Open Compute Project
The Open Compute Project (OCP) is a community that develops and shares designs, specifications, and related engineering work for computing infrastructure. Its work includes servers, storage, networking, racks, power systems, cooling, and infrastructure management. The Open Compute Project Foundation supports that collaboration; manufacturers and service providers build and supply products based on particular contributions. OCP describes itself as a collaborative organization rather than a standards body.[1]
OCP began in 2011 with Facebook's publication of hardware and facility designs for its data center in Prineville, Oregon.[2] Its scope now includes infrastructure for AI accelerators, with work coordinated across physical facilities, IT equipment, and systems management. An OCP contribution may describe an interface, an implementation, or a deployment practice. It is not necessarily a complete server that can be ordered from a supplier.[3][4]
Origins
Facebook announced OCP on April 7, 2011. The initial release included technical specifications and mechanical CAD files for servers, power supplies, racks, battery backup equipment, and the data center building. The company presented open hardware as a way to share infrastructure engineering with other operators and manufacturers.[2]
The launch announcement reported that Facebook's Prineville implementation used 38 percent less energy for the same work and cost 24 percent less to build out than its existing facilities. Those were Facebook's comparisons for that deployment, not savings guaranteed by adopting any OCP specification.[2]
The project expanded beyond those initial designs. Its current community structure includes separate projects for networking, storage, server hardware, firmware, cooling, hardware management, security, sustainability, and other infrastructure areas. Workstreams can address a narrow engineering problem within a larger project, while strategic initiatives coordinate work across projects.[5]
Foundation and governance
The Open Compute Project Foundation is a Delaware corporation organized under section 501(c)(6) of the United States Internal Revenue Code. Its Board of Directors sets the Foundation's strategic direction and charters projects. The Steering Committee reviews proposed specifications, while project leaders organize technical work and help contributors prepare submissions.[6]
The March 2025 operating policy separates project leadership, subproject leadership, and workstream leadership. Project leaders maintain charters and roadmaps, organize meetings, and review contributions. Steering Committee representatives coordinate projects and participate in contribution approval. This structure allows a technical contribution to undergo community review without making its author the governing authority for the whole organization.[7]
OCP's published tenets are efficiency, impact, openness, scalability, and sustainability. These are criteria for community contributions and programs, rather than a single numerical performance benchmark.[8]
Contributions and licenses
Different kinds of technical material
OCP accepts several forms of engineering material. Their purposes differ, so an informational paper should not be read as a mandatory implementation specification.[4]
| Material | What it describes |
|---|---|
| Requirements document | Needs, constraints, and acceptance criteria for a proposed solution. |
| Base specification | Architecture, common interfaces, and system-level constraints. |
| Design specification | More detailed engineering requirements for an implementation. |
| Product specification | Manufacturing and implementation detail for a particular product. |
| Design package | Supporting engineering artifacts, such as drawings, models, manufacturing files, and bills of materials. |
| Source code | Software, firmware, or hardware-description code associated with a contribution. |
| Guidelines or white paper | Recommended practices, technical explanations, or analysis. |
| Tested configuration | Documentation of a particular combination of components and its testing. |
These distinctions appear in OCP's published draft contribution taxonomy, dated June 10, 2026. Actual documents should be checked for their version, release status, supporting files, and applicable license.[4]
Review and publication
For a specification, the contribution process starts with the relevant project group and an abstract submitted through the OCP Contribution Hub. Contributors sign the required agreement, prepare the specification, present it to the project, and address feedback before Steering Committee review. Staff publish the approved work in the contributions database after the applicable intellectual-property process.[9]
The workflow differs for software, design files, and informational documents. Product recognition has a separate application route. A presentation at an OCP event or a document circulating for comment therefore should not automatically be described as an approved specification.[9]
Intellectual property
OCP's operating policy states that contributors retain ownership of their contributed intellectual property. Publication through OCP does not transfer that ownership to the Foundation.[7]
The intellectual-property policy dated August 14, 2026 identifies the Open Web Foundation Contributor License Agreement 0.9 for specification contributions and the Open Web Foundation Agreement 0.9 as the specification license. Software contributions use an Open Source Initiative-approved license. Other contributions may use a public-domain release, a Creative Commons license, or a copyright license agreement.[10]
