Broadcom Optical Standards Are Moving AI Scale-Up From Copper to Fiber

Broadcom optical standards moved into focus when six major AI infrastructure companies formed a group targeting links between accelerators. The March 2026 initiative includes AMD, Broadcom, Meta, Microsoft, Nvidia, and OpenAI. It seeks an open specification for optical connections inside increasingly large AI systems.

That combination makes the announcement more consequential than another networking consortium. Broadcom supplies switch silicon, optical components, signal processors, packaging expertise, and custom accelerator technology. It can influence the specification while selling several product layers needed to implement it.

The conflict is equally important. The consortium promises a multi-vendor market, yet the earliest deployments favor companies that already control tightly integrated technology stacks. Nvidia has its own networking architecture, while Broadcom has spent several product generations combining Ethernet switches with co-packaged optics.

The central question is therefore not whether fiber will enter more AI systems. Copper faces growing reach, bandwidth, and power constraints as accelerator clusters expand. The question is whether open interfaces broaden the supplier market or reinforce the strongest integrated vendors.

Broadcom sits at the center because it has both standards influence and shipping experience. However, specifications alone cannot settle manufacturing yields, thermal management, repair procedures, or the economics of replacing pluggable optical modules.

Broadcom Optical Standards Target the Scale-Up Bottleneck

The new specification moves optical networking closer to the processors that train and run large AI models.

The Optical Compute Interconnect Multi-Source Agreement was announced on March 12, 2026. Its founding members represent accelerator designers, cloud operators, networking suppliers, and major buyers of AI infrastructure.

The group is developing an open optical specification for scale-up connections. Scale-up networking links accelerators within a tightly coordinated computing domain, where latency and predictable performance directly affect job completion.

Scale-out networks serve a related but different purpose. They connect servers or accelerator groups across a larger cluster. Ethernet and optical transceivers already play established roles in that layer.

Scale-up systems have traditionally relied heavily on short electrical connections. Copper works well over limited distances, but higher signaling rates reduce its practical reach. Signal losses also require more compensation and consume additional power.

The consortium’s published roadmap targets 3.2 terabits per second per fiber and beyond. It also covers pluggable, on-board, and co-packaged optical implementations. That flexibility matters because operators are unlikely to adopt one physical design everywhere.

Its specification is intended to remain independent of a single processor protocol. An optical connection could therefore transport different accelerator fabrics without binding every supplier to one proprietary electrical architecture.

The founding companies say the OCI specification will support multiple wavelengths and a multi-vendor supply chain. Those goals still require completed technical definitions and proven interoperability.

Broadcom’s position goes beyond its name on the membership list. The company already sells the SerDes circuits that move data into and out of chips. SerDes converts parallel data into high-speed serial signals and reverses that process at the destination.

It also supplies optical digital signal processors, lasers, photodetectors, switches, and custom accelerator components. Few vendors participate in so many adjacent layers of the proposed connection.

This breadth gives Broadcom useful feedback loops. Engineers working on switch interfaces can coordinate with optical engine teams and packaging specialists. Customers can evaluate a more complete system instead of assembling every component independently.

It also creates a concentration question. An open interface can permit alternative suppliers, but early adopters still need products that operate reliably at production scale. The vendor with the most complete implementation can capture significant value before competition matures.

That is the first source of tension. The specification is designed to loosen proprietary constraints, while the implementation rewards companies with unusually deep integration.

AI Clusters Are Running Out of Electrical Headroom

AI infrastructure needs optical scale-up because bandwidth growth is colliding with reach, power, and connector limits.

Training workloads divide calculations across thousands of accelerators. Those processors repeatedly exchange parameters, activations, and intermediate results. A slow or unstable connection can leave expensive compute resources waiting.

Increasing cluster size therefore raises more than aggregate bandwidth demand. It also increases the number of links that must deliver consistent latency and low error rates. Small inefficiencies multiply across the fabric.

Copper remains attractive because it is familiar, inexpensive, and easy to service at short distances. Yet electrical signals weaken as they travel through circuit boards, connectors, and cables. The problem becomes harder at higher lane speeds.

