Microsoft’s lab in Cambridge has a simple pitch: swap expensive, power-hungry lasers for mass-produced MicroLEDs and pair them with imaging fiber to carry thousands of parallel light channels, and you can radically cut the energy, cost and fragility of the short-range links that glue modern AI datacenters together.
Datacenter networking has always been a balancing act between three hard constraints: reach, power and reliability. Copper cabling (short-reach electrical) is cheap and energy-efficient for a few meters but can’t scale beyond rack distances at high bandwidth. Conventional single-mode optical fiber (laser-driven) spans long distances with high speed, but lasers and high-speed optics are power-hungry, sensitive to temperature and alignment, and expensive when you push density and redundancy to the levels modern AI workloads demand. MOSAIC — Microsoft Research’s Wide-and-Slow MicroLED architecture — reframes that trade-off by using many low-speed optical channels in parallel rather than a small number of ultra-fast serial lanes.
At the same time, Microsoft’s work on Hollow Core Fiber (HCF) — acquired through Lumenisity in 2022 and now being scaled with manufacturing partners — attacks the pistance side: routing light mostly through air rather than glass to reduce propagation delay and, in the lab, dramatically lower signal loss and latency compared with standard silica single-mode fiber. Together, MOSAIC and HCF represent complementary changes in what Microsoft’s engineers call the “digital plumbing” inside and between datacenters.
Key technical points from the MOSAIC demonstration:
Microsoft has already taken steps to industrialize HCF: the Lumenisity acquisition (December 2022) provided the technical foundation, and Microsoft’s public materials show active manufacturing collaboer makers to scale production for Azure deployments. The implication is clear: Microsoft expects to pair MOSAIC-style short-range links inside facilities with HCF across campus-to-campus and region-to-region backbones.
A possible staged rollout:
But the transition is not risk-free. The devil is in production engineering: MicroLED linewidths, optomechanical coupling into imaging fiber, connector ecosystems, and scaled manufacturing all present honest obstacles. Hollow Core Fiber’s dramatic latency and attenuation metrics are real, but scaling HCF to global fiber plant lengths with consistent performance demands large capital investment and process control; Microsoft’s collaborations with established fiber manufacturers mitigate but do not eliminate that risk.
In short: MOSAIC plus HCF is a credible and potentially transformative systems play for AI‑first datacenters. The physics underpinning both ideas is sound, the prototypes and partnerships are real, and Microsoft is investing in manufacturing scale and supplier collaboration. Adoption will depend on how quickly engineering teams convert prototype gains into low-cost, field-serviceable modules and how rapidly the fiber industry standardizes connectors and splicing for imaging and hollow-core cables. Treat the published figures — whether 50% or up to 68% power reductions for MicroLED links, or 30–47% latency improvements for HCF — as strong technical signals backed by peer-reviewed work and lab/industry reports, but allow for the usual caveats around production maturity and supply-chain timing.
Source: Microsoft Source Using inexpensive MicroLEDs, Microsoft networking innovation aims to make datacenters more efficient
Background / Overview
Datacenter networking has always been a balancing act between three hard constraints: reach, power and reliability. Copper cabling (short-reach electrical) is cheap and energy-efficient for a few meters but can’t scale beyond rack distances at high bandwidth. Conventional single-mode optical fiber (laser-driven) spans long distances with high speed, but lasers and high-speed optics are power-hungry, sensitive to temperature and alignment, and expensive when you push density and redundancy to the levels modern AI workloads demand. MOSAIC — Microsoft Research’s Wide-and-Slow MicroLED architecture — reframes that trade-off by using many low-speed optical channels in parallel rather than a small number of ultra-fast serial lanes.At the same time, Microsoft’s work on Hollow Core Fiber (HCF) — acquired through Lumenisity in 2022 and now being scaled with manufacturing partners — attacks the pistance side: routing light mostly through air rather than glass to reduce propagation delay and, in the lab, dramatically lower signal loss and latency compared with standard silica single-mode fiber. Together, MOSAIC and HCF represent complementary changes in what Microsoft’s engineers call the “digital plumbing” inside and between datacenters.