The policy includes a 30-day opt-out period for proposed specifications and a procedure for patent exclusion notices. Release notices identify contributors and applicable exclusions. Consequently, the word "open" does not mean that every associated technology is in the public domain or that every document has the same license. The terms attached to the actual contribution govern its reuse.[10]
Technical work relevant to AI infrastructure
Racks and power
The Rack & Power project covers rack architectures, rack-level power conversion and storage, and mechanical components that support IT equipment. Its scope includes both 19-inch EIA-compatible systems and 21-inch Open Rack systems. OCP therefore is not synonymous with one rack width or one generation of rack hardware.[11]
Open Rack v3 (ORv3) is a family of related specifications. The official document directory separates the rack base and frame specifications from power shelves, power supplies, battery backup equipment, connectors, manifolds, and IT equipment design guides. It also distinguishes documents under review from published design resources.[12]
In its October 2022 ORv3 description, Meta explained a 48 V DC power architecture with shelves that can occupy different positions along the rack's busbar. That implementation supported configurable power and battery arrangements and was designed to accommodate systems such as Grand Teton. The same article described optional blind-mate liquid-cooling connections, which connect equipment to a rack manifold during insertion.[13]
These details are version-specific. A power figure for a particular shelf or historical rack configuration is not a universal limit for all OCP equipment. Integration requires the matching electrical, mechanical, and cooling documents for the chosen design.[12][13]
Accelerator modules
In March 2019, OCP announced the Open Accelerator Infrastructure (OAI) subproject following an OCP Accelerator Module (OAM) specification contribution by Facebook, Microsoft, and Baidu. OAM defines a mezzanine accelerator module form factor and associated interfaces. OAI addresses the surrounding system, including baseboards, host connections, management, trays, and chassis.[14]
The distinction matters when describing an accelerator product. A module specification addresses the hardware integration boundary; it is not the accelerator's instruction set, software development environment, or a guarantee that applications can move unchanged between different chips. OAI's documented scope concerns common infrastructure for heterogeneous accelerator systems.[14][15]
Cooling
The Cooling Environments project covers cold plates, coolant distribution units, immersion, door heat exchangers, and heat reuse. These are different approaches and supporting systems, not requirements that every OCP installation must use together.[16]
The Cold Plate subproject works on the technology cooling system between cold plates and coolant distribution units, including tubing, manifolds, and quick disconnects. Its outputs include interface specifications, requirements, and implementation guidance. The engineering task extends beyond selecting a liquid-cooled server: the fluid path and its connections also need defined operating requirements.[17]
Its public workstream directory distinguishes coolant-fluid guidance, cold-plate performance work, and connector specifications. For example, the Universal Quick Disconnect v2 workstream addresses design, production, and testing gaps in the earlier specification. Draft workstream material should be distinguished from an approved release before being used as a procurement requirement.[18]
Firmware and device security
OCP's security work includes both technical designs and assessment programs. Caliptra is one example of a contributed design: the Caliptra 2.0 specification describes a silicon root of trust for measurement and identity, with subsystem functions for update and recovery. Its purpose is to establish security services inside a system-on-chip, not to certify an entire data center.[19]
The separate OCP Security Appraisal Framework and Enablement (S.A.F.E.) program organizes firmware security reviews by accredited third-party review providers. Its scope includes boot integrity, mutable storage, secure updates, and recovery. The program publishes findings reports for reviewed products. Those reports provide a specific assessment scope; a S.A.F.E. designation should not be treated as proof that a device has no vulnerabilities.[20]
Coordination across AI systems
Open Systems for AI is an umbrella strategic initiative that coordinates work across OCP projects. Its work areas cover data center physical infrastructure, IT infrastructure, and systems management. Workshops address subjects such as power and cooling, silicon, server designs, cluster interconnects, and management interfaces.[3]
This coordination concerns the infrastructure needed to operate AI systems. The initiative's name does not identify a machine-learning model, training dataset, or model-weight license. Its contributions must still be evaluated individually: an initiative's goals do not establish that every planned interface or complete system is available.[3][7]
Sustainability
OCP's Sustainability project addresses equipment reuse, lifecycle measurement, and resource reporting. Its carbon-modeling work distinguishes embodied emissions from manufacturing from operational resource use, and its carbon-disclosure work seeks consistent ways to communicate those estimates through the supply chain. Other work includes data center efficiency metrics, power telemetry, and water-efficiency metrics.[21]
These are measurement and design activities. A claim that an OCP-based installation reduces emissions needs evidence about the actual equipment, workload, power supply, and lifecycle boundaries; membership or use of an open specification is not such a measurement.[21]
Product and facility recognition
Several OCP labels refer to different objects and review processes.
