Catalina (Meta AI rack)
Last reviewed
Jun 3, 2026
Sources
11 citations
Review status
Source-backed
Revision
v1 · 1,371 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Jun 3, 2026
Sources
11 citations
Review status
Source-backed
Revision
v1 · 1,371 words
Add missing citations, update stale details, or suggest a clearer explanation.
Catalina is a high-power, liquid-cooled rack system designed by Meta for training and serving large AI models. Meta unveiled it at the Open Compute Project (OCP) Global Summit in October 2024 and said it would contribute the design to OCP as open hardware. Catalina is built around NVIDIA's GB200 NVL72 rack-scale platform, which pairs Grace CPUs with Blackwell GPUs, and it packages that platform into Meta's own rack and power architecture rather than NVIDIA's reference enclosure. [1][2]
The system is one of the clearest examples of how hyperscale operators adapted their data center hardware in 2024 to absorb the power and cooling load of dense GPU systems, where a single rack can draw more than ten times the power of a conventional server rack.
Catalina is organized as a two-rack "pod." The two information-technology (IT) racks together form a single scaling domain of 72 GPUs, matching the GB200 NVL72 configuration they host. Each IT rack carries 18 compute trays, split between the top and bottom of the rack, plus nine NVLink switches on each side for GPU-to-GPU interconnection. Those NVLink connections tie all 72 GPUs into one coherent memory domain, so software can treat the pod as a single large accelerator rather than a collection of separate nodes. [3][4]
The rack itself is the first high-power implementation of Meta's Open Rack v3 (ORv3) standard, which Meta refers to as the high-power rack, or HPR. Standard ORv3 racks deliver up to about 94 kilowatts (kW) at 600 amps; the Catalina HPR variant raises that ceiling to roughly 140 kW per rack, with power converted from 480-volt feeds down to a 48-volt DC busbar. The full solution is liquid-cooled, and Meta paired it with monitored liquid management, including sensors to detect leaks. [3][4]
| Attribute | Detail |
|---|---|
| Vendor / designer | Meta |
| Debut | OCP Global Summit, October 2024 |
| Compute platform | NVIDIA GB200 NVL72 (Grace Blackwell) |
| GPUs per pod | 72 Blackwell GPUs (two IT racks) |
| CPUs | 36 NVIDIA Grace CPUs (as 36 GB200 superchips) |
| Compute trays | 18 per IT rack (top and bottom) |
| Switching | Nine NVLink switches per side; NVLink coherent memory domain |
| Rack form factor | Open Rack v3 (ORv3) high-power rack (HPR) |
| Peak power | Up to ~140 kW per rack |
| Power distribution | 480 V input converted to 48 V DC |
| Cooling | Liquid cooling |
| Network fabric | Wedge 400 fabric switch (plus management switch) |
| License | Contributed to the Open Compute Project (open hardware) |
Beyond the compute and switch trays and the ORv3 HPR frame, a complete Catalina deployment includes a power shelf, a Wedge 400 fabric switch, a separate management switch, a battery backup unit (BBU), and a rack management controller (RMC). Meta described the layout as modular so that other operators could swap components to fit their own workloads while reusing the same rack, power, and cooling building blocks. [3]
Catalina's compute comes entirely from NVIDIA's GB200 NVL72, the rack-scale member of the Blackwell generation. The "GB200" name refers to the Grace Blackwell superchip, which combines one Grace CPU with two Blackwell GPUs over NVIDIA's 900 GB/s NVLink-C2C chip-to-chip link. The "NVL72" name describes how many of those GPUs are connected: 72. A full NVL72 contains 36 Grace CPUs and 72 Blackwell GPUs, assembled from 36 superchips across 18 compute trays and linked by nine NVLink switch trays running NVIDIA's fifth-generation NVLink fabric. [5][6]
That fabric gives the 72 GPUs a shared, high-bandwidth interconnect (NVIDIA cites roughly 130 TB/s of aggregate NVLink bandwidth and about 13.4 TB of unified GPU memory across the rack), which is what lets the whole assembly behave as one accelerator for very large models. NVIDIA's own reference build fits the NVL72 into a single liquid-cooled rack drawing on the order of 120 kW. Meta instead spread the same 72-GPU domain across two ORv3 IT racks, a choice that fits NVL72 into Meta's existing rack width and power and cooling envelope while still leaving headroom (the ~140 kW HPR ceiling) for the dense compute and the switching. [4][5]
Because the hardware generation matters here, it is worth being explicit: Catalina is a Blackwell-generation system (GB200), not a Hopper-generation one. Meta's earlier and concurrent Hopper deployments used different hardware, discussed below.
Meta introduced Open Rack v3 in 2022 alongside its Grand Teton platform, defining a flexible power and rack architecture for AI hardware. Catalina extended ORv3 into its high-power regime, the HPR, to reach the ~140 kW needed for a Blackwell pod. [3][7]
A common point of confusion is the relationship between Catalina and "Open Rack Wide" (ORW). They are not the same thing. The Catalina system shown at OCP in October 2024 used the ORv3 HPR form factor. Open Rack Wide is a later, separate Meta contribution: a double-wide rack roughly 1,200 mm across (about twice the ~600 mm of ORv3), published as an OCP specification in 2025 to give next-generation AI systems more room for compute trays, liquid-cooling manifolds, and power distribution. ORW became the basis for systems such as AMD's "Helios" rack. In short, ORv3 HPR powered the 2024 Catalina pod, and ORW is the wider successor form factor that arrived the following year. [8][9]
Catalina sits inside a longer line of open AI hardware from Meta. The company introduced Grand Teton, an in-house GPU server platform, at the 2022 OCP Global Summit; its first version used NVIDIA H100 GPUs and was contributed to OCP. Grand Teton is a server (a chassis that integrates CPUs, GPUs, networking, power, and control into one monolithic system), whereas Catalina is a full rack and power design that hosts a rack-scale GPU platform. The two are complementary parts of the same open hardware program rather than competing products. [7][10]
At the same October 2024 OCP summit where it unveiled Catalina, Meta also announced an expanded Grand Teton that adds support for AMD's Instinct MI300X accelerator, and said it would contribute that version to OCP as well. Meta framed both moves as part of scaling its fleet for ever-larger models: the company noted that it trained Llama 3.1 405B on more than 16,000 H100 GPUs and was operating two clusters of roughly 24,000 GPUs each, with much larger build-outs planned. Catalina is the rack-level vehicle for that next phase of growth, giving Meta a standardized, liquid-cooled, ~140 kW enclosure for Blackwell-class clusters. [3][11]
By publishing Catalina, the ORv3 HPR, and the expanded Grand Teton through OCP, Meta extended a pattern it has followed since the Open Compute Project's founding in 2011: releasing its data center designs so that suppliers and other operators can build compatible hardware, which in turn broadens the supply base Meta itself draws on. [3]