Huawei has demonstrated data transmission with an NPO optical module rated at 7.2Tbps at the IPEC booth of the China International Optoelectronic Expo (CIOE). The company disclosed this in an announcement dated September 10, 2026. It reflects a push to commercialize a design that places the components handling optical signals close to the compute chip while keeping them as separate parts. In AI optical interconnects, NVIDIA and Broadcom are moving CPO, which integrates optical components and chips in one package, into production. Evaluating Huawei's demonstration requires looking beyond the size of the bandwidth to where the optical components sit, how they are replaced, and how parts from different manufacturers can be combined.
What the 7.2T demonstration shows
In Huawei's announcement, Man Jiangwei, who leads the company's advanced optoelectronics research division, said Huawei had launched a 7.2T NPO product and successfully performed a dynamic transmission demonstration at the IPEC booth. Huawei calls this an "industry first," but that is Huawei's own assessment, not the result of an independent comparison across the industry.
The product reportedly uses a built-in light source and highly integrated silicon photonics. Silicon photonics refers to technology that builds optical circuits on silicon. Huawei touts high bandwidth and reliability, and also cites low latency, low power consumption, and low cost as advantages.
However, the announcement does not give absolute figures for power consumption or latency, nor the test conditions. The 7.2Tbps figure is the product's rated bandwidth; it does not mean AI training speeds will rise by the same amount, or that this speed will always be achieved in any given communication. Moving from a demonstration to large-scale operation will require evaluation that includes the chips used and the cooling conditions.
NPO itself as an approach is not new. Ruijie Networks also showed a design that mounts optical components and a switch chip separately, in its description of an earlier product. What is new this time is that Huawei demonstrated a 7.2T product and presented the implementation at a forum to discuss standardization.
Bringing optical components close to the chip, but keeping them separate
An optical engine is the component that converts between electrical and optical signals. In Ruijie's description, NPO keeps this optical engine and the switch chip separate and mounts both on the same system board. CPO, by contrast, combines the two in a single package, which changes the length of the paths electrical signals travel and how the parts are handled.
Why does placing them closer reduce power? In its March 12, 2026 technical explainer on NPO, Broadcom explained that placing the optical engine close to the compute chip can reduce the signal loss that occurs in internal electrical wiring. In that design, the DSP circuit that processes signals no longer needs to sit inside the optical module, which reduces the power the circuit uses and the bill-of-materials cost.
The benefit of switching to optical communication is not determined by the fiber segment alone. How far the electrical signal must travel from the chip to the point where it is converted to light also matters. NPO shortens that electrical segment while leaving the optical engine as an independent component.
Separating the parts also affects the choices open to those building systems. Ruijie explains that an independent optical engine interface allows flexible selection of the switch chip and the NPO module. Being able to choose chips and optical components separately matters when considering suppliers and repair methods.
Still, the name NPO does not guarantee interoperability across all products or replacement while a system is running. The range of parts that can actually be swapped, and the shutdown procedures involved, need to be confirmed for each product's implementation. Likewise, the announcement does not reveal specific replacement procedures for Huawei's 7.2T product.
CPO production has begun, and Broadcom also proposes NPO
In its May 31, 2026 announcement, NVIDIA presented Spectrum-X Ethernet Photonics as a CPO switch in production. Broadcom, in its May 15, 2025 announcement, said Tomahawk 5-Bailly had become a volume-production CPO product. Discussing when CPO will spread at scale is a different matter from confirming whether individual products are being produced.
Arranging each company's announcements by product approach and the stage stated at the time of publication gives the following.
| Company / Product | Approach | Date announced | Stage indicated in the announcement |
|---|---|---|---|
| Huawei 7.2T optical module | NPO | Sept. 10, 2026 | Product launch and transmission demonstration |
| NVIDIA Spectrum-X Ethernet Photonics | CPO | May 31, 2026 | In production |
| Broadcom Tomahawk 5-Bailly | CPO | May 15, 2025 | Described as a volume-production product |
| Broadcom 3.2Tbps VCSEL optical engine | NPO | March 12, 2026 | Optical engine concept proposed |
Broadcom is presenting both a volume-production CPO product and an NPO optical engine concept in parallel. Reducing the lineup to Huawei versus NVIDIA and Broadcom, and mapping that directly onto NPO versus CPO as a conflict between approaches, would leave Broadcom's NPO work out of the picture.
The table classifies what each company had made public as of September 11, 2026, by status: demonstration, production, or proposal. Optical modules and optical engines are different units of product from switch products that bundle multiple connections. The table is therefore not a ranking that compares bandwidth directly, and it cannot tell us actual shipment volumes from the companies' statements about production.
For anyone considering Huawei's product as a candidate for adoption, predictions about when a competing approach will become widespread are not enough; a comparison with the implementations available today is needed. Can it meet the bandwidth required for the same use, and can it be operated, including cooling and maintenance? The competition will be over who offers products that meet those conditions.
Replacement and supply conditions that remain beyond standardization
The NPO standardization summit on September 10 was co-hosted by IPEC and OIF. The official CIOE agenda lists a presentation by OIF introducing its next-generation NPO module project and a talk by Broadcom on VCSEL-based NPO. VCSEL is a type of surface-emitting laser, and the program shows that even within NPO there are options for the light source and implementation.
The announcement issued in Huawei's name says the summit reached agreement on international NPO standardization. But sharing a direction is not the same as making products from different manufacturers work together. To let the benefits of separating the optical engine from the chip be widely used, connection specifications must be aligned and interoperability between products must be verified.
Huawei's announcement gives no specific volume-production date, price, or number of customer deployments for the 7.2T product. What comes after the demonstration is information on when it can be procured and operational data that users can compare. Knowing how far a failed unit must be replaced, and how much power an actual system draws, would let designers tie the rated bandwidth to conditions for adoption.
Can the 7.2T bandwidth be turned into a product with an easily replaceable part structure and continuous availability? If that can be confirmed, NPO could become an option that increases communication volume in AI systems while preserving freedom in maintenance and component selection.