| Program or status | What is being recognized or published |
|---|---|
| Approved contribution | A particular technical document, design, or other contribution accepted through the relevant review process.[9] |
| OCP Accepted or OCP Inspired | Product recognition associated with approved OCP contributions and the Foundation's product-review program.[22] |
| OCP Ready | A particular data center facility assessed against the applicable facility requirements.[23] |
| OCP S.A.F.E. | A device covered by the security review program and its findings report.[20] |
OCP Ready involves a facility self-assessment and review by the Data Center Facility project. The program considers matters such as access and logistics, building infrastructure, and connectivity. Facility readiness and product recognition answer different questions: a suitable building does not establish the conformance of every device installed in it.[23]
A supplier's OCP participation also should be distinguished from a particular product's recognition. When comparing products, the relevant evidence is the named contribution, the exact implementation, its supporting documentation, and the scope of any recognition. A general reference to "OCP hardware" does not identify those details.[22]
2026 Global Summit
As of October 4, 2026, the 2026 OCP Global Summit was scheduled for October 12-15 at the San Jose McEnery Convention Center in San Jose, California.[25][26]
NVIDIA's event program listed an October 12 keynote by Ian Buck on open infrastructure for agentic AI. It also listed sessions on 800 VDC power systems and cooling distribution units on October 13, and an MGX power rack reference design session on October 15.[26]
The program advertised an MGX ecosystem exhibit showing the roles of power delivery, cooling, and modular infrastructure in rack-scale AI systems.[26]
Research use
Published specifications can also provide reference architectures for research. A September 2026 preprint by Weqaar Janjua, Eoin O'Connell, and Mihai Penica describes Cnuas, an experimental rack-scale emulation platform based on ORv3. The authors use it to study device, network, firmware, and management interfaces.[24]
The paper explicitly separates functional emulation from physical performance: it does not establish matching hardware throughput, latency, or power behavior. Its accelerator stack and facility-modeling extension are prototypes. This is an example of research using OCP specifications, not evidence of OCP certification or production readiness.[24]
References
- ^Open Compute Project Foundation. OCP home page. Accessed September 27, 2026.
- ^1 ^2 ^3Facebook. Facebook Launches Open Compute Project. April 7, 2011.
- ^1 ^2 ^3Open Compute Project Foundation. Open Systems for AI. Accessed September 27, 2026.
- ^1 ^2 ^3Open Compute Project Foundation. OCP Contribution Taxonomy. June 10, 2026 document, carrying a draft revision entry.
- ^Open Compute Project Foundation. Community Structure. Accessed September 27, 2026.
- ^Open Compute Project Foundation. Foundation Structure. Accessed September 27, 2026.
- ^1 ^2 ^3Open Compute Project Foundation. Operating Structure and Policy. Version 3.2.2, March 21, 2025.
- ^Open Compute Project Foundation. About OCP. Accessed September 27, 2026.
- ^1 ^2 ^3Open Compute Project Foundation. How to Contribute. Accessed September 27, 2026.
- ^1 ^2Open Compute Project Foundation. Intellectual Property Rights Management Policy. August 14, 2026.
- ^Open Compute Project Foundation. Rack & Power. Accessed September 27, 2026.
- ^1 ^2Open Compute Project Foundation. Open Rack specifications and designs. Accessed September 27, 2026.
- ^1 ^2Meta Engineering. OCP Summit 2022: Open hardware for AI infrastructure. October 18, 2022.
- ^1 ^2Open Compute Project Foundation. New Open Accelerator Infrastructure (OAI) sub-project to launch within the OCP Server Project. March 19, 2019.
- ^Open Compute Project Foundation. Open Accelerator Infrastructure. Accessed September 27, 2026.
- ^Open Compute Project Foundation. Cooling Environments. Accessed September 27, 2026.
- ^Open Compute Project Foundation. Cold Plate. Accessed September 27, 2026.
- ^Open Compute Project Foundation. Cooling Environments/Cold Plate workstreams. Accessed September 27, 2026.
- ^Open Compute Project Foundation. Caliptra: A Datacenter System on a Chip (SoC) Root of Trust (RoT). Version 2.0, July 31, 2025.
- ^1 ^2Open Compute Project Foundation. OCP S.A.F.E. Program. Accessed September 27, 2026.
- ^1 ^2Open Compute Project Foundation. Sustainability. Accessed September 27, 2026.
- ^1 ^2Open Compute Project Foundation. Product Recognition Program. Accessed September 27, 2026.
- ^1 ^2Open Compute Project Foundation. About the OCP Ready Program. Accessed September 27, 2026.
- ^1 ^2Janjua, W., O'Connell, E., and Penica, M. Cnuas: A Software-Defined AI/HPC Rack-scale Emulation Platform and Hyperscale Data Center Facility Twin. arXiv:2609.15889, September 14, 2026. Preprint.
- ^Open Compute Project Foundation. Upcoming Events. Accessed October 4, 2026.
- ^1 ^2 ^3NVIDIA. NVIDIA at OCP Global Summit 2026. Accessed October 4, 2026.
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation. Every suggestion is reviewed for sourcing before it goes live.
1 revision · v2 · 2,294 words · full history
Fact-checks are independent of edits: a reviewer re-verifies the article against its sources and stamps the date. How we verify
Research and drafting on this wiki are AI-assisted, under named human editorial standards. How AI is used here
Reviewer note: Independent AI-assisted editorial review checked the published text against cited primary documentation and research. Version-specific behavior and study limitations are stated in the article; this is not a guarantee of runtime behavior or factual infallibility.
Cite this page: AI Wiki. "Open Compute Project." aiwiki.ai, updated 27 Sept 2026, fact-checked 27 Sept 2026. CC BY 4.0. https://aiwiki.ai/wiki/open_compute_project