OCP AI Data Center Announcements: Intel, AMD, Nvidia
The biggest AI data center announcements at the recent OCP (Open Compute Project) Global Summit centered on three things: liquid cooling standards, high-density power delivery, and open networking for AI clusters. Intel, AMD, and Nvidia each used the event to push their specific visions for how to build and scale AI infrastructure, with a clear shift away from traditional air-cooled, 10-20 kW per rack designs toward 100 kW+ liquid-cooled racks and co-packaged optics.
Background: Why OCP Matters for AI Data Centers
OCP started in 2011 as Facebook’s effort to share its custom data center designs. It has since grown into a de facto standards body for hyperscale and enterprise data centers. For AI, OCP is where the industry agrees on mechanical, electrical, and networking specifications that let different vendors’ gear work together in the same rack. That matters because AI clusters are no longer built from a single vendor’s proprietary stack—they mix GPUs, accelerators, switches, power shelves, and cooling loops from multiple suppliers.
The 2024 OCP Global Summit (held in San Jose, CA) was notable because all three major chip vendors—Intel, AMD, and Nvidia—used it to announce or detail hardware that targets the same problem: how to pack more compute into a building without melting it. The common thread is that AI training clusters now need 10 to 20 times more power per rack than a typical enterprise server rack, and air cooling tops out around 30-40 kW per rack. Liquid cooling is the only practical way to go beyond that.
Intel’s OCP Announcements: Gaudi 3 and Open Liquid Cooling
Intel’s OCP presence focused on two areas: its Gaudi 3 AI accelerator and its push for open liquid cooling standards. Intel announced that its Gaudi 3 accelerators, which compete with Nvidia’s H100/H200, would be available in an OCP-compliant baseboard form factor. This means the accelerator can slot into the same open chassis designs that hyperscalers already use, rather than requiring a proprietary server.
More significantly, Intel detailed its participation in the OCP’s new “Open Liquid Cooling” initiative. Intel is advocating for a standard cold plate interface so that cooling loops from different suppliers can attach to Gaudi 3 and future Intel accelerators without custom adapters. The company also showed a reference design for a 100 kW+ liquid-cooled rack using its Gaudi 3 processors, with the cooling distribution unit (CDU) located in the rack itself rather than in a separate room. This is a departure from the older chilled-water plant model and is aimed at making liquid cooling easier to deploy in existing data centers.
AMD’s OCP Announcements: MI300X, Open Networking, and Power
AMD’s OCP announcements centered on the MI300X accelerator and its “Infinity Fabric” interconnect. AMD announced that the MI300X platform now supports the OCP’s Open Accelerator Module (OAM) baseboard specification, which is the same open form factor used by many custom AI chips. This is a direct challenge to Nvidia’s proprietary NVLink and HGX baseboards, which require Nvidia’s own server designs.
AMD also highlighted its work on open networking for AI clusters. The company demonstrated a 400G and 800G Ethernet solution for AI workloads based on the OCP’s “Open Network” specifications, arguing that Ethernet is a viable alternative to Nvidia’s InfiniBand for AI training. AMD’s pitch is that Ethernet is more cost-effective and easier to manage at scale, especially for clusters that mix accelerators from multiple vendors.
On the power side, AMD showed a reference design for a 132 kW liquid-cooled rack using MI300X accelerators. The design uses a rear-door heat exchanger plus cold plates, and it includes a busbar-based power distribution system that delivers 48V DC directly to the compute nodes, reducing conversion losses compared to the older 12V architecture.
Nvidia’s OCP Announcements: NVL72, Co-Packaged Optics, and 800V
Nvidia’s OCP announcements were the most aggressive. The company detailed its NVL72 rack-scale system, which pairs 72 of its next-generation Blackwell GPUs with 36 Grace CPUs in a single liquid-cooled rack. The NVL72 is not just a server—it’s a complete rack-scale compute unit with integrated power and cooling. Nvidia announced that the NVL72 will support the OCP’s new “Open Rack v3” specification, which defines a wider rack (600mm) and higher power capacity (up to 200 kW per rack).
Nvidia also pushed its co-packaged optics (CPO) technology, which integrates optical transceivers directly onto the GPU package rather than using pluggable modules. At OCP, Nvidia announced that its next-generation Spectrum-X Ethernet switch would support CPO, enabling 1.6Tbps per port while reducing power consumption by about 30% compared to pluggable optics. This is critical because in a large AI cluster, the networking gear can consume as much power as the GPUs themselves.
Finally, Nvidia announced that its GB200 NVL72 rack would support an 800V DC power architecture, which is a significant jump from the standard 48V. The higher voltage reduces copper losses and allows thinner, cheaper power cables. Nvidia is working with OCP to standardize this 800V DC bus so that other vendors can build compatible power shelves and battery backup units.
How to Choose Between the Three Approaches
For a data center operator planning an AI cluster, the choice between Intel, AMD, and Nvidia at OCP is not just about chip performance—it’s about the ecosystem around the chip. Here are the key factors to consider:
- Software ecosystem: Nvidia’s CUDA is the most mature and widely used AI software stack. AMD’s ROCm is improving but still has gaps in some frameworks. Intel’s oneAPI is a distant third in terms of community support.
