WhiteFiber and DriveNets Achieve 111.2 Tbps Cross-DC AI Fabric, Breaking Power Constraints
Summary
Key Takeaways
Project Redwood, announced by WhiteFiber, achieves 111.2 Tbps bandwidth and 0.9ms round-trip latency over 83km of dark fiber. The technology stack centers on DriveNets' Ethernet AI fabric, featuring Fabric Scheduled Ethernet (FSE), Virtual Output Queuing (VOQ), deep buffers, and cell-based load balancing. Storage is provided by WEKA NeuralMesh distributed data and memory infrastructure, with compute based on NVIDIA H200 GPUs. Two geographically separated GPU clusters are treated as a single logical supercluster, marking the industry's first commercialization of 'scale-across' AI clusters, overcoming power constraints of single data centers exceeding 100,000 GPUs. This Ethernet-based approach challenges InfiniBand dominance. WhiteFiber has filed related patents. Multiple vendors including NVIDIA, Broadcom, Arista, Cisco, Nokia, DriveNets, and optical suppliers are developing cross-DC metro AI fabrics.
Why It Matters
Beneath the breakthrough veneer, DriveNets and WhiteFiber aim to establish new control points in AI networking. Proprietary FSE and VOQ may lock users into their fabric, limiting interoperability. The claimed 0.9ms latency may still bottleneck AI training requiring microsecond-level synchronization (e.g., All-Reduce), especially over longer distances. Deep buffers and VOQ can increase tail latency, lacking InfiniBand's determinism under congestion. Dark fiber leasing costs and amplifier deployments could erode TCO benefits. WhiteFiber may use this to lock customers into long-term fiber leases, while WEKA's NeuralMesh adds storage layer lock-in. Users risk losing architectural flexibility.
PRO Decision
【Vendors】 NVIDIA should accelerate its Spectrum-X Ethernet offering and highlight InfiniBand's maturity for cross-DC AI, while pointing out the non-standard nature of DriveNets FSE posing interoperability risks. Arista and Cisco should champion standard Ethernet AI fabrics and collaborate with open communities to dilute DriveNets' proprietary moat.
【Enterprises】 CIOs must demand independent benchmarks for cross-DC fabric, especially tail latency and congestion control (PFC/ECN) over distance. Audit FSE compatibility with standard Ethernet to avoid lock-in. Contractually require multi-vendor interoperability to preserve future flexibility. Assess dark fiber leasing costs versus alternatives like DWDM.
【Investors】 Be cautious of WhiteFiber's patent strategy potentially stifling competition. Recognize the long-term trend of Ethernet replacing InfiniBand in AI; vendors backing open standards (e.g., Ultra Ethernet Consortium) may be safer bets. WhiteFiber's fiber assets could appreciate if cross-DC fabric gains traction, but watch for technology disruption.
Get 3-5 key AI infrastructure signals weekly →
💬 Comments (0)