As artificial intelligence models become increasingly advanced, the invisible infrastructure supporting them is becoming just as important as the models themselves. Every breakthrough in generative AI, reasoning systems, and large language models depends on one critical factor: the ability of thousands of GPUs to communicate with each other at lightning speed without interruption.
To meet this growing demand, OpenAI has introduced a groundbreaking networking innovation called Multipath Reliable Connection (MRC)—a next-generation protocol designed to accelerate large-scale AI training while improving reliability, efficiency, and resilience. Developed in collaboration with industry leaders including AMD, Broadcom, Intel, Microsoft, and NVIDIA, MRC is now being shared openly through the Open Compute Project to help the broader AI ecosystem scale more effectively.
This development marks a significant milestone in the pursuit of faster and more reliable AI supercomputing infrastructure.
Why AI Training Needs Better Networking
Training frontier AI models is not just about raw computing power. It is also about how efficiently data can move between GPUs spread across enormous supercomputer clusters.
In modern AI training environments, millions of data transfers occur during a single training step. If even one transfer arrives late, the entire system can slow down. GPUs may sit idle waiting for synchronization, wasting valuable compute cycles and increasing training time.
As AI clusters scale to tens of thousands—or even hundreds of thousands—of GPUs, networking challenges become exponentially more difficult. Common issues such as network congestion, switch failures, and packet loss can disrupt the synchronization required for large AI models to train effectively.
This challenge becomes even more critical in synchronous AI training, where thousands of GPUs must operate together in perfect coordination. A single network failure can impact the entire training process, forcing expensive restarts or delays.
Recognizing this growing bottleneck, OpenAI and its partners reimagined the architecture of AI networking from the ground up.
The Rise of MRC: A New Era of AI Networking
MRC, short for Multipath Reliable Connection, was designed to solve two major problems in AI supercomputing:
- Eliminating network congestion whenever possible
- Reducing the impact of hardware or link failures on AI training jobs
Unlike traditional networking systems that rely on a single path for each transfer, MRC distributes packets across hundreds of possible paths simultaneously. This dramatically improves efficiency, resilience, and performance consistency.
The protocol is integrated into the latest 800Gb/s networking interfaces, enabling AI systems to route traffic dynamically while reacting to failures in microseconds instead of seconds.
This means AI supercomputers can continue training large models smoothly—even when parts of the network fail.
How Traditional AI Networks Fall Short
Conventional AI networking architectures often rely on fixed routing systems where each data transfer follows one specific path. While this approach works at smaller scales, it creates severe limitations in massive AI environments.
Key Problems with Traditional Networking
- Congestion occurs when multiple transfers compete for the same network links
- A failed switch or connection can disrupt entire training jobs
- Dynamic routing protocols may take seconds to recompute paths after failures
- Large clusters require additional switch layers, increasing power consumption and operational complexity
For frontier AI systems operating at extreme scale, these inefficiencies are no longer sustainable.
This is where MRC changes the game.
Multi-Plane Networking: The Foundation of MRC
One of MRC’s most innovative concepts is the use of multi-plane networking.
Instead of treating an 800Gb/s network connection as one massive link, MRC divides it into several smaller parallel links. For example, one high-speed interface can be split into eight separate 100Gb/s paths connected to different switches.
This creates multiple independent network “planes” operating simultaneously.
The advantages are significant:
- Greater redundancy
- Improved fault tolerance
- Fewer switch tiers required
- Lower power consumption
- Reduced hardware complexity
- Better scalability for massive GPU clusters
With this architecture, AI supercomputers can connect over 100,000 GPUs using only two tiers of switches—a major improvement compared to conventional networking designs that may require three or four layers.
The result is a faster, simpler, and more energy-efficient infrastructure for AI training.
Packet Spraying: Eliminating Network Congestion
One of the most revolutionary features of MRC is a technique called packet spraying.
Traditionally, all packets belonging to a data transfer follow the same path through a network. This often creates bottlenecks when many transfers collide on the same link.
MRC takes a completely different approach.
Instead of sending data through one route, MRC spreads packets from a single transfer across hundreds of available paths simultaneously.
