When a cloud provider builds an AI data center, the GPUs get the headlines. But the servers cannot talk to each other — or to the next building, or the next continent — without tiny, high-speed "walkie-talkies" plugged into every machine. Those devices are optical transceivers, and the company making the most of them sits in Suzhou (苏州), in China's Jiangsu province.
Zhongji InnoLight (中际旭创, 300308.SZ) has become the default supplier of the high-speed optical modules that turn electrical signals from chips into light pulses traveling down fiber. According to LightCounting's 2024 ranking, it was the world's No.1 optical-module vendor, with over 40% of the 800G segment and a reported 50–70% of the emerging 1.6T segment.
Why optical modules are the bottleneck everyone plans around
An AI training cluster is a network problem disguised as a compute problem. Thousands of accelerators must exchange gradients continuously; the speed of that exchange often decides how long a training run takes. Optical modules are the interfaces doing that exchange, and they must be qualified and stocked before the GPUs even arrive.
That ordering reality is why InnoLight's business moves ahead of the broader chip cycle. Cloud builders place module orders months early, which is part of why the company's results turned up first among AI-hardware suppliers.
The financial scale
InnoLight's 2025 annual report, published March 2026, shows the magnitude:
- Revenue of RMB 38.24 billion (≈ US$5.27B / HK$41.1B), up 60.25% year on year.
- Net profit attributable to shareholders of RMB 10.797 billion (≈ US$1.49B), up 108.78%.
- Gross margin of 42.04%, up about 8 points, helped by a richer mix of 800G/1.6T and higher silicon-photonics content.
- Overseas revenue above 90% of total, concentrated in North American cloud customers.
The product range runs from 100G to 1.6T modules. The 800G generation became the volume workhorse through 2024; 1.6T began customer shipments in the third quarter of 2025 and started ramping in the following quarters, with several customers' orders extending into 2026–2027 according to the company's investor communications.
Who the customers are
InnoLight serves the largest global compute buyers: Nvidia, Google, Amazon, Microsoft, and Meta are all cited as core customers. Sell-side analysis of Nvidia's Blackwell bill of materials has estimated InnoLight at roughly 80% of Nvidia's 1.6T module purchases, with Nvidia orders around 60% of the company's revenue — figures the company has not confirmed in detail but has not contradicted.
That concentration is both the strength and the risk: deep binding to the top of the AI build-out drives the results, but any shift in a single customer's supply chain would be felt sharply.
Why a Suzhou company leads a "global" market
The surprise is not that optical modules exist — they have shipped for decades — but that one Chinese supplier now sets the pace for the highest-speed tier. The reason is partly industrial ecosystem: around InnoLight sit dense clusters of optical-component, laser, and test suppliers, plus deep manufacturing-process discipline built over years of volume 100G/400G production. When the spec jumped to 800G and then 1.6T, the same lines and supply relationships scaled up rather than starting cold.
The other reason is cadence. Because modules are qualified ahead of GPUs, InnoLight's order book reads as a forward indicator of where accelerators will be deployed months before utilization data appears. That lead time is why its shipment mix is watched as a pulse check on the whole AI-infrastructure cycle.
The next frontier: co-packaged optics
As data rates climb, plugging modules into sockets starts to waste too much power and space. The industry is moving toward co-packaged optics (CPO), where the optical engine sits next to the switch or compute ASIC. InnoLight has demonstrated silicon-photonics-based 800G/1.6T modules and is developing CPO-aligned products; it also notes linear-drive pluggable optics (LPO) samples with sub-4W power for short-reach links.
What readers can do now
- If you follow AI-infrastructure investing, track optical-module shipment mix (800G vs 1.6T) as a leading indicator of accelerator deployment, since modules are ordered ahead of GPUs.
- Watch silicon-photonics and CPO adoption — they are the next cost-and-power battleground and determine which suppliers keep their margins.
- Note customer concentration as the key risk: a diversify-or-depend dynamic that applies to the whole optical supply chain, not just one firm.
Honest limitations
Revenue and profit figures are taken from InnoLight's official 2025 annual report and corroborating exchanges filings; market-share percentages (800G >40%, 1.6T 50–70%) come from LightCounting via media reporting and should be read as analyst estimates, not audited counts. The ~80% Nvidia 1.6T share and ~60% revenue-concentration figures are sell-side estimates the company has neither confirmed nor denied. Geopolitical export and tariff exposure, which materially affect a >90%-overseas-revenue firm, is discussed only qualitatively.
Key takeaways
- Zhongji InnoLight (中际旭创) is the world's No.1 optical-module vendor (LightCounting 2024), with >40% of 800G and a reported 50–70% of 1.6T transceivers.
- 2025 revenue hit RMB 38.24B (≈ US$5.27B), +60%, with net profit RMB 10.797B (≈ US$1.49B); over 90% of sales are overseas to top cloud and AI buyers.
- Modules must be qualified before GPUs ship, making InnoLight's mix a leading signal for AI-cluster build-out — but customer concentration is the central risk.