A Shanghai silicon-photonics startup has raised several hundred million yuan for the second time in thirteen months. Bopu Semiconductor, founded in Shanghai in 2022, closed a Series B of several hundred million yuan, the company announced on November 6, 2026. Huaying Capital led the round, with Kailian Capital, Southeast Space Frame, the Shanghai Science and Technology Venture Capital Group Intellectual Property Fund, and Shunyi Investment participating, according to Dealroom’s account of November 9. No exact figure was disclosed, and no valuation was published. In the usual range of the phrase, “several hundred million yuan” means somewhere between roughly 100 million and one billion yuan — a range wide enough that the round could fund either two comfortable years of development or a single expensive tapeout cycle with little left over. The ambiguity is not a detail. It is the first thing to notice.
This is a financing story, not a results story. The company says the money will fund research and development of silicon photonic switching chips and their processes, plus next-generation high-speed optical transceiver chips and optical engines for chip-level optical interconnection. It says it reached large-scale product shipment in 2025. What it has not disclosed — no customer names, no order volumes, no shipment batches, no yields — is everything that would let an outsider tell whether the company is selling a product or selling a roadmap. The funding is real, as far as multiple outlets reporting the same round can establish it. The technology’s commercial standing is not yet established at all.
The round, and what two large rounds in thirteen months means
The Series B follows a similarly sized Series A in October 2025, led by Shanghai Guohe Investment, according to Dealroom and to Star Market Daily’s November 6 report. Two rounds of several hundred million yuan inside thirteen months is a fast cadence for a four-year-old chip company, and it can be read two ways. The generous reading: investors are topping up against milestones, and the company is hitting them. The sceptical reading: silicon photonics is capital-intensive at every step — tapeouts, packaging, testing, system integration — and a company building a full chain from chip design to mass production burns through several hundred million yuan faster than a software company burns through a seed round. Both readings are consistent with the public record, because the public record contains no milestones, no revenue figures, and no customer list.
The founder: a returnee with the right two backgrounds
Dr. Xing Yufei founded Bopu in Shanghai in 2022 with a team of returnees, according to Dealroom. He earned a doctorate in integrated silicon-based optoelectronics at Ghent University in Belgium, and later built chips and wafer-level test platforms at Xanadu, the Canadian quantum-computing firm. That combination — academic silicon photonics and industrial chip-building and test experience — is genuinely rarer than either alone, and it is the most plausible explanation for the fundraising cadence. A founder who has built chips and the test equipment to validate them is a founder who can credibly promise a full chain from design to mass production.
The company is based in Zhangjiang Science City, Shanghai’s semiconductor and hard-technology cluster. Its business, as Dealroom describes it, is full-link solutions from chip design to mass production: high-speed silicon photonic transceivers, IP licensing and custom development, wafer-level test equipment, optical circuit switching, and optical engines for AI computing. That breadth is the pitch and the risk at the same time: a full chain means control of its own destiny, but also a research budget spread across five product lines, each of which must succeed before any one of them pays.
The technology: optical switching, where light’s advantage is real and narrow
The centre of this round is optical circuit switching — OCS — and it is worth being precise about what OCS is and is not. In a conventional AI cluster, GPUs talk through electrical switches: light carries data in, is converted to electricity, switched, and converted back to light. Each conversion costs latency and power. An all-optical switch keeps the signal in the optical domain, routing light paths inside the chip without the optical-electrical round trip. In principle, that eliminates conversion latency and cuts power substantially.
In practice, OCS has been a laboratory technology for years, and the gap between a working switch and a shippable product is where Bopu’s story will be decided. An all-optical switch must hold insertion loss, crosstalk, switching speed, packaging consistency, and temperature stability inside specification simultaneously — and it must do so on every unit, not just the first. The company’s press release, as analysed in detail by RecodeX, gives the 32×32 OCS model an insertion loss within 5 dB, a switching speed of 100 microseconds, and total system power consumption of 10 W, with the company claiming power “an order of magnitude lower than traditional electrical switching solutions” and key metrics “among the international first tier.” Those are company claims, not independently verified results.
