has confirmed production, published major performance claims, and connected the product to Rubin.

It has not disclosed the volume, yield, deployment, or operating evidence needed to confirm full-scale adoption. Those gaps are normal at the beginning of a hardware ramp, but they remain central to the story.

Over the next three months, watch for orderable partner systems, named customer deployments, and comparable field measurements. Each would strengthen the case that CPO is leaving specialized programs for mainstream AI infrastructure.

If those signals do not appear, production may still be real while commercial scale develops more slowly. That outcome would not invalidate CPO, but it would preserve a larger role for pluggable optics.

For developers and AI product teams, the consequence is indirect but important. Better networking can reduce the infrastructure wasted while accelerators wait for data or recover from unstable links.

For enterprise buyers, the decision sits further upstream. Ask infrastructure providers which parts of their performance claims come from GPUs, networking, software, and workload-specific optimization.

NVIDIA has started the production test. Now customers, system partners, and competing suppliers must show whether co-packaged optics can survive the less controlled test of real data centers.