A two-year-old startup is worth $5.51 billion selling a component that Chinese firms already make at scale, and whose own CEO says will never beat copper on price. The FCC is drafting the rule that makes that math work.
Lumilens came out of stealth on August 6 with more than $700 million in new funding at a $5.51 billion valuation, bringing its total raised past $900 million. It builds optical interconnects that replace the copper wiring between AI chips. Its CEO concedes optics cost several times more than copper and always will. Two days earlier, Reuters reported the FCC is drafting a ban on imports of new Chinese optical transceivers, a category where Chinese vendors hold more than 60% of the global market.
The most interesting sentence in the Lumilens announcement is not about light. It is the part where the founder says he is committed to manufacturing outside of China. In a normal component business that is a footnote about factory siting. In this one, it may be the entire product.
What Happened
Lumilens, based in San Jose, announced more than $700 million in Series C funding at a valuation of $5.51 billion. The round was co-led by Atreides Management, Bain Capital Ventures, Meritech, Seligman Ventures and Spark Capital, with Qualcomm Ventures and JPMorgan among the other participants. Total funding since founding in 2024 now exceeds $900 million.
The company says it is already shipping its first product, an optical transceiver that converts electrical signals into light, into the live data centers of a large hyperscale cloud provider it declined to name. It describes the agreement as worth billions of dollars over several years. That figure is the company’s own, not an audited disclosure, and the customer is unidentified.
Founder and CEO Ankur Singla is not a first-timer. He sold Contrail Systems to Juniper Networks for $176 million in December 2012, nine months after founding it, and later sold Volterra to F5. His Lumilens team is drawn from Cisco, Juniper, Meta, Lumentum and Coherent.
The Backstory
The problem Lumilens is selling into is real and physical. At the signalling rates modern AI clusters run, an electrical signal over copper stays usable for roughly a metre and a half. That is fine for connecting a few hundred chips inside one rack. It is useless for wiring a hundred thousand accelerators into something that behaves like a single computer.
So the industry moves to light. Optical links carry more data further with less heat, and every accelerator you add multiplies the number of transceivers required. In large clusters, cables and transceivers already account for roughly half of total network cost. A dense AI rack can pull up to 600 kilowatts, and optical networking alone can eat about a tenth of that envelope.
None of this is a secret, which is why Nvidia moved first. On March 2, Nvidia announced it would invest $2 billion each into Coherent and Lumentum, paired with multibillion-dollar purchase commitments and future capacity rights, explicitly to expand US-based optical manufacturing. It also joined Ayar Labs’ $500 million round. That is a company with a networking business past $31 billion a year deciding the optical layer was too important to leave to the market, the same instinct behind its $5 billion purchase of research access at SSI.
Lumilens is one of several startups chasing the same shift. Lightmatter, founded in 2017, has raised $850 million and was valued at $4.4 billion in 2024. Celestial AI was pursuing a comparable approach until Marvell acquired it in February for $3.25 billion. Ayar Labs closed a $500 million round in March at $3.75 billion.
The Plan
Lumilens is building two product families. Scale-out optics connect separate server racks across a data center. Scale-up optics connect chips inside a single rack, which is the harder and more valuable problem because that is where copper is failing fastest.
The company has put unusual emphasis on manufacturing rather than chip design alone, describing heavily robotic production lines and automated test and calibration as the mechanism for driving yields up and unit cost down. That emphasis follows directly from Singla’s own framing of the constraint. Optical gear runs several times more expensive than copper, which makes it hard to justify unless the performance gain is large.
And he says he is committed to building outside China.
The Business Model Angle
Read those two commitments together and the valuation stops looking like a bet on photonics and starts looking like a bet on procurement law.
Here is the market Lumilens is entering. According to LightCounting, seven of the world’s ten largest optical module vendors are Chinese, with a combined global share above 60%. Counterpoint Research puts Zhongji Innolight alone at about 27% of the data center transceiver market. Innolight booked RMB 38.24 billion of revenue in 2025, up 60%, with net profit up 109%. Eoptolink grew revenue 187% in the same year. These are not marginal suppliers being disrupted by a better mousetrap. They are the incumbent supply base, they are profitable, and their customers are Nvidia, Google, Meta and Amazon.
On August 4, Reuters reported that the FCC is drafting a measure to bar imports of new models of Chinese optical transceivers, with officials hoping to publish it this year. The reported shape matters: a broad ban on new transceiver models, followed by exemptions for many non-Chinese suppliers. Existing certified models already in volume supply would not be caught. Innolight was added to the Pentagon’s list of Chinese military-linked companies in June.
That structure is the whole ballgame, and it turns on a definition nobody has published yet.
Notice what the trade data already shows. China shipped $61.6 million of optical transceivers to the US in June, just 8.7% of its worldwide total. The US used to be the largest destination. Malaysia now is, at $260 million that month, as data center investment there surges. Chinese optical exports overall rose 27% in the first half of the year. The rerouting happened before the rule was drafted.