- Interconnect: Nvidia’s NVLink and InfiniBand offer the lowest latency and highest bandwidth for tightly coupled training. AMD’s Infinity Fabric and Ethernet are more open but may require more tuning. Intel’s Gaudi 3 uses standard Ethernet, which is the easiest to integrate into existing networks.
- Power and cooling: All three vendors are pushing liquid cooling, but Nvidia’s NVL72 requires it—there is no air-cooled version. AMD and Intel offer air-cooled options for smaller clusters, which can be easier to deploy in existing facilities.
- Openness: If you want to mix vendors or use white-box hardware, AMD and Intel are more aligned with OCP’s open philosophy. Nvidia’s rack-scale systems are more integrated and less modular.
Common Problems and Misconceptions
One common misconception is that OCP announcements mean the hardware is immediately available. In practice, OCP specifications are often ahead of actual product availability by 6-12 months. For example, Nvidia’s NVL72 was announced at OCP in October 2024, but initial shipments were expected in mid-2025.
Another issue is that “liquid cooling” is not a single solution. There are at least three types:
- Cold plates: A metal plate sits on the chip, and liquid flows through channels inside it. This is the most efficient for high-density chips.
- Immersion cooling: The entire server is submerged in a dielectric fluid. This is simpler but harder to service and heavier.
- Rear-door heat exchangers: A radiator on the back of the rack cools the air that has passed over the servers. This is a retrofit option but less effective at very high densities.
Many operators assume that switching to liquid cooling saves water. That is only true if you use a closed-loop system with a dry cooler. If you use a cooling tower, you can consume more water than a traditional data center. The OCP specifications include guidance on water usage effectiveness (WUE), but it is not yet a mandatory metric.
Cost and Practical Considerations
The cost of deploying an OCP-compliant AI rack is significantly higher than a traditional rack, but the cost per FLOP (floating point operation) can be lower. Here is a rough breakdown of the added costs:
- Liquid cooling infrastructure: A CDU (coolant distribution unit) costs $50,000-$100,000 per rack, plus piping and manifolds. For a 100-rack cluster, that is $5-10 million in cooling alone.
- Power delivery: Moving from 48V to 800V requires new power shelves, busbars, and battery backup units. The power infrastructure can be 20-30% of the total data center build cost.
- Networking: Co-packaged optics and 800G switches are more expensive per port than older 100G gear, but they reduce the number of switches needed.
However, the total cost of ownership (TCO) can be lower because liquid cooling reduces fan power (which can be 10-20% of a data center’s energy bill) and allows higher compute density, which means you need less floor space. The OCP publishes TCO models that operators can use to compare scenarios, but the actual numbers depend heavily on local electricity prices and climate.
Frequently Asked Questions
What does OCP stand for, and why is it relevant to AI?
OCP stands for Open Compute Project, a community that shares data center hardware designs. It is relevant to AI because AI clusters need custom power, cooling, and networking that standard server racks cannot provide. OCP specifications let different vendors build compatible components so operators are not locked into a single supplier.
Will Nvidia’s NVL72 work in an existing data center?
Probably not without major renovations. The NVL72 requires liquid cooling and 800V power, which most existing data centers do not have. You would need to install a CDU, upgrade the electrical distribution, and possibly reinforce the floor to handle the weight. Some operators are retrofitting older facilities with rear-door heat exchangers, but that only works for lower-density racks.
Is AMD’s ROCm as good as Nvidia’s CUDA for AI?
ROCm has improved significantly and now supports most major AI frameworks like PyTorch and TensorFlow. However, some niche libraries and tools are still CUDA-only. For production workloads, the gap is narrowing, but CUDA remains the safer choice if you need maximum compatibility with third-party code.
What is co-packaged optics, and why does it matter?
Co-packaged optics (CPO) puts optical transceivers on the same package as the switch chip or GPU. This reduces the distance electrical signals travel, which cuts power use and increases bandwidth. It matters because in a large AI cluster, networking can account for 20-30% of total power, so reducing that is critical for scaling.
Can I mix Intel, AMD, and Nvidia accelerators in the same rack?
Yes, but it is not trivial. The OCP baseboard specifications (OAM and the new UBB) are designed to be vendor-neutral, so you can physically put different accelerators in the same chassis. However, the software stack (CUDA vs. ROCm vs. oneAPI) and the interconnect (NVLink vs. Ethernet) may not work together seamlessly. For most workloads, it is easier to keep clusters homogeneous.
What is the difference between OCP and the Open RAN or other open standards?
OCP focuses on data center hardware (servers, racks, power, cooling). Open RAN is about telecom radio access networks. They are separate initiatives, though both use the “open” philosophy. OCP has a telecom project group, but the core specifications are for general-purpose data centers.
Related
Method: the standard formula for this calculation, rounded as each result states. See how the calculators work or report an error.