Chroma — leaderboard

Optical interconnect built into chips targets AI data bottleneck

Staera Semiconductor has emerged from stealth with US$4.6 million in pre-seed funding to develop LED-based optical interconnects built directly into silicon chips. The company—formerly New Silicon Corporation (NSC)—is targeting the growing bottleneck in AI systems: moving data between chips over the millimeter-to-meter distances spanning packages, boards, and racks.

The round was led by Openspace Capital, with participation from SGInnovate, Black Kite Capital, 28 Holdings, In Group, and Decimus Technologies. Staera’s first product is called Lightwire, which the company says replaces copper links with light generated inside the chip itself, manufactured with the same process and on the same wafer as the chip’s circuits.

Staera says it combines three technology layers on a single wafer: silicon CMOS circuits, GaN LED light sources, and silicon nitride waveguides that route light between links. The company says this is done using standard fab steps and is designed for manufacturing in existing fabs.

Most optical approaches today integrate discrete optical components next to a chip, such as lasers or LEDs made elsewhere. Staera’s approach is to fabricate the light source and the rest of the optical link inside commercial silicon wafers. The company says it has built GaN LEDs into silicon chips in commercial foundries since 2016, with processes that now run at 98 to 99 percent yield.

Lightwire specs and where it fits in packaging

Staera says Lightwire is built inside the bridge or interposer chip that already connects processors and memory in today’s AI hardware, aiming to work with existing packaging approaches. On performance, the company lists 4 to 8 Gbps per link and “more than 600 links per millimeter” of chip edge.

For energy, Staera says Lightwire is designed for under 0.2 pJ/bit, compared with 5 pJ/bit for co-packaged optics. It also claims more than 100x the performance of electrical links across the millimeter-to-meter range and reach “beyond 50 mm,” compared with 2 mm for UCIe-A die-to-die links. Staera also says Lightwire-based HBM is designed to exceed 24 TB/s, compared with 2.8 TB/s per HBM4 stack today.

The blunt takeaway for data center and AI infrastructure teams is that this is aimed at a hard scaling limit: feeding accelerators with enough bandwidth without turning interconnect power into a meaningful chunk of the rack’s thermal budget.

“AI’s bottleneck is no longer how fast chips compute, but how fast and far they can move data,” said Dr Andrew Kim, Chief Executive Officer of Staera.

Staera says it will use the funding to advance its technology, expand its team, and deepen engagement with customers and supply chain partners, targeting commercial release in 2030.

Source: Staera Semiconductor

AFF — below-posts ()

Get Data Center Engineering News In Your Inbox:

By subscribing, you agree to our Privacy Policy and Terms of Use. You can unsubscribe at any time.

Sponsored: Infineon at OCP 2026: Exclusive Content Hub. Executive interviews, technical documents, reference designs, and live expert panel. Unlock the hub.
TDK — tower ()

Popular Posts:

Dewalt-31
The data center critical path runs through the slab, and DEWALT wants a robot on it
DCE-AAF-Promo-Banner-Sentence-Case-1200x675
Webinar: Optimizing data center sustainability and efficiency through smarter air filtration
Vertiv-PowerStation-3100
Vertiv PowerStation 3100 remote power panel scales to 1,200 A for AI racks
Gates-ThermalCool-2100SS-Pump
Gates launches 4U ThermalCool 2100SS pump for rack CDUs, 200 L/min flow
Engineering AI Rack Power: Infineon Answers the Questions Asked at OCP, live expert panel with Carl Smith, Davide Chiola and Michael Williams
AI data center architecture evolution and key findings from OCP

Share Your Data Center Engineering News

Do you have a new product announcement, webinar, whitepaper, or article topic? 

AFF — tower ()

Get Data Center Engineering News In Your Inbox:

By subscribing, you agree to our Privacy Policy and Terms of Use. You can unsubscribe at any time.

AFF — leaderboard ()