What Microsoft built: MicroLEDs + imaging fiber explained
The MOSAIC principle: wide-and-slow instead of narrow-and-fast
MOSAIC abandons the conventional “narrow-and-fast” approach (a few high-speed serial optical channels) and replaces it with a wide array of slower optical channels implemented with MicroLED emitters. Each MicroLED channel operates at modest gigabit speeds, and the aggregate of many channels reaches or surpasses the throughput of today’s 400G–800G links. The architecture is specifically designed to be compatible with common pluggable form factors (QSFP/OSFP) and standard host interfaces, enabling a drop-in replacement model for existing hardware.Key technical points from the MOSAIC demonstration:
- MicroLED channels in the prototype transmit between ~1.3–2.0 Gbps per emitter; the team demonstrated a 100-channel prototype at 2 Gbps per channel over 20 m and showed how that can scale to 800 Gbps aggregate and beyond.
- Imaging fiber is used to carry thousands of parallel channels in a single cable by packing thousands of optical cores inside a single strand; this is the same basic family of fiber used for medical endoscopy and coherent imaging bundles, repurposed for datacenter links.
- The design sidesteps complex digital signal processing (DSP) by using analog-only backends and trading per-channel speed for energy and error budget.
Why imaging fiber matters
Conventional fiber optics are optimized for single-mode or multi-mode serial channels. Imaging fiber — often built as a coherent bundle with thousands of cores — is established technology in medical imaging where a tiny camera must transmit an image through a flexible bundle. MOSAIC leverages that idea: instead of aligning one laser into one core, arrayed MicroLEDs project patterned light (think of photonic “QR codes”) into many cores simultaneously, massively increasing parallelism without proportionally increasing complexity at the connector. This reduces per-channel driver costs and opens redundancy strategies that are infeasible with single-channel optics.Claimed benefits: energy, cost, reliability, and density
Energy savings: what the research shows
Microsoft’s team reports substantial energy reductions versus laser-based optical links. The MOSAIC paper quantifies power advantages — under certain configurations the architecture can reduce link power by up to 68% compared with today’s optical transceivers, largely because MicroLEDs and low-speed electronics avoid the heavy laser drivers and DSP that dominate optical link energy. Microsoft’s public materials and lab summaries presented to press and at conferences frame a conservative, deployable system expectation closer to ~50% energy savings when judged against mainstream laser-driven cables in real datacenter conditions. Both figures are consistent: the paper quantifies the maximum technical gains in prototype conditions while Microsoft’s external communications emphasize a conservative operational estimate.Cost and manufacturability
MicroLEDs are manufactured at scale for displays and are moving quickly down the cost curve. The MOSAIC approach relies on commercially available MicroLED arrays and on existing imaging-fiber manufacturing, which should keep bill-of-materials costs lower than custom laser modules. Microsoft has already run a proof-of-concept with partners (reported cooperation with MediaTek and other suppliers) to miniaturize the MicroLED transceiver into pluggable modules compatible with datacenter equipment. That indicates a pragmatic pathway to production-grade assemblies rather than a long-tail lab-only device.Reliability and operational are, by construction, simpler and mechanically more robust than precision lasers and modulators; they are less sensitive to thermal shifts and particulates. MOSAIC’s very high channel count enables graceful degradation: failed channels can be hot-swapped logically or remapped without losing the entire link, a level of fault tolerance that single-laser, DSP-heavy links cannot easily match. In lab reliability models, MOSAIC’s parallelism yields orders-of-magnitude improvements in mean time between failure for equivalent aggregate link capacity.
How Hollow Core Fiber complements MicroLED cabling
Hollow Core Fiber (HCF) addresses a different part of the network: long-distance propagation and end-to-end latency. By guiding light through a hollow (air) core rather than silica glass, HCF reduces the effective refractive index and increases the propagation speed of light in the cable. Microsoft and its partners report latency reductions on the order of ~30–47% and measured attenuation improvements that in recent labs have crossed the critical threshold of sub‑0.1 dB/km for production-quality HCF designs. That combination means HCF can deliver the same latency as conventional fiber over longer distances or materially lower latency at the same distance — valuable for geographically distributed AI training and latency-sensitive cloud services.Microsoft has already taken steps to industrialize HCF: the Lumenisity acquisition (December 2022) provided the technical foundation, and Microsoft’s public materials show active manufacturing collaboer makers to scale production for Azure deployments. The implication is clear: Microsoft expects to pair MOSAIC-style short-range links inside facilities with HCF across campus-to-campus and region-to-region backbones.