This creates several benefits:
- Reduced network congestion
- Better load balancing
- Faster synchronization between GPUs
- Lower latency during training
- More predictable performance
Even if packets arrive out of order, MRC can still process them correctly because each packet contains its final memory destination.
This innovation is particularly valuable for synchronous AI training workloads, where the slowest transfer can delay the entire system.
By distributing traffic intelligently, MRC minimizes hotspots and keeps GPUs operating efficiently.
Faster Recovery from Network Failures
At hyperscale, hardware failures are inevitable. Links fail. Switches malfunction. Congestion occurs.
The real challenge is not preventing every failure—it is minimizing the impact of failures when they happen.
MRC was built with this philosophy in mind.
How MRC Handles Failures
When the protocol detects packet loss or congestion:
- It immediately reroutes traffic to healthier paths
- It retransmits missing packets automatically
- It probes failed paths to check for recovery
- It dynamically balances load across the network
This process happens within microseconds.
By comparison, traditional network fabrics may require several seconds—or even tens of seconds—to stabilize after failures.
For AI supercomputers training frontier models, this speed difference is enormous.
Packet Trimming: Smarter Congestion Management
Another major innovation within MRC is packet trimming.
Normally, when a network switch becomes overloaded, it may simply drop packets entirely. This can cause systems to assume a network failure occurred.
MRC introduces a smarter solution.
Instead of dropping overloaded packets completely, switches remove only the packet payload while forwarding the header information. This alerts the destination system that retransmission is needed without falsely signaling a hardware failure.
This approach reduces unnecessary recovery actions while improving overall network stability.
Replacing Dynamic Routing with SRv6
Traditional networking systems often depend on complex dynamic routing protocols such as BGP to adapt to failures and determine packet paths.
However, dynamic routing can itself become a source of instability.
MRC simplifies this process using IPv6 Segment Routing (SRv6).
With SRv6:
- The sender directly defines the packet’s path through the network
- Switches simply follow predefined instructions
- Routing tables remain static and simpler to manage
- Failures are handled at the connection level instead of through network-wide recalculations
This dramatically reduces operational complexity while improving determinism and resilience.
In essence, the network becomes more predictable and easier to maintain at a massive scale.
Stargate and the Future of AI Infrastructure
The development of MRC is closely tied to OpenAI’s ambitious infrastructure vision, including the evolution of its large-scale supercomputing initiative known as Stargate.
As AI systems continue to expand, infrastructure efficiency becomes critical not only for performance but also for sustainability and cost optimization.
MRC is already deployed in some of OpenAI’s most advanced AI supercomputers, including systems operating with Oracle Cloud Infrastructure in Abilene, Texas, and Microsoft-powered supercomputing environments.
These deployments are helping train some of the world’s most advanced AI models using NVIDIA and Broadcom hardware.
By releasing MRC openly through the Open Compute Project, OpenAI is also encouraging industry-wide collaboration around scalable AI networking standards.
Why Open Standards Matter for AI Growth
The future of AI will depend on collaboration across hardware manufacturers, cloud providers, networking companies, and research organizations.
Open standards like MRC can help:
- Accelerate AI innovation
- Improve interoperability across platforms
- Reduce infrastructure costs
- Enable broader access to advanced AI systems
- Increase reliability across hyperscale deployments
As AI becomes a foundational technology for businesses, governments, and everyday users, the importance of robust infrastructure will only grow.
Networking may not receive the same attention as AI models themselves, but it is rapidly becoming one of the most critical layers of the AI revolution.
Outlook
The rise of large-scale AI training is reshaping how supercomputers are designed. Speed alone is no longer enough. Reliability, scalability, resilience, and efficiency have become equally essential.
With Multipath Reliable Connection (MRC), OpenAI and its technology partners are pioneering a smarter approach to AI networking—one built specifically for the demands of frontier AI systems.
By leveraging multi-plane architectures, packet spraying, adaptive routing, and SRv6-based infrastructure, MRC represents a major leap forward in how data moves across AI supercomputers.
As the AI industry pushes toward ever larger and more capable models, innovations like MRC could become the backbone of the next generation of intelligent systems.
Read more: Google AI Search Now Features Reddit & Real User Advice
