SK hynix silicon photonics: Why more GPUs still need better connections

SK hynix published a technical note on October 29 explaining why AI can still wait despite having tens of thousands of GPUs. The SK hynix silicon photonics discussion focuses on a growing infrastructure problem: adding computing capacity does not automatically improve the connections between processors. GPUs, the processors used for many AI workloads, need data to arrive before they can perform useful work.

The proposed response combines silicon photonics, which integrates optical functions on silicon chips, with co-packaged optics, or CPO, which places optical engines beside processing chips. The central contest involves conventional pluggable optical modules and tighter optical integration. Both use light, but they convert electrical signals into optical signals at different physical locations.

That distinction now has commercial context. Broadcom highlighted another generation of CPO systems this month, while Marvell included a CPO platform in its September demonstration lineup. The question facing infrastructure buyers is therefore specific: when does moving optics closer to chips justify changing how networks are built, qualified, and repaired?

The SK hynix silicon photonics note lands amid a broader packaging shift

The immediate event is an explanation of an infrastructure constraint, rather than evidence that SK hynix has launched a CPO networking product.

The silicon photonics note appears in SK hynix’s newsroom with an October 29, 2026 publication date. Its title connects waiting GPUs with optical networking and closer integration. The complete page was inaccessible during verification, so detailed technical claims below rely on separately accessible primary sources.

That distinction matters because a technical explainer and a product announcement establish different things. An explainer identifies a problem and describes an engineering direction. A product announcement can establish specifications and intended availability, while operational evidence addresses whether customers obtain the promised benefits.

Readers should therefore treat the SK hynix silicon photonics framing as an entry point into a documented industry transition. It does not establish that every AI cluster suffers the same bottleneck. It also does not establish how much any particular installation would gain from replacing its optical equipment.

The technical tension starts with a familiar limitation of distributed computing. A workload spread across processors must move information between them. If an operation needs information that has not arrived, the available arithmetic capacity cannot resolve that dependency by itself.

This creates an uncomfortable purchasing problem. A buyer can install faster processors while leaving a communication constraint largely untouched. The resulting system has greater theoretical computing capacity, yet the application still encounters delays at the same coordination points.

Optical networking already addresses distance and bandwidth in data centers. The proposed change is therefore more precise than replacing copper cables with fiber everywhere. CPO moves the electrical-to-optical boundary closer to the chip, reducing the electrical distance that high-speed signals must traverse inside equipment.

Conventional pluggable optics place removable conversion modules at a system’s front panel. That arrangement supports accessible replacement and familiar operational procedures. However, the signal must still travel electrically between the main chip and the module.

CPO changes that physical arrangement. Optical engines move into the chip’s package, the assembly connecting semiconductor components to the surrounding system. Fiber then carries signals away from that integrated location.

This is why packaging has become part of the networking discussion. The location of a conversion component affects electrical losses, power requirements, assembly methods, and maintenance choices. A different package can change the practical limits of a network design without changing the AI model running above it.

The news value lies in that shift of attention. Processor specifications remain important, but the equipment connecting processors increasingly determines whether an installation can use its purchased capacity effectively. The relevant unit of evaluation becomes the working system.

Waiting processors put pressure on infrastructure buyers

The strongest case for CPO begins with a measured communication bottleneck, rather than a headline GPU count.

Consider a distributed training job, which divides model development across cooperating processors. During training, processors calculate updates and exchange information needed to keep their work coordinated. The useful result depends on both computation and communication completing when required.

This is a concrete setting where connection quality matters. A faster local calculation provides limited relief when the next operation depends on information from elsewhere. Adding processors can also change communication demands, so larger deployments require attention to how work and data move together.

That observation does not mean networking always determines performance. Storage access, software scheduling, memory capacity, and workload design can also limit progress. An optical upgrade addresses a particular part of the system, so buyers need measurements that identify the limiting dependency.

The first group under pressure consists of cloud operators and organizations building their own AI infrastructure. They must evaluate networking alongside processors, cooling, and power delivery. A configuration that looks attractive through component specifications can disappoint when the intended application runs across the complete installation.

The second group consists of system suppliers. They must translate optical integration into equipment customers can install and maintain. This requires coordination across chip design, optical components, assembly, cooling, diagnostics, and replacement procedures.

