ENG
Insight

Hot Interconnects: New Form Factors, Fibers and Use Cases for Optics

0
August 28, 2026

The technology that was tough to find at Hot Interconnects, a companion conference to Hot Chips that took place August 19-21, was copper. Outside of the occasional mention of copper flyover cables, almost every presentation focused on optics and coming technologies like CPO. But presenters made it clear that, while co-packaged optics (CPO) is imminent, it also won’t be pervasive, at least not the start, as new classes of modules and NPO provide an easier onramp for higher-bandwidth networks.

Key Takeaways

  • Modules aren’t dead: they’re just morphing. Open CPX and the XPO MSA outlined technologies for scale-out, scale-up, CPO, NPO, and pluggables through 12.8T and beyond.
  • The first Multicore Fiber specifications come soon and have room to grow.
  • Marvell and others meanwhile, discussed ways optical into systems.


Corning enumerated the benefits of Multicore Fiber, which combines four optical cores into the same 125-micron fiber that holds only one today: 70% less cable mass in networks, 75% fewer connections, up to 60% less installation time, and greater reliability. The first specification from the MCF MSA, for O-band connections (<2km), comes soon.

Why four fibers? One must start somewhere. The upper limit is likely close to 19 cores, said Corning’s Gabe Sudduth.


Corning Multicore Fiber
Four is better: Corning says Multicore Fiber will eliminate heft from optical networks; room for nearly 5x more fibers exists.



Resolight aims for optical immortality with a platform that combines an optical switch with CPO which co-founder Ofer Shapiro claims can reduce network power by 40x. In 2025, networking accounted for around 10% of power in a scale-up rack. That rises to 50% in 256-576 XPU configurations and 80% in 1,000-2,000 XPU systems. Reducing optical power is also expensive, costing $1 billion to reduce power by 1 picojoule/bit, Shapiro claimed.

Lower power and cost also enable complex scale-up systems to better fit into the confines of edge data centers and reduce the distance to consumers. The technology, however, remains in development.

Marvell discussed how its Photonic Fabric (acquired from Celestial AI) will enable XPUs to share 72GB of HBM or 1-2TB of system DRAM making it a potential successor to CXL. Interesting note: 1 meter of fiber cam add 5ns of latency. The sweet spot for this technology is 30 meters and below.

Marvell Photonic Fabric memory nodule
A potential successor to CXL in future years, Photonic Fabric also features HBM sharing.


XPO MSA, meanwhile, described advances in its supersize module with 12.8Tbps of bandwidth largely for scale-out. Like Open CPX, the XPO MSA plans to come to market by aligning with the existing module ecosystem.

Arista, the primary proponent of the technology, said XPO is compatible with LPO, but admitted that 200G/lane LPO remains a work in progress. Eoptolink, among others, is developing XPO modules containing multiple DSPs and optical engines.

XPO Emulation Program
The SUV of modules. Despite the large size and component heft--four DSPs, 64 channels and integrated liquid cooling (and copper flyover cables)—XPO modules increase bandwidth density.


Open CPX, which is developing open standards for CPO and NPO modules, showcased a prototype combining a 100T Arista switch with a detachable 6.4T Terahop optical engine for CPO, an unusual CPO showing for Arista, often a critic of CPO. 80+ members have joined Open CPX. The organization is already studying 12.8T optical engines (achieved through more lanes or 400G technology).

Terahop’s Ryan Yu also outlined the development of CPX-IAC, an intermediate connectivity scheme between NPO and CPO.

open cpx three versions
IOA? Think of it as optics that sits on the processor substrate but outside the package.


Oak Ridge National Laboratory described Omnistat, a telemetry system for high-bandwidth, large data sets like climate models or nuclear simulations. Today, telemetry data in these systems is viewed retrospectively. Omnistat streams data at terabits per second on a separate network, thereby allowing simultaneous monitoring.

Cloudflare, meanwhile, noted that 58.6% of the traffic it manages for clients now comes from robots versus 41.4% from humans. Just so you know.


Learn More From Counterpoint's Cloud AI Service

  • Cloud AI Capex Tracker & Forecast: Quarterly AI capex by hyperscaler and neocloud with component-level breakdowns; long-term forecasts by operator.
  • AI Accelerator & Deployment Tracker: Quarterly AI accelerator shipments and installed base by vendor, architecture and cloud operator, including in-house ASIC programs.
  • Cloud AI Capacity & Power Tracker: AI data center capacity additions by region with power, cooling and rack density metrics.
  • Cloud AI Services Pricing Tracker: Instance and managed-service pricing across major clouds, with cost-per-token benchmarks for leading model families.
  • Cloud AI Intelligence Report: Monthly analysis of cloud AI announcements, capacity deals, supply agreements and partnership activity.
    • Contact us to sign up for Cloud AI's upcoming newsletter.

Receive our insightful weekly newsletter and stay ahead of the competition.

Author

Michael Kanellos

Michael Kanellos has covered semiconductors, data centers and renewable energy for over 30 years as an analyst, reporter and marketing executive. His work has appeared in CNET, Forbes, Wired, Greentech Media, the New York Times, USA Today, The San Francisco Chronicle, the Chicago Tribune and other publications. Along the way, he’s drunk biofuel, explored underground data centers, almost crashed an electric motorcycle prototype, and been zapped with a Taser to entertain crowds at CES.