Microsoft's MOSAIC: How MicroLEDs Could Replace Lasers and Cut Data Center Networking Power by 68%

Laser-based optical cables in AI data centers consume 9.8–12W per 800 Gbps link, fail every 6–12 hours at 100,000-GPU scale, and depend on a supply chain that Gartner says remains constrained through 2027. Microsoft just published a fundamentally different approach.

MOSAIC replaces lasers with MicroLEDs — commodity display technology running at 100–1,000x lower power per emitter — and distributes 800 Gbps across 400+ parallel channels at 2 Gbps each on medical-grade imaging fiber. End-to-end power: 3.1–5.3W per 800G link. 50-meter reach. QSFP/OSFP compatible. ACM SIGCOMM 2025 Best Paper, with a MediaTek proof-of-concept already built.

Here's how it actually works and what it changes.

The Problem: Power Is the Real Bottleneck

Electricity accounts for 46% of total spending at enterprise data centers and 60% at service provider facilities (IDC). AI data center energy consumption is growing at a compound annual rate of 44.7%, projected to reach 146 TWh by 2027.

In a typical NVIDIA NVL72 pod (72 B200 GPUs), optical link power adds ~20 kW per rack. Optical link failures at scale occur every 6–12 hours. These constraints force engineers to rely on copper DAC cables, which limits all 72 GPUs to a single rack and pushes rack power density to 120 kW — requiring complex liquid cooling.

Interconnect Type	Reach	Power (800G)	Failure Rate	Temp Sensitivity
Copper (DAC)	~2 m	Passive (0W)	Very low	Low
Laser optics (AOC)	100+ m	9.8–12W	High	High (dust/heat)
MOSAIC MicroLED	50 m	3.1–5.3W	Very low	Low

"Power is the biggest bottleneck in AI datacenters today," said Neil Shah, VP at Counterpoint Research.

Wide-and-Slow: The Architecture

MOSAIC flips the conventional "Narrow-and-Fast" optical model. Instead of 8 channels × 100 Gbps (laser-based 800G), it runs 400+ channels × 2 Gbps each. This eliminates three of the most power-hungry components in traditional optics:

Three Core Components

1. Directly Modulated MicroLEDs — Each MicroLED measures a few microns across and consumes only hundreds of microwatts (vs. tens–hundreds of milliwatts for lasers). A monolithically integrated array packs 400+ emitters within 1 mm². Simple ON/OFF (NRZ) modulation at 2 Gbps. Temperature-stable, dust-insensitive.

2. Multicore Imaging Fiber — Borrowed from medical endoscopy, imaging fiber bundles thousands of individual cores inside a single cable. Each MicroLED's signal maps to multiple fiber cores, simplifying alignment. "Imaging fiber looks like a standard fiber, but inside it has thousands of cores," wrote Paolo Costa, MOSAIC's lead researcher. "That was the missing piece."

3. Low-Power Analog Backend — At 2 Gbps per channel with NRZ encoding, you don't need DSP, ADC/DAC, or CDR circuits. The clock rides a dedicated control channel (0.25% overhead for 400 channels). Simple analog equalization handles chromatic dispersion.

Where the Power Savings Come From

Component	Traditional 800G	MOSAIC 800G
DSP / CDR	3.5W	0W (eliminated)
Light source + drivers	4.7W (lasers)	1.2W (MicroLEDs)
Digital backend	In DSP	0.4W
Host interface	0.2–2.4W	0.2–2.4W
MCU / DC-DC	1.4W	1.3W
Total	9.8–12W	3.1–5.3W

The DSP elimination alone saves 3.5W — roughly 30% of a traditional link's power budget. For 1.6 Tbps, MOSAIC projects 10.6W vs. 23–25W for conventional designs.

MOSAIC vs. Co-Packaged Optics (CPO)

CPO is the industry's other major play — NVIDIA and Broadcom are pushing it for 2026 availability. CPO integrates transceivers into the switch/NIC package, cutting power 25–30%.

MOSAIC achieves 56–68% reduction. And they're complementary, not competing:

Feature	Pluggable	CPO	MOSAIC
Power reduction	Baseline	25–30%	56–68%
Light source	Lasers	Lasers	MicroLEDs
Reach	100+ m	100+ m	Up to 50 m
Laser supply chain risk	Yes	Yes	No
CPO-compatible	N/A	N/A	Yes
Available	Now	2026	Late 2027

The supply chain angle is significant. Gartner's Naresh Singh: "Microsoft's MicroLED technology can come as a good alternative" to laser supply constraints. MOSAIC uses commodity MicroLED and CMOS sensor manufacturing — both mature, high-volume supply chains.

Topology Implications

MOSAIC's 50-meter reach at copper-like power opens topology options:

1. ToR switch elimination — servers connect directly to Row/EoR switches, reducing latency and hardware cost
2. GPU fabric disaggregation — instead of confining 72 GPUs to one rack, MicroLED enables cross-rack GPU-to-GPU at low power
3. Advanced topologies — multi-dimensional torus, dragonfly, hypercube become practical when 50m reach removes the copper distance constraint
4. Memory disaggregation — MOSAIC's nanosecond-level latency (no FEC/DSP) supports separating memory pools from compute

"This architectural shift enables Microsoft to scale its Azure GPU clusters more densely than rivals such as AWS and Google Cloud, which remain tethered to power-intensive, heat-sensitive laser systems," said Ron Westfall, VP at HyperFrame Research.

Limitations

Not a silver bullet:

• 50m max reach — MicroLEDs have broad spectral widths causing chromatic dispersion. Can't replace long-haul laser optics.
• Bandwidth ceiling — current sweet spot is 400G–800G. By 2027 the industry targets 1.6T–3.2T. Scaling path exists (more channels, 4–8 Gbps per channel) but isn't proven at scale.
• Ecosystem adoption — without NVIDIA/AMD GPU-side integration, scalability is uncertain. Standardization (MSAs) needed.
• Imaging fiber isn't standard cabling — infrastructure changes add cost beyond the MicroLED components.

Microsoft's Dual-Layer Optical Strategy

MOSAIC is half of a comprehensive optical overhaul:

Layer	Technology	Range	Status
Intra-rack	Copper DAC	<2 m	Current
Intra-facility	MOSAIC MicroLED	Up to 50 m	PoC complete, 2027
Inter-DC	Hollow Core Fiber	Long-haul	In production (Azure)

Hollow Core Fiber (from Microsoft's 2022 Lumenisity acquisition) transmits light through air instead of glass — 47% faster, 33% lower latency vs. standard single-mode fiber.

Bottom Line

MOSAIC attacks the right problem at the right layer. AI data center scaling is hitting a power wall, and networking interconnects are a meaningful contributor. Trading expensive, failure-prone, power-hungry lasers for commodity MicroLEDs on imaging fiber is an elegant architecture — if it can scale to 1.6T+ and gain ecosystem adoption by 2027.

The competitive angle: if MOSAIC works at scale, Microsoft sidesteps the laser supply chain entirely. AWS and Google don't have an equivalent play. That's not just a technical advantage — it's a supply chain moat.

Worth watching closely.

---

Originally published at FirstPassLab. More deep dives on data center networking, AI infrastructure, and network engineering at firstpasslab.com.

---

AI Disclosure: This article was adapted from original research with AI assistance for editing and formatting.