That distinction is important as AI infrastructure scales. The network between accelerators must carry growing amounts of traffic while staying within practical power, cooling, rack-density, and maintenance limits. A high aggregate bandwidth figure is meaningful engineering progress, but data-center operators ultimately need hardware they can source, cool, monitor, replace, and support over a multiyear lifecycle.
What Huawei showed
Huawei demonstrated a 7.2Tb/s near-packaged-optics, or NPO, module at the September 9–11 CIOE event. Reporting on the presentation described the design as using 36 lanes at 200Gb/s per lane. Huawei positioned the module for next-generation SuperPOD systems, its branding for large-scale AI infrastructure.
NPO places an optical engine close to the switch ASIC, the chip that directs traffic through a network switch, while retaining the optical engine as a separate package. That is distinct from co-packaged optics, or CPO, in which the optical engine is placed on the same substrate or package as the switch ASIC. Traditional pluggable optics place the transceiver farther from the ASIC in a removable module.
The underlying objective is to shorten the electrical path before data is converted into light. As signaling speeds rise, longer electrical paths can impose greater signal-integrity, power, and circuit-design demands. Bringing optics closer to the switching silicon is one potential way to reduce those pressures.
Huawei attributed several benefits to its design: eliminating a separate module digital signal processor, reducing latency from 100ns to 10ns, and cutting power consumption by approximately 60%. Those figures could be highly consequential if they held at system scale across large numbers of ports.
They remain vendor-reported claims, rather than independently validated deployment results. The reviewed material does not provide a disclosed comparison baseline, test configuration, workload, thermal conditions, error-rate results, or independent laboratory measurements. It also does not establish whether the figures apply to an entire switch, a module, or a narrower component-level comparison. Buyers should therefore regard them as Huawei’s stated design results, not as settled evidence of fleet-level performance.
Why NPO is attracting attention
CPO has long been viewed as a way to improve bandwidth density and electrical efficiency by putting optics extremely close to the switch ASIC. The trade-off is operational complexity. When optics are more deeply integrated with switching silicon, manufacturing, servicing, and component replacement can become more difficult than with conventional removable transceivers.
NPO is meant to occupy a middle position. It pursues a short electrical path while leaving the optical element separately packaged. Huawei has argued that this makes NPO more mature and easier to maintain than a fully co-packaged approach.
That is a reasonable architectural proposition, but it is not yet a universal operational conclusion. The reviewed record does not supply comparable fleet data for NPO and CPO on failure rates, manufacturing yield, mean repair time, spare-part handling, lifecycle cost, or long-term reliability. Those operational measures will matter as much as headline throughput when an operator decides which optical architecture to deploy.
The more accurate reading is that NPO, CPO, and advanced pluggable designs are competing responses to the same problem: increasingly fast electrical signaling between switch silicon and optics. None has yet been shown in the reviewed evidence to be the single answer for every AI network.
OIF work raises a compatibility question
The Optical Internetworking Forum’s NPO work is important because common interfaces can determine whether a technology becomes a broad ecosystem or remains tied to a particular supplier and system design. Official OIF material from its second-quarter 2026 work confirms that the project includes both 6.4Tb/s and 12.8Tb/s NPO module interfaces.
Huawei’s 7.2Tb/s demonstration does not match either of those aggregate capacities. That mismatch does not make the Huawei design technically invalid or commercially irrelevant. Vendors can demonstrate hardware before a related interface effort is completed, and standards work may accommodate more than one architectural choice.
But interoperability cannot be presumed from aggregate capacity alone. Compatibility depends on much more than the total bit rate: electrical signaling and channel definitions, mechanics, power delivery, cooling, laser arrangements, connectors, management behavior, diagnostics, and service procedures can all affect whether components work together.
The reviewed evidence does not establish how Huawei’s module maps to the eventual OIF interface work. Nor does it show that the demonstrated module will interoperate with equipment designed around either of the OIF project’s stated capacity classes. For tightly controlled, single-vendor environments, a specialized implementation may still be useful. For enterprises seeking interchangeable components, competitive sourcing, and predictable upgrades, interface maturity is a central procurement issue rather than a technical footnote.
Advanced pluggables and CPO remain in the picture
Huawei’s announcement should not be read as evidence that NPO has replaced CPO or conventional pluggables. Multiple approaches continue to develop in parallel.
Broadcom, for example, has publicized both a CPO-equipped 102.4T Ethernet switch and a VCSEL-based 3.2T NPO solution. Its work in both categories illustrates that suppliers see different trade-offs at different switch capacities, densities, cooling arrangements, and operational models.
The broader market is also pursuing ways to extend pluggable optics rather than abandon their serviceability advantages. The reviewed material describes an XPO multi-source effort around a liquid-cooled, 12.8Tb/s pluggable module. The key point is not that one architecture has won, but that removable optics can also evolve through higher capacity and more aggressive cooling.