Five decibels of insertion loss means roughly two-thirds of the optical signal’s power is lost passing through the switching matrix, so the system must budget enough link margin to absorb it. A 100-microsecond switching speed suits reconfiguring a training cluster’s topology between jobs, but it is too slow for protection switching scenarios that need microsecond or nanosecond response. Ten watts is genuinely low against an electrical switch — if it holds at system level, including the control electronics and the laser. And “international first tier” is a claim without a comparison table: no competitor specifications were disclosed alongside it, so there is no way to judge where the 32×32 model stands against comparable products.
The product line Bopu debuted at the 27th China International Optoelectronic Exposition in September 2026 covers OCS switches from 8×8 to 32×32 ports — small and mid-sized switching nodes, suitable for specific subnets or reconfigurable interconnect layers inside AI clusters, rather than for replacing a cluster’s core switching fabric. The company says it will move from on-demand supply to stable volume delivery of OCS products within the next twelve months, and is developing larger 64/72 and 128/144 switching matrices alongside 3.2T–6.4T optical engines for near-packaged and co-packaged optics. But every doubling of port count drives a non-linear rise in crossbar complexity, accumulated insertion loss, and packaging difficulty, and the company has disclosed no figures for the larger matrices — nor whether they have been validated at chip or system level at all. The larger matrices are, for now, a roadmap.
There is also a competitive question that Bopu’s own narrative does not answer. The real competitor to a standalone OCS switch may not be another OCS startup but rapidly improving electrical switching and co-packaged optics: if 800G and 1.6T electrical switches keep improving on power and latency, and co-packaged optics integrates the optical engine directly into the switch package, the market window for a standalone optical switch could be narrower than the funding thesis assumes. Bopu’s simultaneous bet on OCS and on optical engines for near- and co-packaged optics is, in part, a hedge against that risk — but it is also a second roadmap competing for the same research budget.
The mass-production claim, and the twelve-month test
The most consequential claim in the company’s announcement is not a specification but a process claim: that it has completed process and mass-production validation on a domestic 12-inch CMOS silicon-photonics line. A 12-inch line means far more chips per wafer than the 6-inch or 8-inch lines traditionally used for silicon photonics, and unit costs should in theory be lower — but the requirements for waveguide sidewall roughness, lithography alignment accuracy, and wafer-level test coverage are also more demanding. The company has not named the line, disclosed yields, or given wafer-level test coverage or shipment batch volumes. So a gap remains between “mass-production validation completed” and “stable volume delivery capability,” and it is a gap that only order and delivery data can close.
Against that gap, the company has two anchor points that are at least partly checkable. First, it says several of its silicon-photonic PIC chips have already shipped at scale — a claim about chip-level mass production, though without volumes attached. Second, it runs a second business line: optical sensing. According to background reporting from the Series A period, Bopu’s sensing business covers LiDAR, vibration, velocity, and ranging, with an FMCW LiDAR product in joint development with a leading domestic LiDAR maker and a silicon-photonics aortic pulse-wave-velocity monitor in joint development with a listed biopharmaceutical company. Sensing has a shorter customer validation cycle than data-centre switching, so it may generate revenue faster — though the company has disclosed no revenue scale for the sensing business, and the Series B announcement did not mention it at all. Whether sensing can fund the OCS roadmap while switching finds its customers is an open question.
The wider wave: capital is moving into optical interconnects
Bopu’s round is part of a broader capital push into optical interconnects, and the pattern is now unmistakable. Dealroom notes that Guanglian Xinke raised nearly US$206.3 million across two 2026 rounds for chip-level optical I/O. In September 2026, Guoke Guangxin completed a Series B+ of several hundred million yuan led by Sichuan Development, according to Star Market Daily. And this desk covered Zhizi Guangyuan two weeks ago: five seed rounds in six months for reconfigurable all-optical computing. The money is moving in the same direction across different slices of the same problem — getting data between processors with light instead of copper.
