A proposed US ban on new Chinese optical transceivers could disrupt the build-out of AI data centers before Western suppliers have enough capacity to replace Chinese production, raising the risk of higher costs, delayed cluster deployments and stranded GPU capacity.
The Federal Communications Commission (FCC) is reportedly considering restrictions on new China-sourced optical transceiver modules, components that convert electrical signals into light for high-speed fibre links inside data centers. The proposal comes just as AI clusters containing tens of thousands of GPUs are accelerating the shift from copper connections towards 400G, 800G and emerging 1.6T optical networks.
The problem is scale. Counterpoint Research estimates Zhongji Innolight leads global data-center optical transceiver revenue with about 27%, followed by Coherent at roughly 17%, while Eoptolink is another major supplier to US hyperscalers. Chinese manufacturers including Innolight, Eoptolink, Accelink and Source Photonics collectively account for nearly two-thirds of global unit shipments and about 60% of datacom transceiver revenue.
That concentration makes a rapid substitution difficult. Coherent, Lumentum and other Western suppliers have competitive photonics technology, but lack enough cleanroom capacity, automated packaging lines and high-volume yields to absorb Chinese output over the next 12 to 24 months.
AI clusters face an optics capacity gap
For Amazon Web Services (AWS), Microsoft, Google and Meta, the immediate risk is less the price of replacing Chinese modules than whether sufficient alternatives exist at all.
Optical transceivers are not interchangeable commodities. Modules must undergo compatibility testing against network switches, digital signal processors and broader data center architectures. Removing a qualified supplier can therefore trigger lengthy requalification cycles even when replacement hardware is technically available.
That matters more as AI networks migrate towards 800G and 1.6T. A shortage of comparatively inexpensive optical modules can prevent far more valuable GPUs from entering service, delaying the commissioning of billion-dollar clusters and reducing returns on AI infrastructure investment.
Hyperscalers would have more leverage to secure alternative supply than enterprises, regional data center operators and colocation providers, which often rely on distributors and lower-cost Chinese components. Smaller buyers could therefore face sharper price increases and delivery delays while competing against companies such as Microsoft and Meta for scarce non-Chinese capacity.
The exact wording of any FCC measure will be critical. Reports indicate restrictions could focus on new modules, potentially preserving existing installations while constraining next-generation products. A broader approach based on corporate ownership rather than manufacturing origin would have wider consequences because Chinese suppliers have already expanded production outside China.
Innolight and Eoptolink, for example, have built automated manufacturing capacity in Thailand, where final assembly, optical alignment, testing and shipment can be carried out for overseas customers. Restrictions tied to parent ownership could therefore capture products manufactured in Southeast Asia rather than only those exported directly from China.
A supply chain too intertwined to divide cleanly
The optics ecosystem is also less geographically separable than a US-versus-China framing suggests. Chinese module companies integrate high-speed DSPs from Broadcom and Marvell with lasers and optical components supplied by companies including Lumentum, Coherent and Mitsubishi Electric.
Innolight's advantage therefore rests not simply on manufacturing cost, but on high-volume integration of thermal management, precision optical alignment and advanced packaging. Reproducing those yields at 1.6T speeds requires capital investment, customer qualification and time.
Western module suppliers such as Coherent, Lumentum and Applied Optoelectronics could ultimately gain market share and pricing power if restrictions persist. Cisco's Acacia business and suppliers in Japan and Taiwan could also benefit from diversification. But upstream US companies selling components into Chinese modules could simultaneously lose business, illustrating how a ban would disrupt an interdependent supply chain rather than separate two self-contained ecosystems.
The security case is similarly contested. Supporters argue modern transceivers can contain firmware, memory and management interfaces that create risks involving compromised updates, diagnostic manipulation or infrastructure disruption. Critics contend transceivers remain relatively simple devices compared with routers and other network equipment, making blanket restrictions disproportionate to the immediate cyber threat.
Either way, procurement would become more complex. Operators may need greater visibility into firmware origin, subcontractors, component manufacturing and software-update processes rather than simply identifying the company that sells the finished module.
Chinese suppliers also have some room to absorb a US shock. Direct Chinese optical-module exports to the US totalled about US$61.6 million in June, or 8.7% of global shipments, according to data cited in Chinese reporting, reflecting manufacturers' expansion into other markets and offshore production. Malaysia has become an increasingly important destination alongside its rapid data-center investment.
That does not eliminate Chinese vendors' exposure to US AI spending. Their highest-value 800G and 1.6T businesses remain closely linked to overseas hyperscalers, while restrictions based on corporate ownership could undermine attempts to route production through Thailand and other markets.
The dispute could also widen. US scrutiny has increasingly extended beyond chips towards the networking, power and control systems surrounding AI data centers, potentially bringing switches, SmartNICs, data-processing units and management hardware under greater attention.
China has already warned against further restrictions. Near the close of the latest debate, the People's Daily said Beijing would take countermeasures if Washington continued expanding technology controls under national-security grounds.
For the US, the strategic trade-off is becoming harder to ignore. Reducing dependence on Chinese optical hardware may strengthen supply-chain control over time, but moving faster than alternative capacity can be built risks imposing a near-term ceiling on AI infrastructure itself.
In a market where GPUs receive most of the attention, the constraint could come from something less conspicuous: whether enough qualified optical modules exist to connect them.
Article edited by Jack Wu
