UALink Explained: Open AI Chip Interconnect by Google, Meta and Microsoft

Kankan Rai June 1, 2024 10 mins read

Google, Microsoft and Meta were among eight major technology companies that announced the Ultra Accelerator Link (UALink) initiative in May 2024. The group aimed to create an open industry standard for high-speed communication between AI accelerator chips inside large data-centre systems. AMD, Broadcom, Cisco, Google, Hewlett Packard Enterprise (HPE), Intel, Meta and Microsoft formed the original promoter group.

UALink has since moved beyond the announcement stage. The consortium released its first public 200G specification on 8 April 2025 and published a broader set of UALink 2.0 specifications in April 2026. The technology is intended to give data-centre builders another way to connect accelerators from a multi-vendor ecosystem, rather than relying exclusively on a single proprietary interconnect.

This guide explains what UALink does, why accelerator-to-accelerator communication matters, how the 1.0 and 2.0 specifications differ, and what the project means for AI infrastructure. It also explains what UALink does not do: it is not a replacement for every data-centre network, and a published specification is not proof that compatible products are already deployed everywhere.

What is UALink?

UALink stands for Ultra Accelerator Link. It is an open industry interconnect standard for communication between AI accelerators, such as GPUs, and switches inside an AI computing pod. The standard is designed to support low-latency, high-bandwidth data exchange and direct load and store operations between accelerator memories within a defined scale-up domain.

AI training and inference workloads often divide work across many accelerators. Those chips must exchange model parameters, intermediate results and other data. If communication is slow, accelerators can spend time waiting for data instead of performing calculations. A fast interconnect can help keep the processors working together efficiently.

Why AI data centres need faster chip-to-chip communication

Large AI models are commonly distributed across multiple accelerators because one chip may not have enough compute capacity or memory to handle the entire workload efficiently. These accelerators need to exchange information during training and inference.

  • Lower communication latency: Reduce the time accelerators wait for information from other devices.
  • Higher bandwidth: Move large amounts of data between accelerators more efficiently.
  • Memory access: Support direct load and store operations across accelerator memory within the scale-up domain.
  • Scalability: Give system designers a standard way to connect larger accelerator groups.
  • Interoperability: Create a common technical basis for products from different vendors, subject to implementation and compliance testing.

The practical goal is better utilisation of expensive AI hardware. If communication becomes a bottleneck, adding more GPUs does not necessarily produce a proportional increase in useful computing performance.

Which companies formed the original UALink group?

The original UALink promoter group was announced on 30 May 2024 by eight companies:

Company Role in the original announcement
AMD Founding promoter group member
Broadcom Founding promoter group member
Cisco Founding promoter group member
Google Founding promoter group member
Hewlett Packard Enterprise Founding promoter group member
Intel Founding promoter group member
Meta Founding promoter group member
Microsoft Founding promoter group member

The group intended to establish a consortium to develop and maintain the standard. The consortium was incorporated in October 2024. It later expanded its ecosystem: in January 2025, it announced that Alibaba, Apple and Synopsys had joined its board of directors. The consortium said the UALink community had grown to more than 85 member companies by the release of the first 1.0 specification in April 2025.

Membership has different levels, including promoter, contributor and adopter categories. Participation in a standards organisation should not be confused with proof that every member has released a commercial product using the standard.

What does UALink 1.0 provide?

The first public UALink 200G 1.0 specification was released on 8 April 2025. It defines a low-latency, high-bandwidth connection between accelerators and switches in AI computing pods. The specification supports a 200G-per-lane scale-up connection and a pod containing up to 1,024 accelerators.

UALink 1.0 item Published capability
Interconnect type Scale-up communication between accelerators and switches
Per-lane data rate 200G
Maximum scale-up domain Up to 1,024 accelerators in an AI computing pod
Memory operations Direct load and store operations within the scale-up domain
Physical medium in 200G 1.0 Copper-based links, with distance limitations unless repeaters are used

The 1,024-accelerator figure describes the capability defined by the standard. It does not mean that every UALink product or data centre already connects 1,024 accelerators in a working deployment. System architecture, switching, cabling, power, cooling, software and product implementation all affect what can be built in practice.

What changed with UALink 2.0 in April 2026?

On 7 April 2026, the UALink Consortium announced a set of four specifications that extend the original standard. The additions focus on in-network compute, physical-layer flexibility, chiplet integration and system management.

Specification Purpose
UALink Common 2.0 Introduces in-network compute, allowing computation to be performed as part of communication between accelerators.
200G Data Link and Physical Layers 2.0 Separates the data-link and physical-layer specification so link technologies and speeds can evolve without requiring changes to the common specification.
Chiplet 1.01 Defines interfaces, form factors, flow control and management information needed to integrate UALink into chiplet-based systems; the consortium says it aligns with UCIe 3.0.
Manageability 1.0 Defines centralised management approaches and uses established protocols and APIs such as gNMI, YANG, SAI and Redfish.

In-network compute is intended to reduce the amount of data that needs to move between devices for certain operations. That can save bandwidth and improve efficiency for distributed training and inference, depending on how the system and workload are implemented.

UALink and Ultra Ethernet: what is the difference?

UALink and the Ultra Ethernet Consortium (UEC) address related but different parts of AI infrastructure. UALink focuses on scale-up networking: high-performance communication among accelerators inside a computing pod. Ultra Ethernet primarily focuses on scale-out networking: connecting systems, racks and clusters over an Ethernet-based fabric.