Enterprise buyers face a related challenge even when they rent infrastructure. They purchase access to useful computing work, rather than direct ownership of every underlying component. Their practical questions concern completion times, availability, and performance under the workloads they actually intend to run.

For these buyers, an optical architecture is a means to an operational outcome. A provider’s CPO announcement becomes relevant when it supports a measurable improvement in that outcome. The presence of integrated optics alone does not establish a better customer experience.

Workload comparisons also need consistent boundaries. A demonstration using a communication-heavy application cannot establish equal gains for a workload dominated by local calculation. Buyers should ask which operations improved and how much those operations contributed to the complete task.

The pressure reaches power planning as well. Energy used to transport data competes with other demands inside an installation’s operating envelope. Reducing communication overhead can create useful design flexibility, but operators must determine where the savings occur.

An optical component’s efficiency does not describe the electricity consumption of an entire facility. Cooling equipment, processors, memory, and other network elements remain part of that larger total. A percentage attached to one subsystem needs a clearly stated denominator.

The SK hynix silicon photonics discussion is useful here because it challenges a purchasing shortcut. Counting processors provides a visible measure of scale, but it says little about coordination efficiency. Buyers need evidence connecting hardware choices to completed work.

For developers, the implication is equally practical. Performance investigations should distinguish time spent calculating from time spent waiting for required information. That distinction helps determine whether a networking change addresses the observed problem.

Knowledge workers using hosted AI services sit farther from these engineering decisions. Optical integration will not automatically make an answer more accurate or improve access to relevant documents. Its potential contribution concerns the infrastructure supporting delivery, subject to the provider’s actual implementation.

CPO changes where electrical signals become light

The mechanism behind CPO is physical proximity: shorten a demanding electrical path before handing the signal to an optical connection.

A high-speed electrical signal degrades as it passes through conductors, connectors, and other interfaces. Equipment can compensate for degradation, but that compensation requires circuitry and energy. Moving the optical conversion point closer to the source reduces the electrical path that needs this treatment.

NVIDIA describes this approach in its co-packaged photonics design. Its March 2025 technical explanation contrasts conventional electrical paths measuring 14 to 16 inches with paths shorter than half an inch. Those dimensions describe NVIDIA’s comparison, rather than a universal specification for every CPO implementation.

The company also describes eliminating an external digital signal processor, a chip that processes signals to compensate for transmission impairments. Removing that stage changes component requirements and the signal path. The resulting application benefit still depends on the wider network and workload.

This explains why “highways of light” is an incomplete purchasing description. Light already carries data through conventional optical links. The architectural change concerns how much electrical transport and signal conditioning remain before the information enters those links.

Silicon photonics and co-packaged optics also describe different aspects of a design. Silicon photonics concerns the platform used to implement optical functions. Co-packaging concerns the placement of optical engines relative to other chips.

A silicon photonics component can therefore appear in a conventional module. Its presence does not automatically establish CPO. Buyers comparing systems should identify the physical architecture rather than treating these terms as interchangeable product labels.

The conventional arrangement retains a clear operational attraction. A technician can access a front-panel optical module and replace it using established procedures. That modularity provides flexibility when operators qualify suppliers, stock replacements, and investigate individual links.

CPO moves more integration responsibility toward the system supplier. The supplier must deliver a working combination of electrical interfaces, optical engines, fiber attachment, and thermal design. Customers then need to understand which parts remain independently replaceable.

Historical product announcements show that this transition predates the latest AI infrastructure headlines. Broadcom announced delivery of its Bailly CPO switch in March 2024. Its 51.2-Tbps CPO platform combined a switching chip with eight optical engines, each rated at 6.4 terabits per second.

Broadcom claimed 70% lower optical interconnect power consumption compared with pluggable transceiver solutions. That claim concerns the optical interconnect, rather than total data-center electricity consumption. It also represents a supplier comparison, rather than an independently established result for every customer deployment.

The distinction between switching and computing deserves equal care. A network switch directs data among connected devices. Integrating optics with that switch does not establish that the AI processors themselves use the same optical packaging.

This leaves several possible deployment boundaries. An operator can adopt optical integration in network equipment while retaining other connection technologies elsewhere. The appropriate boundary depends on distance, communication requirements, physical constraints, and operational priorities.