Engineers can add retimers or stronger equalization to recover a degraded signal. A retimer reconstructs timing and data before forwarding the signal. That approach extends reach but adds power, cost, and design complexity.

Fiber transports data with lower loss over longer distances. It can also provide greater bandwidth density around a processor or switch. However, the electrical signal must still be converted into light and back again.

Traditional pluggable transceivers place that conversion in removable modules at a system’s front panel. The layout supports replacement and supplier choice. It also leaves a longer electrical path between the switch chip and the optical module.

Co-packaged optics, commonly shortened to CPO, moves optical engines beside the networking chip on a shared substrate. The shorter electrical path reduces loss before conversion. It can therefore reduce the compensation required at high speeds.

Broadcom says its CPO architecture offers more than threefold power savings and lower optical cost per bit. Those figures are company claims, not universal results across every system configuration. Broadcom explains the claimed mechanism in its CPO overview.

The timing also reflects a change in accelerator topology. AI operators want to connect more processors across racks while preserving scale-up behavior. Electrical cables become less practical as the physical domain expands.

Optics can let operators separate compute trays, memory, and switching components without accepting the same distance constraints. It can also help designers reorganize cooling and power distribution within dense racks.

That opportunity pressures several groups at once. Cloud operators must decide when the energy savings justify a less familiar service model. Equipment makers must redesign systems around new optical and packaging requirements.

Optical component suppliers must support higher lane rates while improving production yields. Switch vendors must provide interfaces that work across multiple optical implementations. Accelerator designers must avoid introducing latency that weakens tightly synchronized workloads.

The pressure extends to Nvidia and Broadcom because both want to define the fabric surrounding AI processors. Nvidia can combine GPUs, switches, networking adapters, and software. Broadcom offers Ethernet silicon, custom accelerators, optics, and extensive hyperscaler relationships.

Open optical standards can reduce dependence on a single proprietary fabric. They can also give hyperscalers more control over sourcing. Yet openness does not remove the engineering advantage held by vendors with mature system designs.

That explains why Broadcom’s role attracts investor attention. Optical adoption could expand the company’s addressable content within each AI cluster. Its standards work also helps align that portfolio with customer roadmaps.

The opportunity remains operational rather than automatic. Every additional integrated component introduces qualification, packaging, and supply requirements. A convincing standard must work outside one vendor’s controlled demonstration.

Broadcom’s Advantage Is the Stack, Not the Specification

Broadcom leads because its standards participation connects directly to products spanning switches, signaling, optics, and packaging.

At the 2026 Optical Fiber Communications Conference, Broadcom presented a portfolio built around multiple stages of the data path. Its announcements included a 102.4-terabit Ethernet switch, 400-gigabit-per-lane optical DSP technology, and 200-gigabit Ethernet retimers.

The company also showed near-packaged optics, or NPO. That architecture places optical components closer to the switching chip without integrating them as tightly as CPO. It offers another point between conventional pluggables and full co-packaging.

Broadcom’s Taurus optical DSP operates at a claimed 400 gigabits per lane. The company pairs it with electro-absorption modulated lasers and photodiodes. These components convert electrical data into optical signals and detect incoming light.

Broadcom says that combination supports 1.6-terabit transceivers and prepares suppliers for future 3.2-terabit modules. Its OFC portfolio also covers PCI Express switching and active electrical cables.

This range matters because no single optical component creates a usable AI fabric. The host chip, electrical interface, optical engine, laser, connector, fiber, management software, and cooling system must operate together.

A supplier controlling several layers can tune boundaries that otherwise require negotiation between vendors. That can accelerate early deployments and simplify accountability when a link fails.

The same integration can influence a standard’s direction. Broadcom knows which requirements its manufacturing process can meet and which interfaces customers actually request. It can bring those constraints into consortium discussions.

However, Broadcom does not control the OCI effort alone. Nvidia, AMD, Meta, Microsoft, and OpenAI have different incentives. The hyperscalers want supplier flexibility, while chip vendors want differentiation around their platforms.

