NVIDIA Will Wire 576 Rubin Ultra GPUs Into One Giant GPU Over Optics, Aiming at 10-Trillion-Parameter Models in 2027

NVIDIA has used this year OCP Global Summit to show how far it intends to push the idea that a data center, not a chip, is the right unit of compute. Ian Buck, the company vice president and general manager for hyperscale and HPC, walked attendees through Rubin Ultra NVL576: 576 GPUs, spread across multiple fully liquid-cooled Kyber racks, stitched into a single NVLink compute domain by a next-generation co-packaged optics switch tray that the company demonstrated in public for the first time.

On NVIDIA published roadmap figures, the NVL576 configuration delivers 15 exaflops of FP4 inference and 5 exaflops of FP8 training, which the company puts at 14 times the throughput of the current Blackwell GB300 NVL72 system. Each Rubin Ultra GPU uses a four-reticle die design and carries up to 1TB of HBM4E. The full system is specified at 365TB of fast memory with 4.6 PB/s of aggregate memory bandwidth, a 1.5 PB/s NVLink 7 domain interconnect, and new ConnectX-9 Spectrum parts rated at 115.2 TB/s of network capability. None of this has been independently validated; these are roadmap numbers, and the platform is not expected to ship until 2027.

The rack design is as much a part of the story as the silicon. Each Vera Rubin Ultra compute tray packs two Rubin Ultra GPUs and one Vera CPU, all liquid-cooled. Buck described the resulting machine as a giant GPU in the scheduling sense rather than the literal one: 576 chips that present themselves to software as one coherent resource. NVIDIA is contributing the rack and tray designs to OCP, keeping the tool-less modular scheme introduced with Vera Rubin, which cuts a single-rack installation from hours to minutes.

The optics layer is where the economics get interesting. Earlier this year there were credible reports that co-packaged optics had slipped; NVIDIA has since pushed back hard, and Jensen Huang told the Berlin edition of GTC that CPO racks are now in volume production. A note of caution applies here: that claim covers Spectrum-X Ethernet optics, not the NVLink 7 CPO tray inside NVL576, which remains a demonstration. Rubin Ultra will be the first rack-scale design to lean on the technology end to end if the 2027 timeline holds.

The strategic shift is that the thing customers buy is changing shape again. First it was the GPU, then the NVL72 rack; NVL576 makes it a multi-rack NVLink domain that is provisioned, cooled and networked as one object. For anyone planning to train or serve models in the ten-trillion-parameter range, procurement math starts with whole domains, which concentrates buying power in fewer, larger commitments.

NVIDIA is pairing the roadmap with delivery mechanics. Its DSX Pod reference design puts seven Vera Rubin NVL72 racks inside a standard shipping-container footprint, factory-integrated and pre-tested, with site deployment as short as roughly 14 days once power and facilities are ready. The 14-day figure covers the on-site phase only, not the underlying construction, a distinction worth keeping in mind when the number circulates.

If Rubin lands on schedule in the second half of 2026 and Rubin Ultra follows in 2027, NVIDIA will have moved from 72 connected GPUs to 576 in about three generations. The bet is that optical interconnect makes the leap economic; the risk, as always with roadmap arithmetic, is that exaflop promises meet the same physics and supply chains that every previous generation under-estimated.