Optical compute interconnect is now the explicit target of one of the most significant industry alliances formed in recent AI infrastructure history. Announced on March 12–13, 2026, the Optical Compute Interconnect Multi-Source Agreement — OCI MSA — brings together AMD, Nvidia, Broadcom, Meta, Microsoft, and OpenAI with a single goal: replace copper-based connectivity inside AI clusters with a standardised, multi-vendor optical alternative capable of reaching speeds up to 3.2 Tb/s. This is not a research project or a whitepaper exercise. It is a formal specification effort with founding members who collectively control the hardware and software that defines modern AI infrastructure.
Why optical compute interconnect is the defining AI infrastructure challenge right now
Copper interconnects have served data centres well for decades, but the demands of large-scale AI training and inference have exposed their fundamental limits. Bandwidth density, power efficiency, and scalability all hit a wall with copper at the speeds modern AI clusters require. The OCI MSA frames optical as the necessary successor, targeting a standardised optical PHY — the physical layer that governs how data moves between chips and systems — that can support the full AI rack supply chain through an open, multi-vendor ecosystem.
The use of NRZ modulation in the OCI specification is a deliberate technical choice. NRZ, or non-return-to-zero signalling, is a proven modulation format that prioritises simplicity and reliability at scale. For an alliance trying to build industry-wide consensus across competing vendors, choosing an established modulation scheme over more exotic alternatives reduces friction and accelerates adoption. Whether NRZ can sustain the full 3.2 Tb/s ambition across all deployment scenarios remains to be demonstrated in practice.
What the OCI MSA actually delivers — and what it does not yet
The OCI MSA’s core output is a scalable open specification roadmap for optical scale-up interconnects. Scale-up connectivity refers to the high-speed links between accelerators within the same system or rack, as opposed to scale-out links that connect separate servers across a network. This is the performance-critical tier where copper limitations are felt most acutely in AI workloads, and it is precisely where the alliance is focused.
It is worth being direct about what the OCI MSA has not yet delivered. As of its March 2026 announcement, this is a specification alliance, not a shipping product. The 3.2 Tb/s target is a roadmap aspiration, not a validated deployment figure. The claims around solving AI data centre bottlenecks are forward-looking, and the gap between a multi-vendor specification and a production-grade optical interconnect ecosystem is real. That gap does not make the initiative unimportant — it makes the execution phase the only thing that matters from here.
How the OCI MSA compares to the Open CPX effort
The OCI MSA is not the only optical interconnect initiative announced in this period. A separate but complementary alliance, the Open Co-packaging Multi-Source Agreement — Open CPX MSA — was formed by a different set of companies: Ciena, Coherent, Marvell, Molex, Samtec, and TeraHop. Where the OCI MSA targets optical scale-up interconnects at the rack level, the Open CPX MSA focuses on co-packaged and near-package optical engines, aiming to reduce power consumption, cost, and latency while increasing bandwidth density and reliability at the chip-to-optical interface.
The two efforts are addressing different points in the signal chain. Co-packaged optics, the Open CPX focus, integrate the optical engine much closer to the chip die itself, which reduces the electrical distance data must travel before hitting fibre. The OCI MSA’s scale-up focus operates at a higher system level. Together, they suggest the industry is attacking the copper problem from multiple directions simultaneously, which is either a sign of genuine urgency or of the fragmentation that tends to slow standards adoption. Possibly both.
Is the OCI MSA the alliance that will actually replace copper in AI clusters?
The founding membership is as credible as it gets. Nvidia and AMD together define the GPU landscape for AI training. Broadcom is the dominant force in custom AI silicon and networking ASICs. Meta, Microsoft, and OpenAI represent the hyperscale operators who are both the largest buyers of AI infrastructure and the companies with the most direct pain from current interconnect limitations. When the organisations funding and building AI at scale agree that optical compute interconnect needs a common standard, the industry listens.
That said, the history of multi-vendor standards efforts in data centre networking is a history of slow timelines and competing implementations. The OCI MSA will need to move from specification to silicon faster than most comparable initiatives if it wants to shape the current AI buildout cycle rather than the next one.
What does the OCI MSA mean for AI data centre buyers?
For organisations building or expanding AI infrastructure today, the OCI MSA signals that optical interconnect is becoming the expected direction for scale-up connectivity, not a premium niche. A multi-vendor open specification, if it delivers, means buyers will eventually have competitive sourcing options rather than being locked into a single vendor’s proprietary optical solution. That competitive dynamic is what drives cost down and reliability up over time.
When will optical interconnects replace copper in AI clusters?
The OCI MSA announced its specification roadmap in March 2026, but no shipping timeline for compliant products has been stated. Standards bodies typically require time for specification finalisation, silicon development, and interoperability testing before products reach the market. The Open CPX MSA, targeting co-packaged optics from a different vendor set, is on a parallel but distinct timeline.
The OCI MSA is a genuine inflection point in how the AI industry thinks about interconnect infrastructure — but it is an inflection point at the beginning of a long road, not the end of one. The alliance has the right names on the founding document. Now it needs to ship.
Edited by the All Things Geek team.
Source: Tom's Hardware