Nvidia represents the most important counterweight. It already combines accelerated computing with networking products and a mature software environment. Its ability to coordinate hardware and software makes a proprietary system attractive to customers seeking predictable deployment.

Broadcom’s answer is an open Ethernet-centered portfolio. That approach lets cloud companies mix accelerator designs, switching equipment, and optical suppliers. It aligns with buyers that design custom infrastructure rather than purchasing one complete stack.

The primary contest is therefore integrated openness against proprietary integration. Broadcom wants standard interfaces without giving up the benefits of a coordinated component portfolio. Nvidia can support standards while preserving distinctive system capabilities above them.

This is not a simple Broadcom-versus-Nvidia product comparison. Both companies participate in the OCI group. Both also benefit if optical links expand the size and value of AI systems.

The disagreement concerns where differentiation remains after the optical layer becomes interoperable. If suppliers standardize only the physical transport, vendors can still compete through scheduling, congestion control, topology, and management.

If the standard reaches deeper into system behavior, customers gain more freedom to combine hardware. That outcome can weaken proprietary lock-in but increase the testing burden across mixed environments.

Broadcom is well placed in either case. A narrow specification favors its integrated product portfolio. A broader specification enlarges the market for its switch, SerDes, DSP, and optical components.

That strategic flexibility helps explain its central position. Broadcom does not need every customer to choose identical optical packaging. It can sell into pluggable, near-packaged, and co-packaged designs.

The company’s advantage is still conditional. Customers must believe the components can be sourced and serviced at scale. Standards influence cannot compensate for insufficient optical engines or constrained packaging capacity.

Open Optics Still Has a Proprietary Middle

A shared optical interface does not make complete systems interchangeable, especially when packaging and rack designs remain customer-specific.

Multi-source agreements define enough common behavior for several companies to build compatible products. They do not necessarily standardize every internal component, mechanical arrangement, or management feature.

This distinction is especially important for CPO. Moving optics beside a switch ASIC ties optical engines to the package, thermal design, power delivery, and fiber-routing plan. Those elements vary across equipment makers.

An operator might buy standards-compliant optical components yet remain dependent on one switch package. Another system could use the same logical specification with a different connector or cooling arrangement.

Even the consortium’s broad form-factor support reveals this challenge. Pluggable, on-board, near-packaged, and co-packaged optics involve different repair procedures. They also distribute heat and signal processing differently.

Traditional modules can be removed from a front panel while the rest of the system remains intact. CPO places more optical hardware near valuable silicon. A failure can therefore affect a larger assembly.

External laser sources can improve serviceability by keeping lasers replaceable. They also add connectors, fiber routing, and control requirements. Every architectural choice moves risk rather than eliminating it.

Independent industry work highlights this tradeoff. An IEEE comparison describes LPO as easier to replace and less tightly integrated. It also notes CPO’s data-rate and efficiency advantages.

Linear pluggable optics removes the module’s full digital signal processor while retaining a removable front-panel format. It can reduce power and latency, but the analog link demands careful host and module coordination.

Linear retimed optics restores some signal processing on the transmit side. That choice uses more power than pure LPO but can ease integration. These alternatives show that the market has not selected one universal architecture.

OIF’s work adds another layer of standardization. At OFC 2026, the organization scheduled interoperability demonstrations involving 40 participating companies. Its program covered 224-gigabit and 448-gigabit electrical interfaces, co-packaging, management, and energy-efficient designs.

Those interoperability trials matter because paper compliance does not guarantee stable operation. Vendors must test signal margins, firmware behavior, management interfaces, and fault recovery across real equipment.

Broadcom benefits from that process because it already supplies widely used host silicon. Module makers have a commercial reason to validate their products against Broadcom switches. That installed base can make compatibility self-reinforcing.

Yet the standards market can also create competitors. Once interfaces stabilize, optical specialists can target defined boundaries without building an entire network stack. Equipment makers can qualify second sources for lasers, engines, and modules.

