What is the Open Compute Project (OCP)? Where Nvidia just opened its Vera Rubin racks — 250 kW each
The Open Compute Project (OCP) is a nonprofit that publishes data-center hardware designs, including servers, racks, power and cooling, so many companies can build to the same specifications. Facebook launched it in April 2011 by releasing its own data-center designs. At the OCP Global Summit in San Jose in October 2026, Nvidia contributed rack and compute-tray designs for its next-generation Vera Rubin AI servers through the MGX 1.2 architecture, with racks drawing up to 250 kW
The three lines
- Definition — a nonprofit sharing open data-center hardware designs, launched by Facebook in 2011
- This week — Nvidia opened Vera Rubin rack designs as MGX 1.2: up to 250 kW, liquid cooled
- Meaning — more makers can build the same racks; real adoption remains the open question
Key questions
- Open Compute Project meaning
- **A nonprofit that shares data-center hardware designs the way open-source software shares code.** | Item | Detail | |---|---| | Launched | April 2011, led by Facebook (now Meta) | | Origin | Facebook's 2009 redesign of its data centers to cut cost and power | | Covers | Servers, racks, power, cooling, networking | | Members | About 400 companies (Feb 2025) | | Main event | OCP Global Summit, San Jose |
- Nvidia Vera Rubin OCP MGX 1.2
- **Nvidia contributed Vera Rubin rack and compute-tray designs to OCP through MGX 1.2.** | Item | Detail | |---|---| | Spec | MGX 1.2, third-generation MGX rack | | Rack power | Up to 250 kW | | Cooling | Blind-mate liquid cooling | | Power distribution | Liquid-cooled 5,000 A busbar | | Joint documents | Open Data Center Specification v0.7, Grid Resilience Specification (with AMD, Google, Meta, Microsoft) |
- AI data center rack power 250kW
- **Tens of times a typical server rack.** Denser AI chips push rack power up, forcing a shift from air to liquid cooling. | Rack type | Rough power | |---|---| | Typical enterprise rack | Single digits to ~10 kW | | Grace Blackwell NVL72 generation | Around 100+ kW | | Vera Rubin MGX 1.2 | Up to 250 kW |
The Open Compute Project (OCP) is open source for data-center hardware. Just as software projects publish code, OCP publishes designs for servers, racks, power supplies and cooling, so many manufacturers can build to the same specifications. Data-center operators avoid being locked into one vendor; vendors that follow the spec can sell to the largest buyers.
The organization was at the center of AI hardware news this week. At the OCP Global Summit in San Jose in October 2026, Nvidia contributed the rack and compute-tray designs for its next-generation "Vera Rubin" AI servers — releasing core designs of its own product as an open specification.
1. How OCP started
| When | What happened |
|---|---|
| 2009 | Fast-growing Facebook redesigns its data centers from scratch to control cost and power |
| April 2011 | Facebook releases the designs and launches OCP, an early large-scale open hardware project |
| Nonprofit | Facebook, Intel, Rackspace and others join as founding members |
| February 2025 | About 400 member companies |
| October 2026 | Nvidia contributes Vera Rubin rack designs |
Facebook's reasoning was simple. Custom gear used by one buyer is expensive. Publish the design, and more manufacturers build it, more buyers buy it, and prices fall. Today Google, Microsoft, Meta and other hyperscalers participate alongside chip and equipment makers.
2. What Nvidia opened
| Item | Detail | Why it matters |
|---|---|---|
| MGX 1.2 | Third-generation MGX rack spec covering compute, CPU, storage and networking racks | Multiple manufacturers can build the same Vera Rubin rack |
| Rack power | Up to 250 kW | Tens of times a typical server rack |
| Cooling | Blind-mate liquid cooling (coolant lines connect automatically on insertion) | Air can no longer remove the heat |
| Power distribution | Liquid-cooled 5,000 A busbar | A heavy metal bar carrying enormous current inside the rack |
| Power | 800 V DC reference architecture | Fewer conversion losses |
| Optics | USFP module proposed with Cisco, 400 Gb/s per lane | Targets about 4x bandwidth density |
| Joint specs | Open Data Center Specification v0.7; Grid Resilience Specification | Published with AMD, Google, Meta and Microsoft |
The 250 kW figure is the heart of the shift. As each chip draws more power and more chips are packed into a rack, rack power climbs. That is why cooling is moving from air to liquid and power from AC to high-voltage DC.
3. Why would Nvidia give its designs away?
| What Nvidia gains | Explanation |
|---|---|
| Bigger ecosystem | More manufacturers building Vera Rubin racks means more Nvidia chips sold |
| Setting the standard | Rack, power and cooling standards for AI data centers solidify around Nvidia's design |
| Customer comfort | Large buyers are not tied to one assembler |
| Competitive edge | A broader supply chain than custom-chip rivals such as Google's TPU |
Analysts read it as unbundling the "AI factory" and turning openness into a strength: Nvidia keeps the chip, and lets everyone build the racks and plumbing around it. Whether that leads to real adoption depends on suppliers actually building to the specs. In the same week, Taiwan announced US industrial parks for the small suppliers behind TSMC and Foxconn — the kind of companies that make cooling and power parts for such racks.
4. FAQ and open questions
| Question | Answer |
|---|---|
| Can anyone use OCP specs? | Yes, they are public; membership is not required to build to them |
| Who in the supply chain cares? | Memory, power and cooling component makers design products to OCP specs |
| Link to data-center power? | As rack power rises, whole sites move to gigawatt scale |
- Adoption: No numbers yet on how many manufacturers will build Nvidia's designs.
- Targets vs results: The 4x optics density and two-second grid recovery are targets.
- Founders: Accounts of founding members differ slightly.
Sources
- Wikipedia — Open Compute Project
- OCP blog — A 10 year reflection on the Open Compute Project
- StorageNewsletter — OCP Global Summit 2026: NVIDIA Opens Vera Rubin Architecture to Help Scale AI Factories Worldwide
- Converge Digest — NVIDIA Pushes 800V DC, 400G Optics and Open Rack Designs
- Futurum Group — NVIDIA Opens Vera Rubin Racks via Open Compute Project