It goes further. Innolight and Eoptolink have both built production in Thailand specifically to serve North American demand from outside Chinese export exposure. Innolight earns roughly 90% of revenue overseas. Eoptolink exports about 96%. So if the FCC writes an origin-of-manufacture rule, the incumbents have already stepped around it. If it writes an entity rule that follows the company rather than the factory, those Thai lines do not help and roughly two thirds of global supply becomes unbuyable for US operators overnight.
Now flip it around, because the mirror image is just as awkward. Coherent, the American supplier most often named as the beneficiary of a ban, owns Finisar Photonics in Wuxi, which employed more than 9,700 people in 2025 and was planning a RMB 1 billion expansion in 2026 to reach 10 million units a year of 800G and 1.6T modules, equal to over 7% of projected global AI optical module demand. The designated winner manufactures in China at scale.
Which leaves a very small set: firms that are neither Chinese-owned nor China-manufacturing, at volume, with a hyperscaler already qualified. Lumilens has spent two years and $900 million building itself into that set. That is not a technology moat. It is a compliance moat, and compliance moats are worth the most in the eighteen months before everyone else can copy them. The margin in this stack keeps migrating toward whoever controls the scarce physical layer, and for the moment, scarcity is being manufactured in Washington rather than in a fab.
The Risk
The category may get deleted before it matures. Lumilens’s shipping product is a pluggable transceiver. Nvidia and Broadcom are both pushing co-packaged optics, which moves the optical engine into the switch package and cuts interconnect power from roughly 15 picojoules per bit to around 5. Nvidia’s Quantum-X and Spectrum-X Photonics platforms and Broadcom’s Bailly are the first commercial generations. Analysts currently expect coexistence rather than replacement, with pluggables holding structural dominance across enterprise and lower-bandwidth tiers into the 2030s, but the direction of travel is toward integration. Lumilens is also building scale-up optics, which is the right hedge. It is not yet the revenue.
One customer, one unaudited number. A multibillion-dollar agreement with an unnamed hyperscaler is a single-counterparty revenue base disclosed on the seller’s word. Hyperscalers multi-source deliberately. Nvidia has already bought priority access at Coherent and Lumentum, which means the largest buyer in the industry has pre-committed capacity elsewhere.
The rule might not happen. Reuters’ sources stressed the FCC could modify or shelve it. Commerce shelved a comparable set of data center import restrictions last October during a trade détente. Beijing has said it will respond to actions that materially harm its interests, and it has been building its own outbound investment controls in parallel. A tariff-and-thaw cycle is not a foundation for a $5.5 billion valuation.
The ban’s economics cut both ways. Optical transceivers were 0.2% of China’s total June exports. Restricting them barely dents China’s export engine while raising input costs for the US operators funding the largest capex cycle in corporate history. A policy that slows American data center construction to protect American optics suppliers is not a stable policy.
Quick Questions
What does an optical transceiver actually do? It converts electrical signals into light so data can travel over fiber, and back again at the other end. In an AI cluster it is the thing that lets thousands of accelerators talk to each other beyond the couple of metres copper can handle.
Why not just keep using copper? Copper is cheaper and reliable, and it is still used inside racks. It degrades badly over distance at high data rates, which caps how many chips can act as one machine. Scaling past that cap requires light.
Is Lumilens actually shipping? It says yes, into a large unnamed hyperscaler alongside that customer’s own AI chips. The customer and the contract value are the company’s disclosures, not independently verified.
Would an FCC ban really hit Chinese suppliers that manufacture in Thailand? That depends on whether the rule follows the company or the factory, and the FCC has not published its definition. This is the single most consequential open question for everyone in the sector.
Who else competes here? Lightmatter, Ayar Labs, Marvell via Celestial AI, plus incumbents Coherent, Lumentum, Broadcom, Cisco through Acacia, and the Chinese leaders Innolight, Eoptolink and HGtech.
The Business Model Analyst Take
The honest version of the Lumilens pitch is not that it beats copper. Its own CEO says it does not and will not. The honest version is that a US operator building an AI cluster in 2027 may soon have three requirements to satisfy at once: enough bandwidth, an acceptable cost per bit, and a supplier that survives contact with the Covered List. Most of the industry can satisfy two.
That is a genuinely good business to be in, and it is also a rented one. The scarcity Lumilens is priced against is regulatory, which means it can be widened by an exemption, narrowed by a definition, or erased by a trade deal. Coherent and Lumentum are expanding US capacity with Nvidia’s money. Innolight and Eoptolink are already offshoring. Every month the rule stays in draft, the moat gets shallower.
The useful lesson for anyone building a components business right now is not about photonics. It is that supply chain nationality has quietly become a product feature, and it is one of the few features a competitor cannot ship an update to match. It is also the only feature on the spec sheet that a regulator can revoke.
Lumilens raised $700 million to sell light. What it actually sold investors was a seat on the right side of a rule that has not been written yet.