Technical verification and independent cross-checks
Microsoft’s MOSAIC claims are well-documented in a peer-reviewed SIGCOMM paper that includes a working prototype, measured bit-error-rate (BER) curves, end-to-end experiments with Ethernet/InfiniBand, and a detailed analysis of limits imposed by chromatic dispersion and MicroLED spectral width. The paper explicitly states:- 2 Gbps per microLED channel is feasible in prototype, and 1.3 Gbps/channel provides error-free operation over longer distances (e.g., ~20–30 m) with current hardware; production integration and tighter optics should improve those margins.
Strengths and practical upside
- Power and operating cost reductions. If MOSAIC can deliver even a 40–60% reduction in link power at scale, the cumulative effect across thousands of intra-rack and inter-rack links is a measurable drop in datacenter power usage and cooling load. That reduces operating expenses and carbon footprint for AI-heavy facilities.
- Reduced reliance on exotic lasers and DSP. Transitioning away from high-precision lasers reduces per-port complexity and potentially shortens maintenance windows and spare-part inventories. MicroLED arrays leverage mature display supply chains.
- High fault-tolerance via massive parallelism. Redailed microLEDs or fiber cores can be compensated for without taking a whole link down, simplifying availability engineering.
- Complementary network strategy. HCF reduces propagation delay between datacenters and enables larger AZ-like regions or fewer sites for the same latency budget, which translates into less infrastructure and lower embodied energy over time.
- Compatibility with existing form factors and protocols. MOSAIC’s backward-compatibility with QSFP/OSFP pluggables and Ethernet/InfiniBand stacks eases adoption risk and shortens integration timelines.
Risks, open engineering questions and operational challenges
No technology transition at hyperscale is frictionless. Several non-trivial issues need addressing before MicroLED cabling and HCF become mainstream:- Prototype limits vs. production expectations. MOSAIC’s prototype results are promising, but the paper clearly documents current BER limits at longer distances/higher per-channel speeds (e.g., 2 Gbps per channel shows BER issues at 30 m in prototype hardware). Achieving production-grade transceivers will require tighter optoelectronic integration, improved microLED linewidths and optical coupling — each a real engineering effort. The SIGCOMM paper is explicit about these limits and the need for additional integration work.
- Imaging fiber splicing, connectors and field maintenance. Imaging fiber bundles with thousands of cores are not currently part of the standard fiber connector/splice ecosystem. Field technicians will need new tools and training; connectorization and cold-splice reliability at datacenter scale are nontrivial. These ecosystem gaps are significant operational risks during initial rollouts.
- Manufacturing scale and supply chain risk. MicroLED yields for high-resolution displays are improving but are sensitive to defect rates. Likewise, HCF manufacturing at low loss and high reel lengths is only now moving from lab proof to factory scale. Microsoft’s partnerships with Corning and others reduce risk, but scaling to thousands of kilometers of HCF and millions of MicroLED transceivers will take time and capital.
- Cost parity at volume is still a projection. Microsoft projects lower lifetime costs due to simpler emitters and higher reliability, but initial module costs (optics, miniaturized packaging, imaging-fiber connectorization) could be higher than incumbent transceivers until volumes scale. Early adopters will likely absorb premium costs or accept integration risk.
- Standards and interoperability. Datacenter operators rely on robust standards and multivendor interoperability. MOSAIC’s protocol-agnostic relay design helps, but the industry will need connector standards, test methods and pluggable specifications before broader ecosystem support arrives. This is as much an industry coordination challenge as a technology one.
- Unverified commercialization timelines. Public communications (company press material and reporting) mention late‑stage commercialization planning, with Microsoft describing commercialization with partners in coming years. Independent confirmation of r example, “commercialized late in 2027”) is not fully documented in the peer-reviewed literature and remains a company projection subject to the usual product-development, supply-chain and regulatory contingencies. Readers should treat precise dates as Microsoft estimates rather than settled industry schedule.
