Rumors suggest that NVIDIA will not introduce optical interconnects for its NVLink GPU-to-memory binding until the 2027 “Rubin Ultra” GPU compute engine. This creates a window of opportunity for hyperscalers and cloud builders, who are racing to deploy optical interconnects before NVIDIA to gain a competitive edge in AI workloads.
Because bandwidth bottlenecks between accelerators and memory are so severe, the demand for optical interconnects is surging—and attracting massive venture funding. One of the most notable players is Xscape Photonics, a startup spun out of Columbia University’s photonics research programs.
Columbia University: A Hotbed for Photonics Innovation #
Columbia has long been a hub for interconnect and photonics research.
- Professors Al Gara and Norman Christ pioneered a DSP-driven supercomputer with custom interconnects that won the 1998 Gordon Bell Prize. This QCDSP system later inspired IBM’s BlueGene supercomputers, for which Gara served as chief architect.
- A separate team of Columbia researchers, including Keren Bergman, Alex Gaeta, Michal Lipson, and Yoshi Okawachi, laid the foundation for Xscape Photonics. Their expertise spans silicon photonics, quantum optics, nonlinear photonics, and optical frequency combs.
Interestingly, when it came time to commercialize, they tapped Vivek Raghunathan as CEO. A Columbia outsider, Raghunathan brought hands-on industry experience from Intel, Rockley Photonics, and Broadcom, where he worked on co-packaged optics for 25.6 Tbps and 52.6 Tbps switches deployed by Tencent and ByteDance.
NVIDIA Joins Xscape’s $44M Series A #
Xscape Photonics recently secured $44 million in Series A funding, following a $13 million seed round in 2022.
- The round was led by IAG Capital Partners.
- Other backers include Altair, Cisco Investments, Fathom Fund, Kyra Ventures, LifeX Ventures, Osage University Partners—and notably, NVIDIA.
NVIDIA’s participation is telling. The company already uses copper-based NVLink-NVSwitch interconnects in its GB200 NVL72 rack systems but has hinted at co-packaged optics (CPO) concepts since 2022. By investing in Xscape, NVIDIA signals strong interest in scaling optical interconnects to turn entire data centers into “one giant virtual GPU.”
The Bandwidth Bottleneck Problem #
No matter the architecture, AI accelerators face the same challenge: bandwidth falls off rapidly once data leaves the GPU die.
High-bandwidth memory (HBM) stacks must sit extremely close to the compute engine, but capacity scaling is limited, and HBM remains expensive and scarce. The result: GPUs often underperform simply because they can’t move data fast enough.
As Raghunathan explains, GPU utilization rates in training can drop below 50% due to networking bottlenecks. Meta has reported workloads where 60% of training time was consumed by GPU-to-GPU communication. For inference, utilization may sink to 30–40%, leaving billions of dollars in GPU investments underutilized.
Xscape’s ChromX Platform: One Laser, Many Wavelengths #
Xscape Photonics’ breakthrough is its ChromX programmable photonics platform.
- Instead of relying on multiple lasers, ChromX uses a single laser that generates up to 128 wavelengths, offering up to 32x the bandwidth of conventional 4-color optical interconnects.
- The system uses simpler NRZ modulation to minimize latency versus PAM-4 used in Ethernet and InfiniBand.
- Crucially, ChromX is programmable, dynamically matching wavelength count and distance to specific AI training or inference workloads.
This enables highly flexible architectures:
- Training clusters spanning <2 km
- Cross-data-center AI workloads over 20–40 km
- Inference engines with dense GPU-HBM fabrics over 10–200 m
Toward Disaggregated Compute + Memory Fabrics #
Xscape envisions a future where GPUs and HBM memory pools are disaggregated yet seamlessly interconnected via programmable optical fabrics.
Here, memory is decoupled from GPUs and distributed across racks, while accelerators share data through a programmable optical switch fabric. The result:
- 10x higher interconnect bandwidth
- 10x lower power consumption
- 100x improvement in bandwidth-per-watt efficiency
Why It Matters #
For AI training and inference at hyperscale, bandwidth is destiny. Xscape’s one-laser, multi-wavelength approach breaks both cost and scaling barriers, enabling GPU clusters to function more like unified compute fabrics.
With NVIDIA’s backing, Xscape is well-positioned to shape the next generation of AI infrastructure. If successful, its technology could turn multi-data-center deployments into what appears to be a single massive GPU—unlocking entirely new performance levels for large-scale AI workloads.
✅ Key Takeaway:
NVIDIA’s investment in Xscape Photonics signals a future where optical interconnects unlock 10x bandwidth and 10x energy savings, redefining the economics of AI data centers.