Hua Hong Doubles Down on Optical Chips as AI Data Centres Race for Speed

The Bottleneck Shifts to Light

Hua Hong Semiconductor Manufacturing, China's second-largest contract chipmaker, has announced plans to substantially increase its capacity for silicon photonics, a technology that transmits data using light rather than electrical signals. Speaking during an analyst call on Thursday, chairman and president Bai Peng outlined the company's intention to expand silicon photonics production in its next major capacity build-out, positioning the foundry to serve optical modules used in artificial intelligence data centres.

Silicon photonics sits at the intersection of two urgent infrastructure demands. First, AI training clusters now link thousands of accelerators, and electrical interconnects are approaching physical limits on bandwidth and power efficiency. Second, the technology requires fabrication capabilities that overlap with mature-node logic processes, an area where Chinese foundries retain both equipment and expertise despite export restrictions on advanced lithography tools.

Hua Hong's move reflects a broader recognition among Asian semiconductor manufacturers that the performance ceiling for AI systems is shifting from compute density to data movement. Optical interconnects can carry terabits per second over longer distances with lower latency and heat generation than copper cables, making them essential for scaling training runs beyond tens of thousands of graphics processing units. The technology also underpins high-speed networking between data centres, a segment where demand has outpaced supply over the past eighteen months.

A Foundry's Calculated Pivot

The decision to prioritise silicon photonics represents a strategic recalibration for Hua Hong, which has historically focused on power management integrated circuits, micro-electromechanical systems sensors, and embedded memory. Unlike leading-edge logic, where the company faces both technological and regulatory headwinds, silicon photonics leverages process nodes in the 90-nanometre to 180-nanometre range, geometries that remain accessible to Chinese fabs and where Hua Hong has accumulated decades of manufacturing data.

Bai Peng's comments during the earnings call did not specify investment figures or timeline, but the company's characterisation of the expansion as part of its "next major capacity" phase suggests a multi-year, multi-facility commitment. Hua Hong operates fabs in Shanghai and Wuxi, with the latter site having ramped 300-millimetre wafer production over the past three years. Industry observers note that retrofitting existing 300-millimetre lines for silicon photonics is more capital-efficient than building greenfield capacity, though it requires specialised deposition and etching tools for integrating optical waveguides with standard CMOS transistors.

The timing aligns with a surge in demand for optical transceivers rated at 800 gigabits per second and above, modules that hyperscale cloud operators and AI infrastructure providers are deploying to connect GPU clusters within and across data centres. According to data from the China Semiconductor Industry Association, domestic shipments of optical communication chips grew 47 per cent year-on-year in the first three quarters of this year, outpacing growth in other semiconductor categories.

The Regional Race for Optical Supremacy

Hua Hong's announcement arrives as foundries and integrated device manufacturers across Asia compete for share in the silicon photonics supply chain. Taiwan Semiconductor Manufacturing Company has offered photonics process design kits since 2021 and counts several optical module vendors among its clients. In Japan, NTT and Sony have invested in co-packaged optics, a more ambitious architecture that places photonic chiplets directly alongside processors to eliminate electrical bottlenecks at the package level.

Chinese firms face a distinct set of constraints and opportunities. Export controls imposed by the United States and the Netherlands restrict access to extreme ultraviolet lithography and the most advanced deep ultraviolet immersion scanners, effectively capping domestic logic production at geometries above seven nanometres. Silicon photonics, however, does not require cutting-edge transistors; its performance hinges on the precision of waveguide etching, the quality of germanium or indium phosphide photodetectors, and the efficiency of optical coupling, all areas where mature fab equipment suffices.

This technical reality has made optical chips a focal point for Chinese semiconductor policy. The Ministry of Industry and Information Technology listed silicon photonics as a priority technology in its 2023 guidance on next-generation information infrastructure, and state-backed funds have directed capital toward startups developing lasers, modulators, and packaging solutions for co-packaged optics. Hua Hong's expansion can be read as both a commercial bet and a response to policy signals, a pattern common among foundries with significant government ownership.

Yet the technology's maturity remains uneven. While optical transceivers for data centre networking are in volume production, co-packaged optics and on-chip photonic interconnects are still navigating yield and thermal management challenges. The integration of III-V semiconductor lasers, which provide the light source, onto silicon wafers requires either heterogeneous bonding or epitaxial growth, both of which add cost and complexity. Foundries that can deliver reliable, high-yield silicon photonics processes stand to capture margin premiums, but the pathway to commoditisation is shorter than in logic, where each node transition resets competitive dynamics.

The Economics of Light and Scale

For Hua Hong, the economics of silicon photonics hinge on two variables: the pace at which hyperscalers adopt optical interconnects deeper into their infrastructure stacks, and the company's ability to differentiate its process technology in a market where TSMC, GlobalFoundries, and specialty photonics fabs already compete. The first variable is trending in Hua Hong's favour. Opentechwire has tracked public procurement data from cloud operators in Asia, and purchase orders for 800-gigabit and 1.6-terabit optical modules have increased more than threefold since the start of last year, driven by deployments of Nvidia H100 and H200 systems and their successors.

The second variable is less certain. Silicon photonics processes are not yet standardised, and customers often require foundries to support proprietary modulator designs, custom wavelength channels, or specific packaging interfaces. This places a premium on design enablement and process flexibility, areas where Hua Hong has historically lagged TSMC's ecosystem breadth. The company's ability to attract fabless photonics startups and module vendors will depend on the maturity of its process design kits, the speed of its tape-out cycles, and its willingness to co-invest in early-stage designs, a model more common in Taiwan and South Korea than in China's foundry sector.

Pricing dynamics also matter. Silicon photonics wafers command higher average selling prices than commodity mixed-signal processes, but the market is not large enough to absorb unlimited capacity. If multiple foundries expand simultaneously, wafer prices could compress before demand catches up, eroding the margin advantage that makes the technology attractive. Hua Hong's scale, already smaller than TSMC or Samsung, leaves it less room to absorb a protracted downturn in utilisation.

What the Bet Signals

Hua Hong's silicon photonics expansion is a wager that the infrastructure layer of AI, rather than the processor layer, will be the next arena of competition for foundries operating outside the leading edge. It is a rational wager. The computational demands of large language models and multimodal training are pushing data movement to the foreground, and optical interconnects are the only viable solution at the bandwidths and power budgets that hyperscalers require.

The move also reflects the narrowing menu of options for Chinese foundries. Blocked from sub-seven-nanometre logic and facing a glut in legacy nodes, companies like Hua Hong must identify adjacencies where their existing capabilities translate into new revenue streams. Silicon photonics is one such adjacency, and the fact that it serves AI infrastructure, a sector with sustained tailwinds, makes it doubly attractive.

Whether Hua Hong can execute remains an open question. The company has not disclosed which customers have committed to its photonics roadmap, nor has it detailed the performance specifications of its process technology. Those details will determine whether the expansion becomes a strategic differentiator or an expensive detour. For now, the announcement signals where one of Asia's mid-tier foundries believes the next growth vector lies, and it is not in shrinking transistors.