Deployment scenarios and migration pathways
Microsoft’s engineering narrative is pragmatic: MOSAIC is positioned primarily for intra‑datacenter links — replacing copper or active optical cables used to tie GPUs, NICs and switches inside racks and across short campus distances (tens of meters). HCF targets backbone and inter‑datacenter connectivity where latency and long-distance propagation matter most. That dual-path deployment reduces the need for large rip-and-replace projects: operators can evaluate MOSAIC in controlled pods and incrementally adopt HCF for latency-critical routes.A possible staged rollout:
- Pilot MOSAIC in AI training clusters where density and power are highest and physical distance is limited (racks and pods).
- Standardize pluggable modules and test interoperability with switch ASIC vendors and NIC suppliers (NVLink/CXL awareness is already demonstrated in the paper).
- Introduce HCF between regional aggregation points where latency reductions yield measurable application benefits.
- Expand both technologies as manufacturing yields improve and operator tooling matures.
What the industry should watch next
- Progress on pluggable MOSAIC transceivers from supplier partners compliance with QSFP/OSFP shipping timelines.
- Field deployments reporting real-world power reductions and lifecycle operating data; lab numbers (up to 68% saved power) are exciting, but fleet-level evidence will drive adoption.
- Connectorization and splicing solutions for imaging fiber: without robust field tools and standards, operational costs will slow adoption.
- HCF industrialization metrics: sustained production yields, attenuation performance on long reels, and interoperability with existing amplification and ROADM systems.
- Any public benchmark comparisons showing MOSAIC links integrated into existing switch/NIC ecosystems under realistic datacenter loads.
Final analysis — opportunity vs. reality
Microsoft’s MicroLED + imaging-fiber approach addresses a painfully real problem: datacenter networking is becoming a first-order cost and energy driver as AI scales. The MOSAIC research is unusually complete for an academic-industry hybrid — it includes a working prototype, measured BER/throughput results, and a plausible engineering pathway to pluggable modules compatible with existing form factors. Those strengths make MOSAIC more credible than many lab-only proposals.But the transition is not risk-free. The devil is in production engineering: MicroLED linewidths, optomechanical coupling into imaging fiber, connector ecosystems, and scaled manufacturing all present honest obstacles. Hollow Core Fiber’s dramatic latency and attenuation metrics are real, but scaling HCF to global fiber plant lengths with consistent performance demands large capital investment and process control; Microsoft’s collaborations with established fiber manufacturers mitigate but do not eliminate that risk.
In short: MOSAIC plus HCF is a credible and potentially transformative systems play for AI‑first datacenters. The physics underpinning both ideas is sound, the prototypes and partnerships are real, and Microsoft is investing in manufacturing scale and supplier collaboration. Adoption will depend on how quickly engineering teams convert prototype gains into low-cost, field-serviceable modules and how rapidly the fiber industry standardizes connectors and splicing for imaging and hollow-core cables. Treat the published figures — whether 50% or up to 68% power reductions for MicroLED links, or 30–47% latency improvements for HCF — as strong technical signals backed by peer-reviewed work and lab/industry reports, but allow for the usual caveats around production maturity and supply-chain timing.
Conclusion
The shift from lasers to MicroLED arrays and from solid glass cores to hollow air cores is more than a single-component upgrade; it’s an architectural change in how datacenter networks are built. Microsoft’s MOSAIC project demonstrates that a wide-and-slow optical strategy can overcome many of the reach, power and reliability trade-offs that limit current designs, and Hollow Core Fiber can extend the latency and reach envelope for distributed AI workloads. Together they offer a credible pathway to materially more efficient, denser and lower-latency AI infrastructure — provided that the industry solves the production-scale engineering, connectorization and standards challenges that stand between lab prototypes and fleet-scale deployment. The next two years will be decisive: watch supplier pluggable modules, manufacturing yield announcements, and the first field reports documenting real power and latency benefits at scale.Source: Microsoft Source Using inexpensive MicroLEDs, Microsoft networking innovation aims to make datacenters more efficient
