(By Wang Runlian of Science and Technology Power)
"These chips might just end up sitting idle." John Miller, General Counsel at the Information Technology Industry Council (ITI), told Reuters. As US restrictions on Chinese optical modules loom closer, American companies and industry associations are growing increasingly anxious.
On November 10, Reuters reported that Microsoft and other firms are continuously pressuring the Trump administration to prevent a ban on importing key AI data center components, such as optical module transceivers, from China. The trigger for this situation was a new prohibition being drafted by the FCC three months ago. The agency plans to restrict the entry of new models of Chinese optical transceivers into the US market citing national security concerns, hoping to officially implement related measures within the year.
Sources indicate that Microsoft and several other companies believe such a ban could further slow down data center deployment, irreparably damaging the US advantage over China in the AI race. This is because many US companies have already purchased the AI chips required to build systems, but without sufficient optical transceivers, these chips would be essentially useless.
The impact of missing optical modules is cascading. A large-scale data center requires millions of optical transceivers to connect servers, routers, and switches. According to Cleanview data, as of November 11, there were 4,634 tracked data center buildings in the US, with 1,427 operational, 537 under construction, and another 2,670 in the planning stage.
This means the number of unbuilt projects exceeds twice the current number of operational projects. These future data centers will still require massive quantities of optical transceivers to become operational.
Currently, Chinese enterprises are the primary global suppliers of optical modules, holding approximately two-thirds of the global optical transceiver market share. According to LightCounting's 2025 Global Optical Module Supplier Ranking, Chinese companies occupy seven of the top ten spots, with Zhongji InnoLight and Eoptolink taking the first two places. Chinese firms such as Zhongji InnoLight, Eoptolink, Tianfu Communication, Accelink, Cambridge Industries Corporation (CIC), and HGTECH have become significant forces in the global optical module industry.
For US tech companies, excluding Chinese optical modules is no longer just a matter of switching suppliers; it risks disrupting the entire construction plan for AI data centers.
After official signals of restriction emerged from the US government, the industry reaction was intense. The ITI submitted written comments to the FCC, urging them not to restrict optical module imports. In September, two network equipment vendors publicly warned that a ban could stall US data center construction. During the same period, tech giants like Microsoft directly raised supply chain risks with White House officials, attempting to block the implementation of the ban.
Cignal AI, an optical communications research firm, estimates that even if foundries in other regions expand production, it would take years to absorb the supply volume of leading Chinese manufacturers. Given current capacity and data center construction progress, the US currently lacks the capability to大规模 exclude Chinese optical modules, and filling this gap will remain difficult in the coming years.
Without Optical Modules, Microsoft's Computing Power Is Stuck
Microsoft's biggest headache right now is insufficient computing power. According to Cleanview statistics, Microsoft has 35 data centers under construction and 134 in the planning stage in the US. In late October, Microsoft disclosed that its quarterly capital expenditure reached CNY 45.2 billion (approx. USD 6.3 billion based on context, but sticking to original currency implication or converting? Original says '452亿美元' which is USD. Wait, constraint says: "Original text does not mark currency... default CNY". Here it explicitly says "452亿美元" (452 billion dollars). So I keep USD. Let me re-read carefully. "452亿美元" -> $45.2 billion. Okay.) roughly three-quarters of which were invested in short-lived assets such as CPUs and GPUs. In the same earnings call, Microsoft stated that some third-party computing power originally scheduled for delivery in the quarter had been delayed. The reason was that Microsoft temporarily requested network upgrades to run newer AI models, but suppliers indicated that existing equipment could not meet the requirements.
This computing power supplier is IREN. In August this year, IREN handed over Horizon 1 to Microsoft. By November, the facility was nearly complete, but IREN was still waiting for remaining equipment deliveries and debugging. The online target for the relevant computing power has been postponed to January next year, and if deliveries are further delayed, it may drag on until the end of February. A major reason for the delivery delay is the shortage of key components.
Microsoft's urgency at the White House stems from supply commitments already included in construction schedules. Nomura Securities' survey predicts that Microsoft's demand for 800G optical modules in 2026 will be approximately 3 million to 4 million units. SDxCentral reports indicate that Zhongji InnoLight is one of Microsoft's main suppliers, while also supplying Google. Eoptolink collaborates with Amazon. TrendForce data from February this year shows that Google's demand for 800G and higher-speed optical modules in 2026 will exceed 6 million units, with nearly 80% of orders expected to be won by Zhongji InnoLight and Eoptolink.
