NVIDIA's high-speed GPU-to-GPU interconnect technology that links multiple GPUs into a unified, high-bandwidth memory and compute domain for AI workloads.
Arista's Etherlink SU-144 links up to 144 accelerators in one Ethernet hop and 1,024 across racks, built on the ESUN initiative. Arista says Nvidia-based racks stay on proprietary NVLink for now, so the open choice belongs to buyers of other accelerators.
Reality
NVIDIA's open source Topograph discovers cluster topology from five named clouds or from ibnetdiscover on-premises, then republishes it as node labels, Slurm config or Slinky ConfigMaps whenever a watched part of the cluster changes.
Reality
NVIDIA measured TensorRT LLM holding 96.1 to 98.2 percent of its non-confidential output throughput on Blackwell, and it got there by unpinning host memory on the affected paths, moving decode readback off the scheduler thread, and timing kernel tactics with the GPU's global timer instead of CUDA events.
Reality
DataEnclave uses Nvidia confidential computing so a model developer's base weights and a customer's fine-tuned weights can run in one enclave under separate key control. Vast says it charges nothing extra for it.
Reality
MLPerf Inference v6.1 preview submissions put Vera Rubin NVL72 at up to 3.7x GB300 on Qwen3-VL and up to 2.5x on DeepSeek-R1, on two different inference frameworks. The four-rack 99% scaling result is an offline number.
Reality
The Intel spinout's Active Compute Fabric targets Nvidia's NVLink but reaches customers only through platform builders such as Qualcomm, on terms neither company has described.
Reality
KeyBanc's optical thesis for Marvell and Marvell's own revenue guidance both land near $30 billion, four years apart and measuring different things. The purchase price for the technology is already committed.
Reality
The Windows on Arm port and the compute-capability-107 Rubin preview are the headline items, but the two lines that touch a running cluster are the unbundled driver installer and the new CDMM default on coherent platforms.
Reality
The quarter came in at $96.2B. The argument that compute must now leave the mega-cluster rests on one analyst's own 30 GW estimate, and on nothing Nvidia itself said.
Reality
NVIDIA says Alibaba's largest open-weight model serves over 4K tokens/sec/GPU and 350 tokens/sec/user in FP8 on a GB300 NVL72. That figure is the self-hosting floor, not a benchmark.
Reality