Can Better Interconnects Become a Competitive Advantage?
When people discuss AI infrastructure, the conversation usually starts with GPUs, AI accelerators, memory capacity, and power consumption. Networking is often treated as a supporting component that simply connects these expensive systems. But this is rapidly changing.
Modern AI models are trained across thousands or even hundreds of thousands of accelerators. These accelerators continuously exchange model parameters, gradients, training data, and intermediate results. If the network cannot move this information quickly and reliably, GPUs are forced to wait.
In that scenario, an organization may own one of the most powerful AI clusters in the world, but still fail to use its computing power efficiently.
Microsoft has described this problem as a growing “networking wall.” Memory and networking bottlenecks can reduce GPU utilization, increase training time, consume more energy, and prevent expensive AI infrastructure from reaching its full performance.
This is why new technologies such as Microsoft’s MOSAIC MicroLED optical links and hollow-core fiber are important. They are not simply faster cables. They represent a wider change in how cloud companies may design, connect, and operate future AI datacenters.
The bigger question is: Can next-generation datacenter interconnects become a competitive advantage for AI cloud providers?
Why AI Workloads Put So Much Pressure on Networks
Traditional cloud datacenters usually run millions of relatively independent workloads. One server may host a web application, another may process transactions, while another stores business data.
AI training works differently. A large training job behaves more like one distributed supercomputer. Each GPU processes part of the workload, but it must regularly exchange information with other GPUs. During operations such as gradient synchronization, the system may need to wait until all participating accelerators have completed their work.
If one network link becomes congested, slow, or unavailable, many other GPUs can remain idle while waiting for data.
Microsoft’s Fairwater AI datacenter design uses a flat network that can connect hundreds of thousands of NVIDIA GB200 and GB300 GPUs. Microsoft has also built a dedicated AI wide-area network to connect multiple Fairwater sites and allocate workloads across them.
This shows that the AI infrastructure race is no longer only about how many GPUs a company can purchase. It is also about how effectively those GPUs can communicate.
The Current Copper-versus-Optics Problem
Modern datacenters mainly depend on two physical technologies for high-speed communication: copper cables and optical fiber. Both are useful, but neither provides a perfect solution for the growing demands of AI infrastructure.
Copper connections are generally reliable, energy-efficient, and relatively simple to operate. However, their effective transmission distance becomes shorter as data speeds increase. In many modern high-bandwidth environments, copper links may only work efficiently across distances of around two metres or less. This makes them suitable for connections within a rack, but less practical when accelerators need to communicate across multiple racks.
Traditional optical connections solve the distance problem because they can carry high-speed data across tens of metres or more. This makes them useful for connecting servers, switches, and accelerator systems across a larger datacenter area. However, conventional optical links usually depend on lasers, digital signal processors, error-correction systems, and other electronic components. These additional components increase power consumption, hardware cost, heat generation, and operational complexity.
Reliability is another concern. According to Microsoft’s research, conventional optical links may experience higher failure rates than passive copper connections. In a normal cloud workload, a failed link may affect only a limited number of requests. In a large AI training cluster, however, one network failure can interrupt communication between many GPUs and potentially delay a job that has already been running for several days or weeks.
This leaves AI datacenter designers with a difficult trade-off. Copper offers low power consumption and strong reliability, but its short reach forces GPUs and networking equipment to be placed very close together. This can create extremely dense racks with major power, cooling, and space requirements.
Optical networking provides the longer reach required to connect equipment across racks, but it can introduce higher power consumption, additional components, greater cost, and more maintenance requirements. Hollow-core fiber may improve longer-distance communication by reducing propagation latency, but it also requires new manufacturing, installation, splicing, and testing processes.
The challenge is therefore not simply choosing between copper and optics. Datacenter operators need an interconnect that can provide optical-level distance while maintaining the energy efficiency and reliability normally associated with copper.