The result will probably remain mixed. Open specifications can create interchangeable optical layers while packaging stays customized. Operators can gain some sourcing flexibility without obtaining complete system portability.

Readers should therefore treat “open” as a technical scope, not a guarantee. The important questions concern which interfaces are standardized and which qualification data are shared.

A credible open environment also needs failure reporting and management consistency. Operators must identify a failing component without assigning blame across several vendors. Otherwise, integration costs can erase procurement savings.

Broadcom’s central role carries responsibility here. Its components will often sit on one side of an interoperability test. The company must show that openness extends beyond partners using Broadcom’s preferred implementation.

That test will determine whether Broadcom optical standards enlarge a competitive market or merely define entry points into Broadcom-centered systems.

Manufacturing Will Decide Whether CPO Escapes Limited Deployment

The largest near-term risk is not bandwidth theory; it is producing complex optical packages with acceptable yield and reliability.

A standards document can align electrical and optical behavior. It cannot instantly expand silicon photonics fabrication, packaging tools, or experienced engineering teams. Those resources require long investment cycles.

CPO optical engines combine modulators, photodetectors, waveguides, couplers, and electronic control. Small defects can reduce the yield of an expensive assembly. Thermal changes can also alter optical performance.

These difficulties become more serious beside a high-power switch ASIC. The package must manage heat without disturbing alignment or shortening component life. It must also preserve signal quality across manufacturing variation.

TrendForce reported in July 2026 that Nvidia had begun shipping Spectrum-X CPO switches to selected partners. It described Broadcom’s 51.2-terabit Bailly CPO switch as entering volume manufacturing with Delta Electronics and Micas Networks.

However, the same production analysis identified optical engines, silicon photonics, and advanced packaging as expansion bottlenecks. That warning qualifies any claim of immediate mass adoption.

TrendForce said Broadcom’s system had completed deployment validation with Meta. It also reported a claimed power reduction of up to 70 percent against conventional pluggable transceivers.

That figure should not be treated as a universal data-center saving. Optical interconnect power represents only one part of a switch or cluster. Results depend on traffic, reach, cooling, topology, and comparison methodology.

Limited shipments still represent meaningful progress. CPO spent years in presentations and prototypes while operators questioned repairability. Production systems create the field data needed to evaluate those concerns.

Broadcom also has a generational advantage. It has developed multiple CPO switch designs rather than making a single experimental product. Repeated designs let engineers improve fiber attachment, external lasers, packaging, and diagnostics.

But production maturity must be measured through sustained volumes and failure rates. Announcing a shipping product does not reveal yield, customer concentration, or qualification duration.

Supply constraints could also shape competition unexpectedly. A strong specification can increase demand faster than component suppliers can respond. Vendors with reserved packaging capacity would then hold leverage regardless of interface openness.

Nvidia’s close work with TSMC provides one model. Broadcom uses its own supplier relationships and partner network. Large hyperscalers may reserve capacity or support optical startups to reduce dependence on either path.

Advanced packaging adds another conflict. CPO systems compete for some manufacturing resources also needed by AI accelerators and high-bandwidth memory assemblies. Those products can carry different margins and strategic priorities.

Reliability remains equally important. A pluggable module failure can often be addressed by replacing one accessible unit. A co-packaged failure can require more involved service, even when lasers remain external.

Operators will need data covering link flaps, thermal cycles, connector contamination, and field replacement. A link flap is a brief connection loss that can interrupt synchronized computing work.

AI training makes such interruptions expensive. One unreliable link can delay many accelerators, depending on the workload and recovery process. Buyers will prioritize predictable operation over a favorable laboratory efficiency number.

This risk does not invalidate CPO. It explains why adoption will proceed through selected systems and demanding customers. Hyperscalers can absorb integration work that ordinary enterprise buyers cannot.

Broadcom is positioned for that initial market because it already works closely with large infrastructure operators. The harder test comes when equipment must support broader deployment and more varied operating environments.

