With generative AI driving exponential growth in computing power demand, the "last mile" within data centers—data transfer between servers—has become a critical bottleneck for performance. (Microsoft Research)AnnounceSignificant breakthroughs have been achieved in the development of next-generation Active Optical Cable (AOC) technology in collaboration with MediaTek and multiple suppliers, with commercialization expected by the end of 2027 in partnership with industry partners. This revolutionary design, based on MicroLED light sources, is expected to reduce power consumption for data center interconnects by approximately 50% while achieving reliability comparable to copper cables, paving a new path for the deployment of large-scale AI training clusters.
The "Digital Water Pipe" Revolution Amid the AI Craze: How Will MicroLED Disrupt the Existing Architecture?
As AI models continue to grow in scale, the number of graphics processing units (GPUs) that need to be connected within data centers often reaches tens of thousands. Traditional transmission solutions are facing severe physical limitations: although copper cables are energy-efficient and reliable, their transmission distance is limited to about 2 meters, and they can only be used for interconnection within a single rack; laser-driven optical fibers can transmit from tens of meters to several kilometers, but their high power consumption, temperature sensitivity, and susceptibility to dust make their failure rate up to 100 times that of copper cables, becoming a pain point for large-scale cluster maintenance.
The MicroLED solution developed by Microsoft's Cambridge Research Lab in collaboration with the Azure core team, Azure Hardware Systems and Infrastructure, and the Microsoft 365 team aims to solve the "impossible triangle" of distance, power consumption, and reliability. Paolo Costa, the project's lead researcher, explained that the core concept of the new technology is to replace the traditional "narrow and fast" approach with a "wide and slow" one.
Traditional laser fiber optic cables transmit data via pulses at extremely high speeds through a few channels, much like a narrow but turbulent stream. MicroLED systems, on the other hand, utilize inexpensive and commercially available MicroLED components, paired with special cables called "imaging fibers"—optical fibers originally used in medical endoscopes that contain thousands of independent cores—to create thousands of parallel channels, transmitting data simultaneously at a lower speed. Paolo Costa likened it to "a wide, slow-moving river that, compared to a narrow, turbulent stream, ultimately carries the same volume of water."
Half the energy consumption, reliability comparable to copper cables: Five major technological breakthroughs of MicroLED.
According to laboratory tests and simulations conducted by Microsoft's research team, if this technology were actually deployed, energy consumption could be reduced by approximately 50% compared to current mainstream laser fiber solutions. This energy-saving benefit comes from two aspects: firstly, MicroLED itself is a high-efficiency light source; secondly, the new architecture eliminates the need for the complex and power-consuming digital signal processor (DSP) found in traditional optical cables.
Doug Burger, Microsoft's Senior Vice President and Technical Fellow, pointed out: "In the early days, transmitting data with LEDs at a lower cost and with lower power consumption sounded like a pipe dream. This breakthrough has the potential to fundamentally change almost every aspect of computing infrastructure."
This collaboration with MediaTek and other suppliersCollaborative development technologyThey have successfully miniaturized complex optical components from a laboratory workbench and integrated them into a metal transceiver about the size of an adult's thumb, which can be directly plugged into existing data center servers. Key breakthroughs include:
• Energy saving and power saving:By eliminating the need for a DSP, power consumption is reduced by up to 50% compared to traditional VCSEL (Vertical Cavity Surface-Emitting Laser) active optical cables.
• Copper cable reliability rating:MicroLEDs are simple in structure, durable, and insensitive to temperature, achieving stability far exceeding that of lasers.
• Extended transmission distance:It can reach tens of meters, meeting the needs of cross-rack interconnection of AI training clusters.
• Scalability:The bandwidth can be vertically expanded by increasing the number of optical channels or increasing the rate of a single channel.
• Advanced packaging integration:By employing single-crystal CMOS wafer integration and heterogeneous bonding technology, MicroLED arrays are directly bonded to CMOS wafers, achieving extremely small pitch and ultra-high density channel arrays.
Vince Hu, Vice President of MediaTek, said that this collaboration combines the deep technical capabilities and industry insights of both parties to successfully miniaturize MicroLED technology and integrate it into transceivers compatible with existing data center equipment, allowing the industry to seamlessly adopt this new technology.
