AI networking is scaling rapidly, but the manufacturing supply chain for optical interconnects is not keeping pace. In my latest @650Group blog, I look at one of the less-discussed constraints on AI infrastructure growth: indium phosphide (InP) and the industry's ability to manufacture optical components at the volumes AI factories will require. AI networking is on a path to exceed $300B this decade and approach $1T early next decade. But getting there will require more than faster switches, new architectures and open standards. We need to rethink how optical components are manufactured, moving from telecom-era production models to semiconductor-scale manufacturing capable of supporting billions of lasers. The industry made major investments in advanced packaging and HBM to keep pace with GPU demand. Networking needs a similar shift. I explore why heterogeneous integration of silicon photonics and III-V materials will be critical to scaling AI infrastructure, and why manufacturing must advance alongside networking innovation. Read the full blog: 650group.com/blog/ai-buildou… #AIInfrastructure #AINetworking #OpticalNetworking #SiliconPhotonics #650Group

Heading into #OCPGlobalSummit, open networking and SONiC will be a big part of the AI infrastructure conversation. I shared my perspective in @Cisco's blog on where scale-across is heading and why we expect this market to exceed $50B by the end of the decade. Worth a read: blogs.cisco.com/datacenter/s… #SONiC #AINetworking #650Group

I’m looking forward to moderating @HPE's “Reimagining SONiC to Power AI at Scale” webinar and digging into how SONiC is evolving to support the next generation of AI infrastructure. As AI clusters grow, the network has to scale in multiple directions at once — scale-up, scale-out and increasingly scale-across. That is creating new requirements around openness, automation, operational simplicity and flexibility. I’ll be joined by Sridhar Katere, VP Engineering at HPE, and Marty Ma, Director of Product Management at HPE, for a discussion covering: • Where SONiC adoption stands today • Deployment and operational challenges • SONiC across scale-up, scale-out and scale-across architectures • Agentic AIOps for production networks • Deployment and support models • The role of the broader SONiC community There’s a lot changing quickly as AI reshapes networking architectures. I’m looking forward to a practical discussion about where SONiC fits and what operators need to think about as these networks move to production. Register here: rsvp.hpe.com/reimaginingsoni… #SONiC #AIInfrastructure #Networking #AIOps #DataCenter #HPE

Great to be in San Diego this week for @nokia's SReXperts Americas. One point from Ron Johnson’s keynote really stood out: the optical industry needs to move ZR/ZR+ transceiver production to a whole new level to keep pace with demand. Today, production is measured at roughly a dozen units per hour. Where the market is headed will require several units per minute. And we think the demand curve could be even steeper as Hyperscalers and Neoclouds scale AI infrastructure and agentic workloads begin to dominate traffic. The implications extend well beyond optics. The entire supply chain will need to think differently about manufacturing scale, capacity and how quickly the industry can ramp. More to come. We’ll be sharing a @650Group blog with our key takeaways from SReXperts Americas soon. #SReXperts #AIInfrastructure #OpticalNetworking #ZR #DataCenters #650Group

Multi-core fiber could change how optical capacity is packaged and scaled, but what will it take to deploy? At 3:45 p.m. PT on Oct. 7, I’ll moderate a @EthernetAllianc #TEF2026 panel on standards, lane breakouts, transceivers and its role in Ethernet AI networks. ethernetalliance.org/tef-202… #TEF2026

A new optical market is forming between the data center fabric and traditional long-haul coherent networking. At @650Group, our latest forecast projects the Coherent Lite market will surpass 10 million units within the next three to five years. The driver is straightforward: AI clusters are getting physically larger. As they spread across buildings and campuses, operators increasingly need high-capacity links across roughly 2-10 kilometers, and traditional ZR optics carry more reach, power and cost than those shorter links need. That creates room for a purpose-built architecture. What this points to: – Scale-across is becoming part of the AI system architecture, not conventional WAN plumbing. – 800G is moving outside the building, across short unamplified links where full ZR is overkill. – Power per bit favors optics built for the actual reach instead of over-provisioning every connection. A multi-vendor ecosystem and clear standards will matter as volumes rise. AI infrastructure is expanding the definition of the network, and the fabric no longer stops at the data center wall. Read the new 650 Group forecast: 650group.com/press-releases/… #CoherentLite #DCI #ScaleAcross #OpticalNetworking #AIInfrastructure #DataCenter #650Group

ICYMI: AI models—not just GPUs—are rewriting the rules of network infrastructure. MoE, expanding context windows and disaggregated inference demand larger pods, flatter fabrics and high-radix switching. My latest @650Group blog: 650group.com/blog/mixture-of… #AIInfrastructure

Faster GPUs are only part of the AI infrastructure story. MoE models, expanding context windows and disaggregated inference are creating all-to-all traffic that demands larger pods, flatter fabrics and high-radix switching. My latest @650Group blog: 650group.com/blog/mixture-of… #AIInfrastructure