A truly open optical market needs repeatable manufacturing across several suppliers. It also needs components that can be qualified without rebuilding each system around them.

Until those conditions emerge, Broadcom’s integration advantage remains stronger than the standard’s promise of interchangeability.

Three Signals Will Show Whether Broadcom’s Lead Lasts

Specification detail, production evidence, and competitive interoperability will reveal whether Broadcom has established a durable position.

The first signal is a usable OCI specification accompanied by implementation commitments. Membership announcements establish intent, but engineers need defined electrical, optical, mechanical, and management boundaries.

The most informative release will explain what vendors can change without breaking compatibility. It should also identify supported wavelengths, link distances, error handling, and qualification expectations.

Broad participation would strengthen the open-market thesis. Implementations from multiple optical engine and switch suppliers would show that the standard supports more than founding-member roadmaps.

A specification shaped mainly around one available platform would weaken that argument. It could still accelerate deployment, but customers would gain less practical supplier freedom.

The second signal is Broadcom’s production and reliability evidence. Shipment labels matter less than sustained volume, repeat customers, and disclosed operating experience.

Watch for information about optical engine yield, field failures, system availability, and partner capacity. Customer deployments across more than one hyperscaler would carry particular weight.

Broadcom’s product claims deserve reporting, but third-party validation deserves greater weight. Independent testing should compare complete systems under similar traffic, distance, cooling, and redundancy conditions.

Evidence of stable production would strengthen Broadcom’s lead because manufacturing is harder to copy than a standards vote. Persistent packaging constraints would delay revenue and preserve demand for pluggable alternatives.

The third signal is a competitive response from Nvidia, Marvell, Arista, optical specialists, and system vendors. The most revealing products will combine OCI compatibility with different physical implementations.

A competing CPO switch would test whether suppliers can match Broadcom’s package-level integration. A successful NPO or LPO platform could instead show that customers prefer serviceability over maximum density.

The outcome does not require one architecture to eliminate the others. Scale-up links, scale-out fabrics, and longer data-center connections have different constraints. Multiple optical designs can coexist within one facility.

Broadcom will remain central if its components appear across those designs. Its switch silicon, DSPs, SerDes, lasers, and custom accelerators give it several ways to participate.

Its position becomes less secure if standards let customers replace entire Broadcom layers without significant integration work. Strong competition could compress component value even as the optical market expands.

Investors should also separate infrastructure adoption from near-term financial expectations. Technical leadership does not disclose customer timing, contract structure, margins, or production costs.

Enterprise buyers have a different concern. Most will not purchase CPO switches directly in the first wave. They will still feel the effects through cloud capacity, service pricing, and the availability of larger AI systems.

Developers should care because network behavior affects model training efficiency and distributed inference. Better bandwidth does not automatically improve every application, but communication-heavy workloads benefit most.

Cloud architects should examine failure domains and observability. An optical fabric needs clear telemetry, predictable recovery, and service procedures. Peak throughput cannot compensate for uncertain operations.

The deeper significance of Broadcom optical standards lies in where system boundaries are being redrawn. Optics is moving from a removable network accessory toward part of the compute package.

Broadcom enters that transition with products, customers, and manufacturing experience. The OCI consortium adds an open framework around those assets, while also inviting competitors into defined parts of the stack.

That combination makes Broadcom central without making it unassailable. Its lead rests on delivering interoperable systems at scale, not merely helping write the specification.

Over the next several months, follow the specification’s technical scope, independent deployment evidence, and genuinely mixed-vendor systems. Those signals will show whether openness expands choice or consolidates influence.

If you operate AI infrastructure, ask suppliers which interfaces are actually interchangeable and how failures are isolated. Request system-level power measurements rather than component estimates. Examine replacement procedures before accepting density claims. Developers should track whether larger optical scale-up domains improve real workload completion times. Investors should separate consortium membership from qualified production revenue. Broadcom currently connects more pieces of the optical stack than most rivals, but the market is still testing those pieces together. The defining question for Broadcom optical standards is practical: can another vendor enter the system without forcing customers to redesign it?