This is especially relevant for operators that value front-panel replacement, established sparing practices, and simpler fault isolation. A pluggable design may consume more power or require more physical space in some configurations, while offering a service model that infrastructure teams already know how to run. Conversely, NPO or CPO may offer compelling density or electrical-path benefits while requiring a different approach to thermal engineering and repair.
The appropriate choice will depend on the system. Port speed, rack power, cooling capability, switch design, workload topology, fault-tolerance goals, supplier strategy, and operational staffing all influence whether a more integrated optical design makes sense.
A practical checklist for Windows Server and hybrid AI teams
For Windows-focused enterprise teams, optics may seem distant from the operating system. In practice, organizations building Windows Server AI environments, hybrid clusters, or accelerator-backed services can be directly affected by network decisions made below the application layer. A faster optical interconnect does not automatically improve a workload; it must fit the servers, switches, fabrics, cooling systems, monitoring tools, and support processes that make up the complete environment.
Before treating an NPO roadmap announcement as a capacity plan, infrastructure and procurement teams should ask several practical questions:
- Is there a defined production and qualification path? Request a clear distinction between a demonstrator, a customer-evaluation unit, and a product supported for volume deployment.
- What exactly is the interoperability position? Establish whether the hardware conforms to a finalized or emerging interface, what equipment it has been tested with, and whether alternative suppliers are possible.
- How is a failure handled? Determine whether optical components are field-replaceable, which parts must be stocked, who performs service, and whether a fault can require removal of a larger switch assembly.
- What are the actual power and cooling requirements? Ask for complete switch and rack power envelopes, thermal limits, cooling assumptions, and operational behavior under sustained load—not only module-level claims.
- How will it be monitored? Confirm the availability of telemetry, fault reporting, firmware-update procedures, and integration with the organization’s network-management and incident-response workflows.
- What is the support boundary? Clarify responsibility across the server vendor, switch vendor, optical supplier, systems integrator, and on-site support provider.
- What is the fallback plan? A hybrid AI design should include a supported alternative for capacity expansion or hardware replacement if an announced optical architecture slips, changes form factor, or remains supplier-specific.
These questions are not unique to Windows Server, but they matter to Windows administrators who may be asked to support AI services whose stability depends on a shared accelerator fabric. The operating system cannot offset inadequate cooling, immature optics, difficult replacement procedures, or a network interface without a clear support path.
SuperPOD intent is not a delivery commitment
Huawei’s stated intention to use the NPO module in future SuperPOD iterations is strategically meaningful. Earlier in 2026, Huawei identified Atlas 950 SuperPoD and TaiShan 950 SuperPoD as its latest SuperPOD products, and said Atlas 950 integrated 64 NPUs per cabinet and could scale to 8,192 NPUs.
However, the reviewed material does not establish that this exact 7.2Tb/s module is committed to a named successor product. It does not verify a formal launch schedule, customer qualification milestone, price, shipment target, or deployed-fleet result. Reporting on the CIOE presentation also said Huawei declined to provide a SuperPOD availability date.
That leaves the module in the category of a relevant roadmap signal, rather than near-term deployable capacity. Organizations planning AI infrastructure can consider it in longer-range architecture discussions, but should not treat it as a confirmed procurement option until product, support, and qualification details are available.
Export controls are context, not a demonstrated outcome
Huawei and many of its non-U.S. affiliates have been on the U.S. Entity List since May 16, 2019. The U.S. Commerce Department states that the listing imposes license requirements for items subject to U.S. Export Administration Regulations supplied to listed entities, subject to specified exceptions.
That background is relevant to advanced networking supply chains, which can involve components, manufacturing equipment, and service relationships across multiple countries. It does not establish that such rules have specifically blocked NPO components, altered OIF work, or reduced adoption of Huawei’s approach.
The practical consequence is a need for disciplined supply-chain assessment, not an assumption of a proven direct restriction on this product. Buyers with cross-border operations may reasonably assess export-control exposure, alternate sourcing, spare-part continuity, and long-term support arrangements alongside the technical evaluation.
What would turn the demonstration into stronger evidence
Huawei’s 7.2Tb/s module highlights a real industry challenge: at very high signaling rates, the electrical connection between a switch ASIC and its optics can become costly in power and complexity. NPO is a plausible answer, and Huawei’s demonstration is a notable technical statement.
The next evidence needed is less dramatic than a bandwidth headline. Buyers should watch for interface alignment, mechanical and thermal specifications, reproducible latency and power testing, error-rate data, customer qualification, service procedures, volume availability, and long-term reliability information. Comparable evidence for CPO, NPO, liquid-cooled pluggables, and other advanced designs will be needed to make an architecture decision on operational rather than promotional grounds.
For now, Huawei’s hardware is best understood as an announced NPO demonstrator with a stated future role in SuperPOD planning. It shows why optical integration is becoming more important in AI networking, but it does not yet prove broad interoperability, superior lifecycle economics, or a settled winner among competing optical architectures.