Microsoft’s MOSAIC project is designed to address this exact problem. By using hundreds of parallel MicroLED channels instead of a small number of extremely fast laser-based channels, MOSAIC aims to offer longer-distance optical communication with lower power consumption, simpler electronics, and greater resilience. Although the technology is still moving from research prototypes toward large-scale production, it could eventually remove one of the most difficult infrastructure trade-offs in AI datacenter design.
What Is MOSAIC MicroLED Networking?
MOSAIC is an experimental optical interconnect developed through collaboration between Microsoft Research, Azure, Microsoft 365, and hardware suppliers.
Most existing high-speed interconnects use a “narrow-and-fast” architecture. For example, an 800 Gbps connection may use eight individual lanes, with each lane transmitting at 100 Gbps.
MOSAIC takes the opposite approach. Instead of using a small number of extremely fast channels, it uses hundreds of lower-speed channels operating in parallel. Microsoft calls this a wide-and-slow architecture.
Each channel uses a tiny MicroLED as a light source. These MicroLEDs send data through a multicore imaging fiber to an array of CMOS sensors at the receiving end.
MicroLED technology was originally developed for high-resolution displays such as smartwatches and head-mounted devices. Because MicroLEDs are extremely small, hundreds of them can fit into a very small physical area. They can also be directly switched on and off at several gigabits per second without requiring the same complex laser electronics used in conventional optical transceivers.
Imagine that a traditional optical cable is a motorway with eight extremely fast lanes. MOSAIC is closer to a motorway with hundreds of slower lanes carrying data at the same time. The individual lane is slower, but the total combined bandwidth can still be extremely high.
What Microsoft Has Demonstrated
Microsoft’s MOSAIC research prototype uses 100 optical channels operating in parallel. During testing, each channel successfully transmitted data at 2 Gbps over a distance of 20 metres. The prototype also achieved 1.6 Gbps per channel over 30 metres, which shows that the wide-and-slow MicroLED approach can support useful datacenter distances without depending on a small number of extremely fast laser-based channels.
Based on these results, Microsoft researchers have outlined a potential path toward an 800 Gbps pluggable transceiver. Such a design would use hundreds of MicroLED channels working together to produce the required total bandwidth. The same architecture could potentially scale further to support future 1.6 Tbps and 3.2 Tbps interconnects as MicroLED performance, sensor technology, packaging, and manufacturing processes improve.
However, it is important to separate what has already been demonstrated from what is still projected. The physical research prototype has shown successful communication over distances of up to 30 metres. The frequently discussed 50-metre reach and complete 800 Gbps configuration are based partly on simulations, system modelling, expected component improvements, and more advanced packaging. They do not yet represent a mass-produced 800 Gbps MOSAIC cable available for commercial deployment.
Microsoft estimates that a production-ready MOSAIC link could consume up to 68% less power than mainstream optical links. This could save more than 10 watts for every cable used inside a datacenter. At the scale of a large AI cluster containing thousands of connections, these savings could reduce both electricity consumption and the amount of heat that cooling systems need to remove.
The proposed system could also reach distances of up to 50 metres while providing much greater reliability than conventional optical links. Microsoft estimates that the use of redundant MicroLED channels could make MOSAIC links up to 100 times more reliable in certain configurations. If one optical channel becomes faulty, the system could move the data to an unused spare channel rather than treating the complete cable as failed.
Another important advantage is protocol compatibility. MOSAIC is designed to transport technologies such as Ethernet, InfiniBand, PCIe, and CXL without inspecting or terminating the traffic. This means it could potentially be introduced into existing and future datacenter architectures without requiring cloud providers to redesign every networking protocol around the new cable.
These power, reach, and reliability figures should still be treated as research estimates rather than guaranteed specifications of a commercially available product. Microsoft has stated that it is working with hardware suppliers to productize MOSAIC and prepare the technology for mass manufacturing. Its real impact will therefore depend on whether the design can maintain the same benefits when produced and deployed at datacenter scale.