The central trade is therefore concrete. Conventional modules preserve a familiar separation between switching hardware and removable optics. CPO reduces that separation to improve integration, while requiring the complete assembly to meet customer expectations.

That is the mechanism worth evaluating. Buyers do not need to assume that every connection will become optical. They need to identify where a shorter electrical path produces enough benefit to justify a different equipment design.

Broadcom and Marvell show why the timing matters

Recent supplier activity places CPO inside current infrastructure planning, although demonstrations and announcements do not establish widespread deployment.

Broadcom’s October 8 announcement for the 2026 Open Compute Project Global Summit included its third-generation TH6-Davisson CPO technology. The Open Compute Project is an industry organization developing shared infrastructure designs. Broadcom’s announcement placed integrated optics alongside switches, server connectivity components, and rack systems.

Its planned optical networking demonstrations included Tomahawk 6-Davisson systems from Celestica, Delta Electronics, and Micas Networks. These named system partners matter because buyers require complete equipment rather than isolated semiconductor specifications. However, the announcement described planned displays and should not substitute for shipment or installation records.

Marvell provided another recent reference point on September 20. Its announcement for the European Conference on Optical Communication included a 102.4-terabit-per-second CPO platform among its demonstrations. The conference took place in Málaga, Spain, from September 20 through September 24.

The CPO demonstration lineup also included advances in other optical technologies. That breadth matters because integrated optics compete for adoption alongside improvements to conventional optical equipment. Operators are evaluating moving targets on both sides of the architectural choice.

These announcements establish activity across multiple suppliers. They do not provide a controlled comparison between Broadcom, Marvell, and NVIDIA systems. Matching a headline bandwidth figure would still leave questions about configuration, power boundaries, cooling, software, and serviceability.

Nor should the announcements be treated as evidence that conventional modules have stopped improving. A buyer considering CPO must compare it with the conventional equipment available for the same installation. Comparing a new integrated design with an outdated alternative can exaggerate the practical advantage.

The current moment therefore has a specific character. Optical integration has moved beyond a single supplier’s proposal, and system partners are presenting equipment around it. The remaining evaluation concerns qualification, repeatability, operational support, and workload results.

That creates pressure on vendors to make their comparisons auditable. A useful disclosure identifies the equipment configuration, the comparison system, the measurement boundary, and the workload. It also explains whether the result came from a demonstration, a qualification environment, or customer operations.

The same discipline applies to deployment language. “Announced,” “demonstrated,” “available,” and “deployed” describe different milestones. Customers should expect suppliers to distinguish them clearly when presenting adoption claims.

For infrastructure planners, the sensible response is to keep both architectures under active evaluation. An existing installation may favor conventional modules because its operational requirements are already well understood. A new design may justify closer optical integration if measured constraints support the change.

This is also why the story extends beyond a component race. Decisions made during equipment selection can affect maintenance training, replacement inventory, qualification schedules, and supplier relationships. Those consequences remain relevant even when the electrical and optical specifications look compelling.

The competitive comparison should therefore remain neutral. Broadcom’s current partner activity, Marvell’s recent demonstrations, and NVIDIA’s published architecture document different parts of the transition. None alone establishes which design best serves a particular operator.

Reliability evidence needs the same scrutiny as efficiency claims

A lower-power connection earns its place only when the complete system meets reliability and maintenance requirements.

Optical integration changes how components share space and operating conditions. Buyers must assess temperature behavior, fiber handling, assembly consistency, and diagnostics. These requirements do not disappear because the electrical path becomes shorter.

Serviceability is especially important. The relevant question is which component fails and what must happen to restore service. Replacing an accessible laser source presents a different maintenance task from addressing an inaccessible optical engine or package defect.

An external laser source supplies light from outside the integrated optical assembly. NVIDIA’s technical explanation describes front-panel laser modules that technicians can replace separately. This addresses a specific maintenance concern, while leaving buyers responsible for evaluating the complete recovery procedure.

Reliability claims now include some published testing evidence. In October 2025, Broadcom reported Meta testing totaling one million cumulative 400G-equivalent port device hours without a link flap. A link flap is a brief loss and restoration of connectivity.

