Hanul Photonics closed a $145 million round on Saturday co-led by SK Hynix Ventures and Lightspeed, with participation from Korea Development Bank and Samsung Catalyst Fund. The Seoul company builds co-packaged optical interconnects, moving the optical transceiver from a pluggable module at the edge of a switch onto the same substrate as the switching chip or the accelerator. Its claim is 3.1 picojoules per bit against roughly 7 for current pluggable optics at comparable bandwidth, plus lower latency and higher density. Two hyperscalers are evaluating the technology and one has committed to a production deployment in 2028, though Hanul would not say which.
The problem with wires
Training a large model requires accelerators to exchange gradients constantly, and the interconnect has become the limiting factor in cluster design. Copper works well over short distances and fails quickly as bandwidth and distance increase, which is why the reach of an electrical link at 224 gigabits per second per lane is roughly a meter. That constraint shapes the entire data center: racks must be dense, and anything beyond a rack requires optics.
Pluggable optics solve the distance problem and cost energy. A conventional transceiver consumes 15 to 20 watts, and a large cluster has hundreds of thousands of them. Estimates from several hyperscalers put interconnect power at 20 to 30 percent of total cluster consumption, which at a gigawatt-scale facility is a number worth attacking with real capital. Co-packaged optics attack it by shortening the electrical path to nearly zero.
What Hanul built
The company's approach uses a silicon photonics die flip-chipped onto the same organic substrate as the compute or switch die, with an external laser source delivered through a fiber connector. Keeping the laser off-package is the design choice most of the industry has converged on, because lasers degrade with heat and a package running at 90 degrees Celsius is a poor place for one. Hanul's differentiation is in the coupling structure between fiber and chip, which it claims achieves 1.1 decibels of loss against a typical 2.5.
Chief executive Ji-woo Park, who spent eleven years at SK Hynix working on high bandwidth memory packaging, said the coupling work came from applying assembly techniques developed for memory stacking to a photonics problem. That heritage explains the SK Hynix investment and the manufacturing arrangement: Hanul assembles at an SK facility rather than building its own line, which keeps capital requirements down and creates an obvious dependency.
The story is rarely the launch. It is what breaks, what ships, and who owns the mess at 2 a.m.
The competition is large and well funded
Broadcom shipped co-packaged optics in its Tomahawk 5 generation and has the switching market position to force adoption. Nvidia announced photonic switches at its 2025 developer conference with silicon from a partner. Ayar Labs, backed by Nvidia and Intel among others, has been building optical input-output chiplets since 2015 and has raised more than $370 million. Lightmatter and Celestial AI both compete in adjacent parts of the problem.
Hanul's argument for existing in that field is geographic and structural. Korean and Japanese memory and packaging supply chains are where high bandwidth memory is made, and co-packaged optics is fundamentally a packaging problem. A company embedded in that ecosystem with access to SK Hynix's assembly capability starts from a different place than a Silicon Valley startup contracting with a Taiwanese assembler. Whether that advantage is durable depends on how much of the value is in the photonics design rather than the assembly.
Timeline realities
Semiconductor startups measure progress in tape-outs and qualification cycles rather than in quarters. Hanul has silicon working in a laboratory at the claimed performance and has not run the reliability qualification that a hyperscaler requires, which involves thousands of hours of thermal cycling and accelerated aging. Park estimated eighteen months to complete that, which puts a production commitment in 2028 at the earliest, matching what the company disclosed.
The risk in that window is that the standards move. The Optical Internetworking Forum and several vendor consortia are still arguing about interfaces for co-packaged optics, and a company that bets on the wrong interface spends a year adapting. Hanul is participating in the standards work, which is what everyone in this position does and which provides visibility rather than protection.
Why investors are writing these checks
The capital flowing into interconnect startups reflects a straightforward calculation. If AI cluster spending continues at anything like the current pace, and if interconnect is a quarter of the power and a meaningful share of the cost, then a technology that halves it addresses a market measured in tens of billions annually by the end of the decade. Even a small share of that supports a large outcome.
The counter-scenario is that the hyperscalers build it themselves or that Broadcom's position proves unassailable, in which case these companies become acquisition targets at modest multiples. Several investors we spoke with described that as an acceptable downside, which is a revealing thing to hear about a $145 million round and consistent with how semiconductor investing has always worked. The pattern in photonics specifically has been that the technology arrives roughly three years after the market expects it and then becomes mandatory very quickly, which rewards whoever is qualified when the switch happens and punishes everyone who ran out of money waiting. Hanul now has enough capital to reach the far side of that gap, which is most of what this round buys.
Skarvonix will keep following this beat with reporting grounded in how systems behave outside the launch keynote.
- Funding