Why MOSAIC Could Change Datacenter Design
The possible value of MOSAIC goes beyond replacing one type of network cable with another. Its greater impact could come from giving infrastructure architects more freedom in how they arrange AI accelerators, racks, power systems, and cooling equipment inside a datacenter.
Copper connections provide strong reliability and relatively low power consumption, but their limited reach forces AI accelerators to be positioned extremely close together. As accelerator density increases, this can result in racks that consume hundreds of kilowatts of power. Such designs are difficult to cool, maintain, upgrade, and operate safely.
A low-power optical connection with copper-like reliability could allow GPUs and other AI accelerators to be distributed across several racks while still participating in the same high-bandwidth scale-up network. The accelerators would continue to operate as part of one large computing system, but they would no longer need to be packed into the smallest possible physical space.
This additional distance could give datacenter architects more flexibility when deciding where racks should be placed and how cooling systems should be arranged. Power distribution equipment could be positioned more effectively, and high-density compute components could be separated to reduce concentrated heat generation.
The design could also improve fault isolation. Instead of placing a large number of critical accelerators inside one highly dense rack, operators could distribute them across several physical zones. A power, cooling, or hardware problem affecting one rack would therefore be less likely to disturb the entire accelerator group.
Greater physical flexibility may also simplify future hardware upgrades. New accelerator generations often require different power, cooling, and networking configurations. If the interconnect can cover longer distances, operators may be able to replace or expand individual racks without redesigning the complete scale-up network.
Network topology and datacenter floor planning could become more flexible as well. Architects would have more options for connecting switches, servers, accelerators, and storage systems while still maintaining the high-speed communication required by distributed AI workloads.
Instead of forcing every GPU into one extremely dense rack, MOSAIC could make it possible to spread computing equipment over a larger physical area without losing the benefits of tightly connected accelerator communication. This may reduce mechanical, electrical, and cooling constraints while also allowing cloud providers to build larger accelerator domains.
In this way, MOSAIC could influence not only networking performance, but the overall physical architecture of future AI datacenters.
Hollow-Core Fiber: Moving Light Through Air
MOSAIC is mainly designed for shorter connections inside datacenters. Hollow-core fiber addresses a different part of the networking problem.
Traditional optical fiber sends light through a solid glass core. Hollow-core fiber, or HCF, guides most of the light through an air-filled central region surrounded by carefully designed glass structures.
Light travels faster through air than through glass. Microsoft says this allows HCF to provide up to 47% faster signal propagation than conventional silica fiber. Microsoft has also described this as approximately 33% lower propagation latency for the same physical distance.
This does not mean every Azure application automatically becomes 33% faster. Application latency also depends on switches, routing, software, storage, congestion, processing time, and many other components.
However, for long fiber routes where propagation delay represents an important part of total latency, the improvement can be valuable.
Microsoft also says HCF can extend transmission distance by up to 1.5 times while remaining within a similar latency envelope to standard single-mode fiber. This could allow datacenters to serve customers over a wider geographical area without creating the same latency penalty.
HCF Is Already Carrying Azure Traffic
Unlike MOSAIC, which is still moving from research into product development, hollow-core fiber has already been deployed inside Azure’s operational network.
Microsoft reports that HCF links are carrying live customer traffic across multiple Azure regions. In one publicly discussed deployment, two Azure datacenters were connected through two metropolitan network routes, with each route extending for more than 20 kilometres.
However, deploying hollow-core fiber required much more than replacing standard fiber with a new cable. Microsoft and its technology partners had to create an entire operational ecosystem around the technology. This included outdoor-grade HCF cables that could survive real environmental conditions, new cable-joint enclosures, specialized patch tails, and field-splicing techniques suitable for hollow-core structures.
New testing and fault-location equipment was also required because existing tools and repair processes were originally designed for conventional solid-glass fiber. Engineers additionally had to ensure that HCF could integrate with Azure’s existing Dense Wavelength Division Multiplexing, or DWDM, systems without requiring a complete redesign of the surrounding network infrastructure.
