Ayar Labs Joins NVIDIA's NVLink Fusion: How Co-Packaged Optics Will Power Million-GPU AI Factories

The Bottleneck: Why Copper Has No Future in AI Factories

Copper interconnects fail across three dimensions as AI clusters grow :

• Bandwidth density: Trace lengths beyond a few meters lose signal integrity fast, forcing engineers to use narrower links or power-hungry retimers. Ayar Labs' optical chiplets sustain high bandwidth across rack-scale distances, unifying thousands of GPUs into a single cluster .
• Power per bit: Electrical SerDes at 100+ Gbps consume untenable amounts of energy. Optical I/O slashes power consumption per transmitted gigabit, which becomes the dominant constraint when million-GPU clusters would otherwise burn massive energy just moving data between chips .
• Physical reach: Copper chains racks together by the length of a cable. Optics decouple distance from signal quality, enabling disaggregated, rack-spanning topologies that are impossible with electrical links alone .

Ayar Labs' approach is distinct because it packages photonics and electronics together in a single super-chiplet, placing the optical I/O directly next to the compute die. This eliminates the power-sapping electrical-to-optical conversion at the switch faceplate that pluggable optics require, delivering better energy efficiency at higher bandwidth .

A Multi-Vendor Future: Why AMD and MediaTek Invested

What makes this partnership strategically unusual is Ayar Labs' investor list. Its $500 million Series E round, closed on March 3, 2026, brought together not only NVIDIA but also AMD, MediaTek, Alchip, and a consortium of institutional investors led by Neuberger Berman . The company's total funding reached $870 million at a $3.75 billion valuation .

This coalition of chipmakers has concrete implications for heterogeneous compute deployments:

• A single optical fabric across architectures: Ayar Labs' CPO can connect NVIDIA GPUs, AMD GPUs, and other accelerators within the same optical fabric. Hyperscalers operating mixed fleets can leverage NVLink Fusion's capabilities without accepting a proprietary, single-vendor interconnect stack .
• Open-ecosystem signaling: AMD and MediaTek co-investing alongside NVIDIA sends a clear market signal that the industry wants a standardized CPO layer that works across silicon architectures, not a proprietary NVIDIA-only solution .
• Manufacturing scale across foundries: Ayar Labs partners with GlobalFoundries, Intel Foundry Services, and TSMC for manufacturing and assembly, giving it a multi-source production base that is essential for serving diverse heterogeneous deployments at volume .

The Bigger Picture: NVIDIA's $6.5 Billion Bet on Light

The Ayar Labs deal is a single tile in a much larger mosaic. Since early March 2026, NVIDIA has deployed at least $6.5 billion across five photonics companies: $2 billion each in Lumentum, Coherent, and Marvell; up to $3.2 billion in Corning; and participation in Ayar Labs' Series E . This spending spree, compressed into a single quarter, makes NVIDIA the world's largest investor in optical interconnect technology.

Jensen Huang stated at GTC 2026 that "the silicon photonics capacity we need far exceeds the current supply" . For NVIDIA, the math is straightforward: electrical links between GPUs cannot scale to the exaflop-class clusters needed for next-generation AI training without blowing through power budgets and physical space. Optical interconnect is not a speculative technology anymore; it is the binding constraint on AI's next phase .

Ayar Labs' NVLink Fusion qualification is the first concrete, standard-level integration of CPO into NVIDIA's scale-up fabric. It provides a reference design for how future racks, rows, and entire data center halls will be stitched together, not with copper cables but with light from the chip package itself—connecting GPUs as a single unified system regardless of vendor or physical location.