Moreover, the restrictive signals released by the US have not affected orders for Chinese domestic companies. On November 6, Zhongji InnoLight disclosed that shipments of its third-quarter 800G and 1.6T products continued to expand, with revenue growth from 1.6T products accelerating further. Orders from the company's major clients have basically covered the full year of 2027, with some clients providing clear demand forecasts for 2028.
In other words, as demand for optical modules surges, Chinese optical module companies are not lacking orders, and some cloud service providers are locking in future capacity through long-term contracts. Under these circumstances, if the US restricts companies like Microsoft and Google from using Chinese optical modules, they will be at a severe disadvantage in competition with other cloud service providers.
US domestic companies are certainly trying to expand capacity, but progress is far slower than needed, and it is difficult to break away from the Chinese supply chain.
Applied Optoelectronics Inc. (AOI) has successively expanded factories in Texas this year. Disclosed on November 5, the company's monthly production capacity for 800G and 1.6T optical modules has approached 330,000 units, with plans to increase it to over 1 million by the end of 2027. However, only about 30% of this is expected to be produced in Texas, while the remaining capacity still relies on factories in Ningbo, Taiwan, and elsewhere.
In March, optical device giant Lumentum acquired a factory in North Carolina owned by Qorvo, preparing to convert it into an indium phosphide laser production line, but scaled-up production is not expected until 2028.
On November 2, Marvell Technology announced its intention to acquire Lumentum for an enterprise value of approximately USD 115 billion. The former specializes in high-speed signal processing chips, while the latter possesses lasers, optical modules, and related manufacturing technologies. The two companies hope to integrate these segments to shorten product development and delivery cycles. However, this transaction is expected to close in the first half of 2028.
This means that although US companies are expanding production and making acquisitions to complete their supply chains, the release of new capacity will take at least one to two more years. Meanwhile, US cloud service providers' data center construction is already accelerating, making distant water unable to quench near thirst.
In the long term, even if these factories come online sequentially, raw materials still cannot bypass China. Lasers require indium phosphide substrates, and US-listed company AXT has its relevant production capacity located in Beijing. In June this year, Coherent prepaid approximately USD 22.29 million to AXT to secure its six-inch indium phosphide substrate capacity expansion from 2026 to 2028. In July, Lumentum also signed a six-year supply agreement with AXT, with the initial deposit alone reaching USD 43.5 million.
China began implementing export license management for relevant indium phosphide materials starting in February 2025. One of the key issues focused on by Coherent's CEO during his visit to China in May was license approval.
US Has Optical Device Giants, But Cannot Fill Finished Product Gaps
In fact, the main issue preventing US domestic optical module capacity from rising quickly lies not in Chinese materials, but in structural problems within the US industry itself.
For example, AOI leased a building adjacent to its Sugar Land headquarters, originally planning to move existing module production lines there in December to free up space for laser expansion at the headquarters. However, the construction of subsequent new production lines has been delayed by several months because transformers, cables, and other power supply equipment are unavailable.
AOI's CFO explained that numerous US data centers are competing for these devices. In other words, data centers urgently need optical modules, while the factories producing them struggle to obtain timely power supply equipment for expansion due to the rapid pace of data center construction.
Beyond power supply issues, talent is another significant gap. In February this year, AOI announced plans to invest up to USD 300 million to expand its Texas factory and add 500 jobs. These positions primarily involve programming and operating automated production lines, roles that ordinary workers cannot immediately fill.
Photonics Spectra pointed out that what the US optical communications industry lacks most are technicians capable of participating in production, testing, and troubleshooting. A study estimates that by 2030, the US photonics industry will need to add approximately 42,000 mid-skilled technicians. At the current rate of graduate training, around 100 additional relevant training programs are needed to meet this demand.
Secondly, complex validation procedures pose a barrier. Lumentum stated in its annual report that some customers only purchase small sample quantities before a production line passes certification. Even moving already-sold products to a new factory or changing an OEM may require re-validation of the new line. Once at the data center, it must also be confirmed whether the modules can run stably on the customer's selected switches and software.
This is why, despite having optical device giants, the US struggles to rapidly boost optical module finished product output.
The predicament of the US optical module manufacturing industry was foreshadowed more than twenty years ago. After the dot-com bubble burst in 2000, global demand for optical communications plummeted, and product prices crashed. To save costs, US optical communication giants like Finisar and JDSU gradually transferred manufacturing operations to Asia.
