Data Center Interconnect (DCI) Is Now a Four-Layer Stack with Open Standards Advancing, but NVIDIA Is Setting the Pace

Artificial Intelligence (AI) Data Center Interconnect (DCI) is quickly becoming a considerable bottleneck in new data center rollouts, in some cases more than Graphics Processing Units (GPUs). This market is now separated in four distinct layers in terms of innovation and reach:

1. Die-to-die interconnect, also referred to as scale-within. The emerging integration of chiplets and the Universal Chiplet Interconnect Express (UCIe) 3.0 aim for chip-level disaggregation and there are multiple new optical technologies being developed, including Celestial AI and Lightmatter.
2. Scale-up refers to the connections within a rack or pod where NVIDIA’s NVLink dominates the market, which is now transitioning from copper to optical. There is little competition for NVIDIA here as it’s a captive market; NVIDIA GPUs dominate AI workloads and are coupled with NVLink.
3. Scale-out connects different racks within a cluster and NVIDIA’s InfiniBand has dominated the market, except for Tier One hyperscalers, which developed their own systems, including Amazon Web Services (AWS) proprietary stack and Google Virgo. Co-packaged optics are aiming to replace coherent pluggables as the next evolutionary transition.
4. Scale-across is currently using repurposed long-haul optical transport links to connect geographically distributed data centers, within tens of thousands of fiber strands connecting them. The optical market is developing multiple innovations to improve this domain, including hyper-rail platforms that reduce the need for inline amplification.

The throughput and capacity decreases from die-to-die to scale-across, while the link distance increases. There are structural and market differences across these four layers, but Broadcom and NVIDIA are now controlling multiple layers.

Although each of the four layers above is rapidly developing, Broadcom and NVIDIA have a strong presence in multiple layers due to their long-term strategic planning. For example, NVIDIA has NVLink for scale-up, InfiniBand and Spectrum-X for scale-out, and, of course, it has full control for its die-to-die interconnect (e.g., dual reticle dies for Rubin and other GPUs). Broadcom has developed a competitive Serializer/Deserializer (SerDes) for scale-up and scale-out, and has its Scale-Up Ethernet (SUE) fabric and switches for scale-up and its Jericho switch for scale-out. As the market is now dominated by Tier One AI hyperscalers that design their own systems, having expertise and consistent behavior across multiple layers is a formidable competitive differentiation, something that smaller vendors cannot compete against.

The second major implication is that all these innovations and market developments are secondary to manufacturing and execution. For example, the transition to Co-Packaged Optics (CPO) in scale-up will not likely reach large-scale until mid-2027 or even 2028, and new entrants will likely delay their entry into the market due to manufacturing and component constraints, including memory. For example, Arista Networks announced its entry into the scale-up market using Ethernet for Scale-Up Networking (ESUN), but the company also stated that it will come in 2027 or later. Even UCIe for chiplet interconnect, which was specified 4 years ago, experienced its first cross-vendor proof point this year. So, the binding constraint this year is the production of reliable optical components at volume, and this sits largely in Asian supply chains.

Scale-out optical and CPO are near-term transitions and are generating revenue today. This will likely continue in the next few years as the copper-to-optical migration reaches an inflection point. Scale-across DCI is in a steep growth phase driven by campus distribution and scale-up optical is a 2027/2028 transition still mostly on roadmaps. Finally, die-to-die optical interconnect is the longest-dated of the four and will likely take the most time due to NVIDIA essentially having a captive market in the GPU space. For smaller data center operators, scale-within and scale-up are essentially beyond their control, but there are choices to be made for scale-out and scale-across.

For scale-out, Ethernet is becoming a viable alternative to NVIDIA’s InfiniBand and Spectrum-X, but only if the data center operator can afford the implementation and optimization effort with Broadcom’s Tomahawk switch and its Thor Network Interface Controller (NIC). This requires optimization in line with the NVIDIA Collective Communications Library (NCCL) to ensure networking does not become a bottleneck for processing. Otherwise, NVIDIA’s products are considered the gold—albeit, more expensive—standard. For scale-across, smaller operators can lease dark fiber capacity, rather than build out their own network, bypassing the fiber scarcity and current market challenges. Once they reach scale, their scale-across network build-out can develop in parallel with their infrastructure expansion strategies.

Most of the developments covered above apply to Tier One hyperscalers, which can set an aggressive strategy and implementation plans, and can reach a large enough DCI scale to even drive vendor roadmaps. For the rest of the market, and the myriad other smaller companies deploying AI infrastructure space, identifying how to address each networking layer is its own challenge. Smaller data center operators will not be able to match the aggressive deployment and innovation strategies of hyperscalers, and they will also not benefit immediately from key innovations in the market (e.g., chiplet interconnects), but they need to be mindful of silicon developments as these innovations will likely influence long-term roadmap decisions that happen today.

As of 2026, the market is dominated by NVIDIA and this will likely remain in the coming years, as NVIDIA’s GPU roadmap continues, with networking evolving alongside. Tier One hyperscalers are the only ones that can design and implement their networking stack, and many of these will likely release their innovations in the open source community, much like OpenAI and others did with Multipath Reliable Connection (MRC). However, smaller data center operators will likely not be able to utilize these innovations as their scale may not require it or networking products for these innovations will simply not be available. The market must now congregate around open projects, including UALink and the Ultra Ethernet Consortium (UEC), and push for more vendor choice and an open supply chain. Only then will they be able to deploy an open data center networking stack, but this is much easier said than done, especially with NVIDIA’s dominance in the GPU market.