Broadcom’s CPO reliability results described high-temperature laboratory characterization and reported 65% lower optics power than pluggable module solutions. These are supplier-published results concerning a defined evaluation. They should not be generalized into a universal field failure rate.

The cumulative-hours wording deserves attention. Hours aggregated across ports do not describe a single system operating continuously for that entire period. Readers also need the equipment population, operating conditions, and failure definitions before comparing reliability claims.

Similarly, an absence of link flaps addresses one observable behavior. It does not automatically establish the absence of every hardware fault, software interruption, or maintenance event. A broader operational assessment needs evidence covering the service as customers experience it.

These qualifications do not dismiss the testing. Defined reliability observations are more useful than unsupported assurances. The next question is whether comparable behavior persists across deployment conditions, equipment populations, and operational procedures.

Manufacturing consistency presents a separate issue. A working demonstration establishes that an assembly can function. Buyers evaluating scale need confidence that delivered systems meet the same specifications consistently and receive adequate support.

Thermal qualification also needs realistic boundaries. A component’s behavior depends partly on surrounding equipment and cooling arrangements. Operators should evaluate the supplied system under conditions matching the intended installation, rather than assuming laboratory results transfer unchanged.

The skeptical position is therefore specific rather than ideological. CPO addresses a real engineering constraint, but integration moves some complexity into packaging and system qualification. Whether that exchange benefits a buyer depends on evidence.

The SK hynix silicon photonics framing should be read with that boundary intact. Explaining why processors wait does not establish that a particular optical architecture resolves every cause. The appropriate conclusion follows the measured bottleneck and the tested implementation.

That approach also prevents unsupported promises about AI services. Better networking can address communication overhead where it exists. It does not by itself establish better reasoning, more reliable factual answers, or improved handling of a user’s private information.

Three signals to watch through January 2027

The next useful evidence will connect supplier announcements with delivered equipment, application results, and maintainable operations.

First, watch for delivery and qualification disclosures covering the systems highlighted at recent industry events. Named customers, identifiable configurations, and clear acceptance milestones would strengthen the case that CPO is advancing into routine infrastructure choices. Additional demonstrations alone would provide weaker evidence of adoption.

Broadcom’s named system partners create concrete places to look for that progress. Their disclosures should distinguish an exhibited system from equipment available for ordering or operating in customer environments. A clearer progression would make the commercial picture easier to assess.

Second, watch for workload measurements that connect optical integration to useful application performance. The strongest comparisons will explain the model or task, system configuration, networking arrangement, and power boundary. They should also identify which hardware or software variables changed.

A result showing improved completion time alongside measured communication savings would support the central thesis. A result limited to an optical component would establish a narrower benefit. Both can be informative, provided the reported conclusion matches the measurement.

Buyers should also watch for workloads that show little improvement. Such results help define where CPO adds value and where other constraints dominate. A technology becomes easier to deploy responsibly when its limits are documented alongside its advantages.

Third, watch for service and reliability disclosures from operational settings. Useful evidence includes replacement procedures, diagnostic coverage, recovery behavior, and clearly defined failure observations. These details would strengthen the argument that closer integration can satisfy the practical requirements of an operating network.

Maintenance limitations would not erase the electrical advantages. They would change where the architecture makes sense and how an operator should support it. The outcome depends on the complete operating model rather than a single favorable measurement.

Together, these signals provide a disciplined way to follow the story without assuming a predetermined winner. Delivery establishes access to equipment. Application measurements establish usefulness, while service evidence establishes whether operators can sustain that usefulness over time.

Teams tracking these developments should preserve the original claims alongside later results. A searchable technical knowledge base can help reviewers compare specifications, qualification findings, and maintenance guidance. Keeping those materials connected makes it easier to identify when a supplier’s evidence has materially changed.

For your next infrastructure review, ask the provider to connect its optical architecture with your actual workload. Start with measured waiting time, then examine the proposed remedy and the maintenance plan. Keep component efficiency separate from system efficiency, and request the configuration behind every comparison. The SK hynix silicon photonics discussion offers a useful starting question: what prevents the installed processors from doing useful work? Follow that question through the network, the software, and the operational evidence before treating another hardware specification as the answer. Which upcoming disclosure would change your decision: a qualified shipment, a reproducible workload result, or a documented recovery procedure?