US companies initially aimed to reduce costs by relocating manufacturing, but this inadvertently gave Chinese companies opportunities to accumulate production experience and refine their supply chains. With the development of China's telecommunications network construction and cloud computing markets, local manufacturers expanded their scale. By 2020, many US and Japanese suppliers had exited the optical transceiver market. Zhongji InnoLight and Eoptolink developed products alongside top US cloud vendors for years, accumulating orders, skilled workers, and manufacturing support.
LightCounting data shows that in 2021, the combined revenue of China's top ten optical device and module suppliers exceeded that of Western competitors for the first time.
Today, rebuilding production lines in the US still relies on this division of labor. In October, US manufacturer Starman and Taiwan's Wistron announced a joint venture plan to invest USD 45 million to produce 800G and 1.6T modules in the US. According to both parties' estimates, mass production will begin no sooner than six months later.
Cignal AI estimates that even if mature foundries in other regions expand production, it would take years or even longer to absorb the supply volumes of Zhongji InnoLight, Eoptolink, and HGTECH.
Ban May Be Postponed, Next-Generation Orders Already Contested
In reality, the US government is aware that a complete ban on Chinese optical modules is unrealistic. The Senate bill proposed in September has left a loophole. The bill targets federal critical systems, proposing implementation five years after enactment, requiring an assessment of design and production capabilities in the US and partner countries first, while leaving room for exemptions.
The draft even states that implementation should leave procurement space for compliant third-party modules to avoid proprietary certifications from switch original manufacturers excluding other suppliers. The FCC equipment authorization rules effective in October involve specific hardware components produced by entities on controlled lists.
Recently, the House Select Committee on China named Zhongji InnoLight, urging the FCC to continue advancing restriction measures. ITI's opinion to the FCC is that if action must be taken, the scope should be limited to companies already on relevant controlled lists. Since even federal critical system procurement considers that alternatives take years, pushing restrictions to commercial data centers first requires someone to provide qualified goods.
Some experts believe that even if the draft lands, the range of affected products should be limited, leaving transition periods or exemption space for current supplies. The plan disclosed by the US in August centered on new model Chinese optical modules, originally intended for publication within the year. By November, discussions were still ongoing. Reuters reported that one option under consideration is setting the restriction threshold at 3.2T.
According to expectations cited in reports, this generation of products will not see volume ramp-up until 2028, with widespread deployment in 2029. Coherent demonstrated 3.2T pluggable modules in September, while companies under Japan's NTT and Furukawa Electric are developing mass production processes for optical chiplets through NEDO projects. Setting the threshold at the next generation allows 800G and 1.6T products to continue supplying according to existing arrangements, giving US manufacturers more time for factory commissioning and customer validation.
Whether to use production location as the restriction criterion is also debatable. The Thai government discussed local suppliers and talent cultivation with Lumentum in October. Foundry Fabrinet has already generated revenue from production lines in Building No. 10 in Chonburi and plans to start constructing Building No. 11 in January next year. Eoptolink and Zhongji InnoLight are also expanding production in Thailand.
If restrictions are based on origin, products manufactured by Chinese companies in Thailand may still enter the US market, while AOI's products made in Ningbo might fall under the ban. There remains significant room for discussion here.
However, US manufacturers do not intend to wait until 2028 to contest orders. In October this year, AOI signed a multi-year agreement with an existing cloud client. Subject to meeting supply capability conditions, the client promised to allocate at least 10% of its 1.6T optical module demand to AOI between June 2026 and May 2029.
For US manufacturers, such long-term contracts help lock in customers early and provide guarantees for capacity expansion. But Chinese companies are also vying for the next-generation market. Zhongji InnoLight has started R&D on 3.2T optical modules, with some next-generation products undergoing joint development and sampling with key clients.
For cloud service providers like Microsoft and Google, supplier selection for new-generation optical modules often begins before large-scale procurement. Only those who pass validation earlier and deliver stably have the chance to win subsequent orders.
This also means that even if the US ultimately raises the ban threshold to 3.2T, giving domestic companies one to two years for capacity expansion, it cannot guarantee that orders will flow to US manufacturers. Chinese companies will not stop R&D just because the US delays restrictions, and cloud service providers' computing power construction will not automatically slow down because US factories are not yet ready.
What the US truly needs to solve is whether domestic companies can deliver sufficient quantities of optically validated, reasonably priced optical modules before the next round of large-scale product deployment. Otherwise, when 3.2T starts mass deployment, the supply difficulties Microsoft raised with the White House today will likely reappear.