Beyond MicroLED: Hollow fiber optic cables have been deployed in Azure, increasing speeds by 47%.
It's worth noting that MicroLED is not Microsoft's only foray into data center networking. Another technology called "Hollow Core Fiber" (HCF) is currently being used in some Microsoft Azure regions and is being deployed in more regions around the world.
Unlike traditional optical fibers that transmit light signals through solid glass, HCF signals travel through air within a hollow core, enabling faster light speeds. According to published research, HCF can increase data transmission speeds by up to 47% and reduce latency by approximately 33% compared to traditional single-mode fiber (SMF). This means data centers can be located at greater distances without sacrificing the speed and responsiveness users are accustomed to. This technology, developed by the University of Southampton and further refined by Lumenisity, a company acquired by Microsoft in 2022, has been named one of Time magazine's best inventions of 2025.
Frank Rey, General Manager of Azure Hyperscale Networking, pointed out that MicroLED and HCF are complementary technologies. MicroLED will be mainly used inside data centers to connect servers and GPUs; while HCF has the ability to cover long distances, serve customers, connect data centers, and may also play a role inside data centers in the future.
Analysis: The "Invisible War" of AI Infrastructure – Microsoft Shifts from Computational Expansion to Efficiency Optimization
Microsoft's announcement of two network technology breakthroughs, MicroLED and HCF, reflects a strategic intent far deeper than a simple technological upgrade.
First, the AI computing power race has entered the deep waters of "efficiency per watt." While GPU computing power is growing several times over each year, data centers are facing physical limits in terms of power supply and heat dissipation. Microsoft's emphasis on halving the energy consumption of MicroLEDs directly addresses the most pressing cost pain point for operators—power consumption. In training clusters of mega-parameter models, even a slight reduction in power consumption can translate into millions of dollars in operating cost savings.
Secondly, "reliability" will become an invisible barrier for large-scale AI clusters. The report specifically mentions that the failure rate of laser fiber optic cables is up to 100 times that of copper cables, which is an unbearable burden in AI training at the 10,000 kilowatt level. A single training task may take weeks, and any single point of failure may cause the task to be interrupted or even restarted from scratch. MicroLED achieves "copper-level reliability," which means that the "mean time between failures" of large-scale clusters will be greatly improved. For large AI laboratories and cloud service providers that pursue training stability, this value even exceeds the simple bandwidth specification.
Furthermore, MediaTek's key role signifies the increased influence of the Asian semiconductor supply chain in AI infrastructure. Traditionally, the data center optical communication market has been dominated by US and Japanese manufacturers. MediaTek not only participated in miniaturization engineering but also integrated complete electronic functions such as MicroLED drivers and transimpedance amplifiers into a single-crystal CMOS chip, demonstrating its technological strength in high-speed transmission and advanced packaging. This also means that the future standards for AI data centers will be jointly defined by "hyperscale cloud providers" and "Asian chip design companies."
Finally, there's a strategic shift from "increasing computing power" to "optimizing transmission." Doug Burger, Technical Fellow at Microsoft Research, highlights the key point: "This breakthrough has the potential to fundamentally change almost every aspect of computing infrastructure." For the past few years, the focus of AI hardware has been almost entirely on GPU computing power. Now, with Moore's Law slowing, the marginal benefits of simply pursuing the performance of computing units are diminishing. Instead, "system-level optimization"—how to make thousands of GPUs work together efficiently—has become the new deciding factor. Microsoft is simultaneously betting on MicroLED (short-range high reliability) and HCF (long-range low latency), essentially creating a comprehensive "high-speed network three-dimensional warfare" solution for AI data centers.
Overall, this technology, slated for commercialization in 2027, is not only a milestone in Microsoft's internal R&D but also foreshadows a shift in the competition for AI infrastructure from a "computing power arms race" to an "invisible war of transmission efficiency." For cloud giants and enterprises planning their next-generation data centers, whoever achieves the best balance between power consumption, distance, and reliability will occupy a more advantageous position in the computing landscape of the AI era.
