What NVLink Fusion Optics Actually Changed
NVIDIA’s NVLink Fusion program lets hyperscalers build custom AI systems around its scale-up fabric instead of buying a fixed GPU-to-GPU topology. The partner list has now grown to include photonics vendors, and IEEE Spectrum frames this NVLink Fusion optics shift as NVIDIA “quietly dipping its toes” into optics for scale-up connectivity rather than just scale-out.
Until now, data centers split the job cleanly: optics moved data across the room between racks (scale-out), while copper NVLink moved data between GPUs inside a rack (scale-up), because copper kept latency low over short reach. That boundary is what the new optics partners are starting to test.
The concrete change is an ecosystem addition, not a new NVIDIA reference architecture. Ayar Labs announced it joined the NVLink Fusion ecosystem, making its co-packaged optics (CPO) products optically and electrically compatible with NVIDIA’s own optical and SerDes technologies.
That compatibility claim matters more than the announcement’s marketing framing. It means a hyperscaler building a custom XPU or CPU into an NVLink Fusion rack can now choose Ayar Labs’ CPO as the physical link layer instead of, or alongside, copper NVLink traces.
Why Copper Needed a Photonics Escape Hatch
Astera Labs, an existing NVLink Fusion lead partner since the platform’s Computex 2025 unveiling, has already documented what is straining copper. Its own channel-budget analysis cites 40 dB of channel budget across 1.75 meters at 224G signaling, describing exactly the kind of link that gets harder as racks scale from an 8-GPU chassis to 72, 144, or more accelerators.
Ayar Labs’ framing lands on the same problem from the optics side. “AI infrastructure is being co-designed from the ground up, and customers need more options to scale performance efficiently as bandwidth continues to rise,” said CEO Mark Wade.
NVIDIA’s own comment on the addition, from Vice President Ashish Karandikar, stays deliberately narrow: NVLink Fusion combined with Ayar Labs’ CPO “gives customers more options to build heterogeneous AI factories”. That is an options statement, not a roadmap commitment to optics by default.
The same announcement discloses that Ayar Labs closed a $500M Series E funding round with NVIDIA participating as an investor, a financial signal worth separating from the technical compatibility claim.
What Copper NVLink Still Delivers Today
Optics has to earn its place rather than simply replace copper, because NVIDIA’s current copper-based NVLink generation is not standing still. HGX Rubin NVL8 pairs eight Rubin GPUs with sixth-generation NVLink interconnects and 28.8 TB/s of NVLink Switch bandwidth for inter-GPU communication.
NVIDIA states that configuration delivers 10x more token factory throughput than HGX B200 and matches its training performance with 4x fewer GPUs, backed by 400 PFLOPS of NVFP4 compute and 176 TB/s of memory bandwidth. These are NVIDIA’s own projected figures, explicitly labeled as projected performance subject to change, not independently verified benchmarks.
Separately, NVIDIA’s Rubin architecture disclosures describe the chip itself as built on HBM4 memory with NVLink 6 and 336 Billion Transistors, the silicon generation the optics conversation sits on top of, not a replacement for it.
Copper, Optics, or Both: The Vendor Split to Track
The ecosystem now spans two distinct connectivity approaches under one NVLink Fusion umbrella, and they solve different problems rather than compete head-to-head.
| Vendor | NVLink Fusion role | Connectivity type | What it targets |
|---|---|---|---|
| Ayar Labs | Ecosystem partner added in 2026 | Co-packaged optics (CPO) | Bandwidth and power headroom once copper’s electrical reach runs out |
| Astera Labs | Lead partner since Computex 2025 | PCIe/CXL retimers, Ethernet smart cable modules | Extending copper reach, including 7m links over flexible cables for multi-rack clustering |
Reading the table plainly: Astera Labs extends how far copper can still go, while Ayar Labs offers an optical alternative once copper’s channel budget is exhausted. Buyers do not have to pick one exclusively, since NVLink Fusion is designed to let custom silicon connect through either path.
Choose the copper-extension path, Astera-style retimers and cable modules, when the rack topology stays within a few meters and the priority is preserving existing NVLink-based designs without a new physical layer to validate. Choose the CPO path when the design integrates heterogeneous, power-constrained custom accelerators where copper’s electrical budget is the binding constraint.
What to Verify Before Committing
Ayar Labs states it will work with customers on “system architecture, validation requirements, and platform timelines,” language that signals integration is still being worked out per deployment rather than shipping as a fixed, qualified rack SKU.
Nothing in the available announcements states a shipping date for an NVLink Fusion rack with CPO installed by default, nor a power or cost delta against copper NVLink at equivalent bandwidth. Those figures are not established by the evidence at hand.
For scale context on the deployments NVIDIA is currently selling into, its Vera Rubin platform is powering infrastructure builds such as a Noetra Corp. project sized at 13,750 Vera CPUs and 27,500 Rubin GPUs, none of which is described as optics-connected at the scale-up layer.
Before signing on to an NVLink Fusion design that assumes co-packaged optics, confirm three things directly with the vendor: the validated power-per-bit figure against the copper baseline, the actual channel-budget headroom the CPO link buys back relative to the 40 dB / 1.75 meter / 224G copper limit Astera Labs has documented, and a committed platform timeline rather than an ecosystem-membership announcement.