That convergence has a real driver. As AI training clusters scale from thousands to hundreds of thousands of GPUs, the network interconnect is replacing the compute chip as the ceiling on a cluster’s effective compute, and electrical switching is approaching physical limits in bandwidth density, latency, and power. Light is the obvious answer. But belief is not evidence: the largest sums are going to interconnect and switching — the nearer-term, more measurable problem — while the more speculative all-optical computing bets raise in smaller, faster tranches. Bopu sits on the nearer-term side of that divide, which makes its claims more checkable and its failures more visible.
What would make this story matter
As with every photonic AI announcement this desk covers, the right response is a list of checkable developments, not a verdict.
First, volume delivery: the company’s own twelve-month commitment — stable volume delivery of OCS products — is the clearest test it has ever set itself. Watch for the number of units delivered, to whom, and against what quality standards. If it delivers standardised products at volume, the round will look cheap in hindsight; if it can only deliver small batches of engineering units to a handful of joint-development customers, the “core foundation of domestic silicon photonics” narrative will need revising.
Second, a named data-centre customer: background reporting from the Series A period says Bopu had signed joint research or supply contracts with a leading domestic LiDAR maker, a listed biopharmaceutical company, and optical module companies in China and overseas. Those were sensing and module relationships, and no update has been disclosed showing them extending to volume data-centre orders for OCS switches. A named cluster operator buying OCS switches — with enough detail to distinguish a pilot from a purchase — is the transition that matters.
Third, larger matrices with numbers: the 64/72 and 128/144 switching matrices need published insertion loss, switching speed, and power figures, and evidence that the 5 dB loss of the 32×32 model does not deteriorate as port counts double. A matrix that works at 32 ports and fails at 128 is a product for subnets, not for clusters.
Fourth, the optical engines: the 3.2T–6.4T engines for near- and co-packaged optics are Bopu’s hedge against the standalone-OCS window closing. Whether they arrive as products or remain a roadmap item will tell readers how seriously to take the company’s claim to a full chain.
Until those appear, Bopu belongs in a category this desk watches closely but does not celebrate: well-funded, well-pedigreed, and commercially unproven. Two rounds of several hundred million yuan in thirteen months buys a great deal of time. It does not buy a customer, a yield, or a delivered switch. The next twelve months — the company’s own deadline — will say whether the money bought a product or just more time to promise one.
Sources and further reading
Primary sources: Dealroom, “Bopu Semiconductor raises several hundred million yuan for AI data-center photonics,” November 9, 2026 (round size, announcement date of November 6, investor group, founder background, product lines, use of funds, Series A timing, large-scale shipment claim); Star Market Daily, November 6, 2026, via Longbridge (investor group including Southeast Space Frame, use of funds, Series A led by Shanghai Guohe Investment, Guoke Guangxin Series B+ in September 2026). Company claims — including the 32×32 OCS specifications (insertion loss within 5 dB, 100-microsecond switching, 10 W system power), the “international first tier” characterization, mass-production validation on a domestic 12-inch CMOS line, the twelve-month stable-delivery commitment, the 64/72 and 128/144 matrix roadmap, the 3.2T–6.4T optical engine plans, and the large-scale shipment claim — are reported as company statements in its press release, as analysed by RecodeX Pro, November 6, 2026, and are not independently verified results. Background on the optical sensing business and the Series A-era customer relationships is from reporting of the 2025 Series A period and reflects the situation at that time; no update on those items was disclosed with the Series B. No exact round amount, valuation, customer names, order volumes, yields, or shipment batches were publicly available at the time of writing. Note on names: Dealroom renders the lead investor as “China Vision Capital” and one participant as “Southeast Grid”; the Chinese press gives the lead as Huaying Capital and the participant as Southeast Space Frame (东南网架), the listed steel-structure company, which this article follows. For the desk’s earlier coverage of the same funding wave, see the November 4 piece on Zhizi Guangyuan’s five seed rounds.