AI Optical Networking Rewires Data Centers for GPU-Scale Performance
Runaway AI training demand is pushing data center fabrics past their limits, making optical networking the bottleneck to unlock GPU-scale performance and efficiency.
AI compute growth is outpacing data center networks
Enterprises and hyperscalers are stitching together ever-larger GPU clusters where bisection bandwidth, microsecond latency, and lossless transport determine training throughput and job economics. As accelerators scale from racks to supernodes to multi-site clusters, the interconnect—not the chip—often gates utilization. This is shifting value from servers to networking, with optical links replacing copper at shorter reaches and higher speeds to meet AI’s traffic patterns.
AI cluster expansion models: scale-up, scale-out, and scale-across
AI infrastructure growth follows three paths. Scale-up connects more GPUs within a box or across tightly coupled racks to form supernodes with ultra-low-latency fabrics. Scale-out links many of these systems via switching topologies, now spanning 100,000+ GPUs in leading AI clusters. Scale-across extends connectivity across data centers and metros for geo-distributed training and inference. Each step increases optical reach, lane rates, and system “dollar content” in networking.
Goldman Sachs Forecasts $154B AI Optical Networking Market by 2028
A new forecast from Goldman Sachs positions optical networking as the next mega-trend in AI infrastructure, with spend growing an order of magnitude as clusters densify.
Scale-up drives spend as optical content per port surges
The analysis projects total addressable market expanding to roughly $154 billion by 2028 across scale-up and scale-out domains, up dramatically from mid‑decade levels. Scale-up networking—inside and across racks forming supernodes—accounts for about 69% of that total (near $106 billion), reflecting the premium on ultra-low latency and high radix fabrics. As architectures shift toward tighter coupling, spend on optical modules and engines rises sharply, with the optical content opportunity expanding more than tenfold versus today’s scale-out centric designs.
CPO adoption rises with 800G/1.6T pluggables and LPO
Co-Packaged Optics (CPO)—integrating optical engines with switch ASICs or accelerators—features prominently in the growth outlook. Under reasonable adoption assumptions, CPO could represent a majority share of the incremental market value by 2028 due to higher content per port and energy savings, even with partial penetration in scale-out. Expect a technology mix that also includes pluggable 800G/1.6T optics and emerging Linear Pluggable Optics (LPO) to reduce DSP power at short reaches. The transition will be phased by reach, serviceability needs, and ecosystem maturity.
AI Fabric Stack: Switch Silicon to Co-Packaged Optics
AI fabrics are being re-architected end-to-end, from link speeds and packaging to congestion control and telemetry.
Ethernet, InfiniBand, and silicon roadmaps define AI fabrics
Ethernet and InfiniBand remain the primary choices for AI scale-out. Ethernet is advancing with RoCEv2, enhanced congestion control, and initiatives like the Ultra Ethernet Consortium to harden lossless behavior at AI scale. InfiniBand continues to win in the highest-performance training clusters with adaptive routing and offloads. Inside supernodes, proprietary links such as NVIDIA NVLink/NVSwitch dominate scale-up. On the merchant silicon front, Broadcom Tomahawk/Jericho, Marvell Teralynx, and Cisco Silicon One roadmap ports and buffers to 800G today and 1.6T next, while NVIDIA’s Spectrum-X targets Ethernet-optimized AI fabrics. DPUs/SmartNICs (e.g., NVIDIA BlueField, AMD Pensando) are increasingly essential for transport offload and telemetry.
800G today, 1.6T next with new optics and form factors
IEEE 802.3 standards are moving the industry from 100G to 200G per lane, enabling 800G and 1.6T ports across OSFP and QSFP‑DD ecosystems—with OSFP‑XD and 1.6T variants on deck. Active Optical Cables are displacing DACs even intra-rack as reaches grow and thermal budgets tighten. CPO and Near‑Package Optics reduce electrical trace losses and energy per bit, but introduce new challenges in serviceability, yield, and repair. Vendors to watch span switch platforms (Arista, Cisco, Juniper), optical engines and modules (Coherent, Lumentum, InnoLight, Eoptolink), silicon photonics (Intel, Broadcom, Marvell), fiber/connectivity (Corning, Molex), and transport players (Ciena, Infinera) as data center optics intersect metro backbones.
Implications for Telcos, Cloud Providers, and Enterprise IT
The AI fabric does not stop at the rack—transport, peering, and edge nodes must evolve in lockstep to deliver end-to-end performance.
Converging data center and WAN optics for AI workloads
Training datasets, checkpoints, and inference traffic spill into WAN and metro networks, driving upgrades in 400G/800G coherent transport and SRv6/segment routing for deterministic paths. Telcos operating sovereign or federated AI clouds will need tight coordination between leaf‑spine fabrics and optical line systems to maintain latency budgets and meet power envelopes. Edge sites hosting RAG/CV inference benefit from short‑reach optics and efficient top‑of‑rack designs aligned with 5G MEC buildouts.
Supply chain risks and CapEx planning for AI optics
Lead times for 800G optics, next‑gen switch ASICs, and advanced cooling are volatile. Multi‑sourcing pluggables, validating CPO/LPO pilots, and aligning with OCP and UEC guidance can de‑risk rollouts. Expect higher networking dollar content per rack and rising facility costs for power and liquid cooling; budget models must weight port speed transitions and optics mix more heavily than in past refresh cycles.
Execution Risks and Open Questions in AI Optics
The opportunity is large, but technical, operational, and ecosystem hurdles could shift adoption timelines.
CPO maturity, serviceability, and multi-vendor interoperability
While CPO promises lower energy per bit and denser fabrics, field-replaceability, module yield, thermal coupling, and interoperability across multi-vendor environments remain open issues. Standards and MSAs are improving, but many operators will favor pluggables and LPO for near-term flexibility, using CPO selectively in the most demanding supernodes.
Power density and liquid cooling constraints
Optics power per port is rising even as energy per bit falls. Moving to 1.6T, liquid cooling and advanced chassis designs become prerequisites. Facility upgrades—power distribution, heat rejection, and structured cabling—can be on the critical path and should be treated as part of the network program, not an afterthought.
Action Plan for Network Leaders Building AI Fabrics
Pragmatic steps today can preserve optionality while positioning teams to capture AI-driven value in 12–24 months.
Standardize on 800G now; evaluate 1.6T, LPO, and CPO next
For clusters delivering value in the next 12 months, standardize on 800G Ethernet with RoCEv2 and proven congestion control, or use InfiniBand where ultra-low latency warrants it. In parallel, lab-evaluate 1.6T optics, LPO, and CPO/Near‑Package Optics for the next refresh. Design supernodes to be interconnect-agnostic where possible.
Align vendors to silicon and optics roadmaps and multi-source
Down-select platforms around two switch silicon families to ensure supply resilience, and multi-source 800G OSFP/QSFP‑DD optics across at least three vendors. Participate in UEC and OCP workstreams, and engage with module makers on thermal envelopes, reliability data, and RMA models for CPO. Build testbeds for AI traffic profiles, not just RFC benchmarks.
Engineer for AI traffic with visibility, control, and offloads
Deploy ECN/PP marking, dynamic buffer tuning, and per-flow telemetry to keep tail latencies in check. Use DPUs for transport offloads and security isolation among tenants. Instrument job-level metrics (goodput, convergence time) to link network investments directly to AI outcomes and guide iterative upgrades.
Bottom line: AI is pulling optics into the heart of system performance. With a potential $154 billion market by 2028, winners will pair disciplined near-term deployments with aggressive validation of next-wave optical technologies.