The deployed links support commercially available transponders and multi-terabit-per-second data capacity. During Microsoft’s published evaluation period, the HCF routes handled changing volumes of customer traffic without reported link outages or link flaps. This is important because a networking technology may perform well in laboratory testing but still struggle when exposed to real traffic patterns, outdoor conditions, maintenance activity, and operational failures.
The deployment therefore demonstrates that hollow-core fiber is moving beyond an experimental stage. It has already shown that it can operate as part of a production cloud network while remaining compatible with existing optical networking equipment.
Microsoft has also entered manufacturing partnerships to increase HCF production capacity. These collaborations are intended to improve manufacturing scale, reduce deployment barriers, and support the wider use of hollow-core fiber across more Azure datacenters and regional network routes.
Its long-term impact will depend on whether production costs, installation processes, repair methods, and supply chains can scale globally. However, compared with MOSAIC, hollow-core fiber is already further along the path from technical research to real infrastructure use.
MOSAIC and HCF Are Complementary Technologies
MOSAIC and hollow-core fiber should not be viewed as competing technologies because they are designed to solve networking problems at different physical distances.
MOSAIC is primarily intended for shorter connections inside the datacenter. These connections may exist between GPUs, servers, network switches, and racks. Its purpose is to deliver high bandwidth across distances of tens of metres while reducing the power consumption, complexity, and reliability concerns associated with conventional optical transceivers.
Hollow-core fiber is more suitable for longer-distance communication. It can be used to connect different buildings, datacenters, availability zones, metropolitan network locations, and regional cloud facilities. Its main advantage comes from allowing light to travel mostly through air rather than solid glass, which can reduce propagation latency over longer routes.
This means MOSAIC could improve communication inside an individual AI datacenter, while HCF could improve communication between datacenters and regional network facilities. One technology focuses on short-reach, high-density AI infrastructure, while the other focuses on faster transmission across wider geographical distances.
Together, these technologies could help Microsoft create a layered optical infrastructure that begins at the GPU level and extends across the wider Azure network. Data could move from one accelerator to another, then across racks, throughout the datacenter, between Azure regions, and eventually across connected AI datacenters.
At the local level, MOSAIC could help GPUs communicate with lower power usage and stronger reliability. At the regional level, hollow-core fiber could reduce propagation delay and allow datacenters to be connected over longer distances without introducing the same latency as conventional fiber.
This layered approach could optimize each section of the network according to its specific requirements. Short-distance connections could focus on bandwidth density, energy efficiency, and fault tolerance, while long-distance connections could focus on propagation speed, geographical reach, and compatibility with existing optical transport systems.
The result could be an AI infrastructure where networking is optimized from the accelerator itself to the wider cloud region. This would help reduce communication delays, improve GPU utilization, support larger distributed AI clusters, and make it easier to operate AI workloads across multiple connected datacenters.
How Networking Becomes a Cloud Competitive Advantage
Customers may not choose a cloud provider because it uses MicroLED cables or hollow-core fiber. However, they will notice the results through faster AI training, lower costs, improved reliability, and better service performance.
GPUs are among the most expensive resources in an AI datacenter. When network delays force accelerators to wait for data, both computing capacity and electricity are wasted. A faster and more predictable network keeps GPUs active for longer, allowing more productive work to be completed with the same hardware. For customers, this can mean shorter training times and lower infrastructure costs. For cloud providers, it means better utilization and more value from every accelerator.
Networking improvements can also reduce model training time. Large AI workloads repeatedly exchange and synchronize data across thousands of accelerators. Even small delays can become significant when repeated millions of times. Microsoft’s AI WAN and connected Fairwater datacenters are designed to support large training jobs across multiple locations as one connected AI infrastructure.
Energy efficiency is another important advantage. Networking equipment consumes electricity and generates heat, which increases cooling requirements. If MOSAIC achieves its projected power savings at scale, cloud providers could reduce networking costs and use more of their available power capacity for GPUs and other accelerators.
