Lightmatter joined Nvidia's NVLink Fusion program to supply photonic interconnects for next-generation AI infrastructure. The technology targets power and latency bottlenecks in GPU clusters housing tens of thousands of accelerators.
Key facts
- Lightmatter joined Nvidia NVLink Fusion for photonic interconnects.
- Photonic interconnects claim 80% power reduction vs electrical.
- Nvidia Kyber rack delayed to 2028 on PCB issues.
- Vera Rubin rack costs $7.8M; memory is key cost driver.
- Nvidia renting back GPU capacity from neoclouds.
Lightmatter, a startup developing photonic interconnects for AI workloads, has joined Nvidia's NVLink Fusion program, according to a press release covered by AOL.com. The partnership aims to integrate Lightmatter's silicon photonics technology into Nvidia's GPU cluster networking fabric, which currently relies on electrical interconnects from providers like Broadcom and Marvell.
The power problem
As GPU clusters scale to tens of thousands of nodes—Nvidia's Vera Rubin rack costs $7.8 million, with memory as the primary cost driver per MarketScale—interconnect power consumption becomes a dominant cost. Lightmatter claims its photonic interconnects can reduce power by up to 80% compared to electrical alternatives, while increasing bandwidth density. Nvidia's NVLink Fusion is the company's latest attempt to standardize the fabric connecting GPUs within and across racks.
Timing and competitive landscape
The announcement comes amid Nvidia's ongoing manufacturing challenges. According to Tom's Hardware, Nvidia's Kyber rack for Rubin Ultra—its highest-density design—has been delayed to 2028 due to PCB midplane manufacturing issues. A stopgap solution was also scrapped after customer pushback. Separately, CNBC reported Nvidia's next-gen AI rack system delayed to 2028 on manufacturing snags, citing SemiAnalysis.
Lightmatter's inclusion in NVLink Fusion signals Nvidia's recognition that electrical interconnects are hitting fundamental physics limits as cluster sizes balloon. But the timing is awkward: Nvidia is simultaneously renting back GPU capacity from neoclouds amid softening demand, suggesting hyperscalers are rethinking their buildout pace. Photonic interconnects might solve a future problem that arrives after the current capex cycle peaks.
Key Takeaways
- Lightmatter joins Nvidia NVLink Fusion with photonic interconnects.
- Nvidia Kyber rack delayed to 2028.
What to watch
Watch Nvidia's Q3 2026 earnings call for updates on Kyber rack timeline and any customer commitments to photonic interconnect adoption. Also watch Lightmatter's next funding round—the company has not disclosed its valuation or recent financials.
Source: news.google.com
[Updated 08 Jul via gn_infiniband]
Ayar Labs, a competitor to Lightmatter, has also joined Nvidia's NVLink Fusion ecosystem, bringing co-packaged optics for rack-scale AI [per AOL.com]. This dual-supplier move signals Nvidia is hedging its bets on photonic interconnect technology, pitting Ayar Labs' optical I/O chipsets against Lightmatter's silicon photonics.