Area UALink Ultra Ethernet
Primary focus Scale-up accelerator interconnect Scale-out data-centre networking
Typical communication Accelerator-to-accelerator and accelerator-to-switch inside an AI pod Server-to-server and system-to-system communication across a cluster
Main objective Low latency and high bandwidth within a tightly coupled accelerator domain Efficient and scalable networking across larger distributed systems
Relationship Can complement a wider data-centre network Can complement UALink in a larger AI infrastructure design

The two standards are not direct substitutes. A large AI data centre may use a scale-up interconnect within accelerator pods and a scale-out network to connect those pods to other systems and storage.

How does UALink compare with NVIDIA NVLink?

NVIDIA’s NVLink is a proprietary high-speed interconnect family used to connect NVIDIA GPUs and related system components. UALink is being developed as an open industry standard intended to enable a broader multi-vendor ecosystem.

That difference can matter to data-centre operators seeking flexibility in accelerator selection and system design. However, an open standard does not automatically make products interchangeable. Vendors still need compatible implementations, suitable switches and software, validation, compliance and interoperability testing.

Feature UALink NVIDIA NVLink
Governance Developed through an industry consortium Developed and controlled by NVIDIA
Design goal Multi-vendor accelerator interoperability High-performance connectivity in NVIDIA GPU platforms
Ecosystem Consortium members and technology partners NVIDIA GPUs, switches, systems and software ecosystem
Deployment status Specifications published; product development and deployment are progressing Commercial products available across multiple generations

NVIDIA was not part of the original eight-company UALink promoter group. That is a factual description of the 2024 announcement; it does not by itself establish why NVIDIA chose not to participate. AWS was also not among the original promoter companies. Avoid interpreting the founding membership list as a permanent list of all future UALink participants or as proof of a specific company’s business strategy.

When will UALink products reach data centres?

Publishing a specification is an important milestone, but hardware suppliers must still develop compatible accelerator interfaces, switches, system designs, testing tools and software support. The UALink Consortium says products using an open standard typically reach the market one to two years after the initial specification release. Its March 2026 roadmap described development activity across IP, accelerators, switches, test equipment and management solutions, with commercial deployments targeted through 2026 and 2027.

In April 2026, the consortium also highlighted a first UALink switch and IP product design and prototype verification milestone by Netforward. Such milestones demonstrate ecosystem progress, but they should not be described as widespread commercial adoption across major cloud data centres.

What is planned for UALink 3.0?

In September 2026, the UALink Consortium said work was underway to scope the UALink 3.0 specification. The areas being explored include a 400G data rate with optical support, optical interconnects, end-to-end resiliency and richer management features.

These are development focus areas, not a final, released UALink 3.0 specification. The consortium has not published a confirmed release timetable for the next specification iteration. Its current FAQ says future changes will depend on industry requirements and the work of its technical groups.

Who could benefit from UALink?

  • Cloud providers: More choice when designing large AI training and inference systems.
  • AI chip developers: A common interconnect standard for building accelerator products for multi-vendor systems.
  • Server and switch manufacturers: A published interface around which compatible hardware can be developed.
  • Enterprise data centres: Potential flexibility in selecting accelerators and infrastructure as products become available.
  • AI researchers and operators: The possibility of more efficient large-scale computing systems when compatible hardware and software are deployed.

The benefits will depend on real-world performance, total system cost, software maturity, power efficiency, product availability and interoperability. Buyers should compare complete systems rather than selecting hardware based solely on the presence of a UALink logo or specification reference.

Frequently asked questions

What does UALink stand for?
UALink stands for Ultra Accelerator Link. It is an open industry standard for high-speed communication between accelerators and switches in AI computing pods.

Which companies founded UALink?
AMD, Broadcom, Cisco, Google, Hewlett Packard Enterprise, Intel, Meta and Microsoft announced the original promoter group in May 2024.

When was UALink 1.0 released?
The first public UALink 200G 1.0 specification was released on 8 April 2025, later than the original 2024 announcement’s expected schedule.

How many AI accelerators can UALink 1.0 connect?
The specification supports a scale-up domain of up to 1,024 accelerators in an AI computing pod. Actual configurations depend on implementation.

Is UALink an alternative to NVIDIA NVLink?
It is an alternative open-standard approach to accelerator interconnects. NVLink is NVIDIA’s proprietary technology. Actual competition depends on compatible products, performance, software and commercial deployments.

Is UALink the same as Ultra Ethernet?
No. UALink focuses on scale-up communication inside AI accelerator pods, while Ultra Ethernet focuses primarily on scale-out networking between systems and across clusters.

Is UALink 3.0 available?
No public 3.0 release timetable has been announced. The consortium said in September 2026 that it was scoping future work, including 400G data rates, optical links and greater resiliency.

Does UALink guarantee that chips from different companies will work together?
No. A common standard is designed to support interoperability, but compatible products still require compliant implementations, testing, software and system-level integration.

Official sources and further reading

Editorial note: This article distinguishes the original 2024 promoter announcement, the published UALink 1.0 and 2.0 specifications, and future features still under development. Consortium specifications and statements were checked on 20 November 2026. Actual product availability and performance depend on vendor implementations and deployments.