Reliability also directly affects AI performance. Large training jobs may run for several days or weeks, and a failed optical link can interrupt the workload or force it to restart from a previous checkpoint. MOSAIC’s use of spare MicroLED channels could allow traffic to move away from a failed channel without replacing the complete cable, resulting in fewer interruptions and more predictable completion times.
Hollow-core fiber may also give cloud providers greater flexibility when selecting datacenter locations. Facilities could potentially be positioned farther apart while still meeting latency requirements. This allows providers to consider power availability, renewable energy, land cost, water usage, customer location, and data-residency requirements without facing the same networking limitations.
The real competitive advantage therefore does not come from one cable or fiber technology alone. It comes from combining networking, accelerator systems, cooling, routing, workload scheduling, manufacturing partnerships, and datacenter design into one optimized platform. Cloud providers that control and improve the full infrastructure stack may be able to deliver AI services faster, more reliably, and at a lower cost than competitors.
The Wider Industry Is Moving in the Same Direction
Microsoft is not the only company treating networking as a strategic AI technology.
NVIDIA has announced Spectrum-X and Quantum-X switches using co-packaged silicon photonics. NVIDIA claims its photonics designs can provide 1.6 Tbps per port, significant energy savings, and greater networking resilience for large AI factories.
Google’s TPU infrastructure combines custom accelerator interconnects, optical circuit switching, datacenter networking, and specialized topologies. Google describes TPU Pods as large groups of accelerators connected over purpose-built high-speed networks.
AWS uses Elastic Fabric Adapter networking and petabit-scale non-blocking networks in EC2 UltraClusters to connect thousands of accelerated instances.
These investments show that networking is becoming one of the main areas of differentiation among AI infrastructure providers.
Can Datacenter Interconnects Really Become a Competitive Advantage?
Yes, but not because customers care about the cables themselves. The real advantage appears when better interconnects improve training speed, inference latency, GPU availability, cluster reliability, energy efficiency, and the overall cost of running AI workloads.
MOSAIC could give Microsoft more flexibility inside AI datacenters by offering longer optical reach with lower projected power consumption and improved reliability. This could help cloud operators keep GPUs active for longer, reduce networking energy use, and design less physically constrained accelerator clusters.
Hollow-core fiber could provide a similar advantage across longer distances. By reducing propagation latency and extending the useful range of optical connections, it may improve communication between datacenters, Azure regions, and large distributed AI systems.
These technologies become more valuable when combined with Microsoft’s wider infrastructure, including Fairwater datacenters, AI WAN, workload scheduling, custom silicon, and its global fiber network. Together, they could improve how quickly and efficiently Microsoft deploys and operates AI capacity.
However, MOSAIC is still a developing technology rather than a finished replacement for existing optical systems. Hollow-core fiber is already being used in Azure, but its wider impact will depend on manufacturing scale, installation cost, long-term reliability, and global deployment.
The competitive advantage will therefore come not from a single cable, but from how effectively the entire networking and computing stack works together.
References
- Microsoft Research, Breaking the Networking Wall in AI Infrastructure.
- Microsoft Research, MOSAIC: Breaking the Optics versus Copper Trade-off with a Wide-and-Slow Architecture and MicroLEDs.
- Microsoft Azure, How Hollow Core Fiber Is Accelerating AI.
- Microsoft Azure Networking, The Deployment of Hollow Core Fiber in Azure’s Network.
- Microsoft, Infinite Scale: The Architecture Behind the Azure AI Superfactory.
- Microsoft, From Wisconsin to Atlanta: Microsoft Connects Datacenters to Build Its First AI Superfactory.
- NVIDIA, Spectrum-X Photonics and Co-Packaged Optics Networking Switches.
- Google Cloud, Tensor Processing Units and TPU System Architecture.
- Amazon Web Services, Amazon EC2 UltraClusters.
