Microsoft’s laser-free cable tech promises to slash AI data center networking power bills

Microsoft Research has developed MOSAIC, a MicroLED-based optical interconnect system for data centers that it estimates will use about 50% less energy than mainstream laser-based optical cables when deployed, the company said in a blog post.

The figure is based on lab tests of the system combined with deployment estimates, Microsoft said in the post.

MOSAIC was built at Microsoft’s Cambridge, UK, lab in collaboration with Azure Core, Azure Hardware Systems and Infrastructure, and Microsoft 365 teams, Microsoft said. Unlike conventional fiber optic cables that rely on a handful of high-speed laser-driven channels, the new system uses hundreds of parallel low-speed channels powered by cheaper, more temperature-stable MicroLEDs. It carries the signal through commercially available imaging fiber — the same multi-core cable used in medical endoscopy, Microsoft said in the blog post.

The team has completed a proof-of-concept with MediaTek and other suppliers that miniaturizes the full MOSAIC system into a thumb-sized transceiver compatible with existing datacenter equipment, Microsoft said. Commercialization with industry partners is expected by late 2027, the company added.

Electricity accounts for 46% of total spending at enterprise data centers and 60% at service provider facilities, according to IDC, which forecasts AI data center energy consumption growing at a compound annual rate of 44.7% to reach 146 terawatt-hours by 2027.

“Power is the biggest bottleneck in AI datacenters today,” said Neil Shah, VP for research and partner at Counterpoint Research, in an interview. “Microsoft’s use of inexpensive MicroLEDs is a good approach which could keep the thermal bottleneck in check within the power-hungry AI data center, thereby reducing TCO for hyperscalers and eventually CIOs renting the infrastructure.”

The power problem, Microsoft argues, starts with the cables themselves.

How MOSAIC works

Copper interconnects top out at roughly two meters at high data rates, limiting them to within a single rack. Laser-based fiber optic cables go further but consume more power and are sensitive to temperature and dust, Microsoft said in the post. MOSAIC reaches up to 50 meters while drawing less power than either, the company added.

“Imaging fiber looks like a standard fiber, but inside it has thousands of cores,” Paolo Costa, a Microsoft partner research manager and the project’s lead researcher, wrote in the post. “That was the missing piece. We finally had a way to carry thousands of parallel channels in one cable.”

MOSAIC is not Microsoft’s only optical networking bet, and it is not the one furthest along.

HCF is already in production across Azure regions

MOSAIC arrives alongside Hollow Core Fiber (HCF), a complementary technology Microsoft is already deploying globally. HCF carries optical signals through air rather than glass, delivering up to 47% faster data transmission and 33% lower latency than conventional single-mode fiber, according to published research from the University of Southampton cited by Microsoft.

Frank Rey, Microsoft’s general manager of Azure Hyperscale Networking, said in the post that the two technologies are complementary — HCF for long-distance inter-datacenter links, MOSAIC for in-facility GPU and server connectivity.

“With MicroLED, you have the pure efficiency of LED over a laser,” Rey said in the post. “That has a pure bottom-line impact to power usage at any given datacenter.”

Competitive context — and questions of scale

MOSAIC enters a crowded field. Nvidia and Broadcom are advancing co-packaged optics (CPO) as the preferred path to cutting interconnect power, with Nvidia’s CPO-based switches promising up to 3.5 times lower power consumption over pluggable transceivers and slated for commercial availability in 2026.

But CPO relies on lasers, and lasers are in short supply, with supply chain gaps expected to persist through 2027, said Naresh Singh, senior director analyst at Gartner, in an interview. “Microsoft’s MicroLED technology can come as a good alternative, in this context,” Singh said. He also flagged standardization as a broader adoption hurdle. Traditional optical interconnects have benefited from Multi-source Agreements that define industry standards for transceivers and modules. “Recent interconnect offerings have to aim for some standardization to drive faster and sustained adoption,” Singh said, pointing to the Open CPX MSA — an initiative to standardize CPO optical engines — as a step in that direction.

Counterpoint’s Shah, however, said MOSAIC faces challenges of its own: chromatic dispersion could limit its effective reach; specialized cabling and rack design changes add cost beyond the MicroLED components; and without buy-in from Nvidia or AMD, scalability remains uncertain. A bandwidth ceiling is also a risk. MOSAIC’s current sweet spot is 400G to 800G, and by its 2027-2028 deployment window, the industry may have moved to 1.6T or 3.2T targets, Shah said.

“While MicroLEDs can bring superior power benefits, there are a lot of other factors that could potentially limit widespread adoption,” Shah said. “But there could be some heterogeneity still possible if one or two major players adopt it in their portfolio, maybe replacing copper rather than silicon photonics or CPO.”