Hardware/Explainer

What Is UALink? The Open Standard Taking On Nvidia's NVLink

An explainer on UALink, the open scale-up interconnect backed by AMD, Intel, Google, Microsoft and Meta to link AI accelerator GPUs inside a rack, and the Nvidia NVLink lead it is still chasing.

The Ultra Accelerator Link (UALink) and UALink Consortium logos
The Ultra Accelerator Link and UALink Consortium logos. Image: UALink Consortium.

UALink, short for Ultra Accelerator Link, is an open interconnect standard built to link AI accelerator GPUs directly to one another inside a server rack, so hundreds of chips can share memory and work as one giant accelerator rather than many separate ones. It exists because Nvidia's proprietary NVLink currently owns that job in almost every AI data center, and a group of Nvidia's biggest rivals and customers — AMD, Intel, Google, Microsoft, Meta, Broadcom and others — decided the industry needed an open, royalty-bearing-free alternative they could all build to instead of depending on one vendor's closed fabric.

UALink is a specification, not a product. It defines how AI accelerators — GPUs, custom AI chips, and the switches that connect them — talk to each other with direct load, store and atomic memory operations across a rack-scale "pod" of up to 1,024 accelerators. The goal is low-latency, high-bandwidth, cache-coherent-style communication so a training or inference job can treat a whole rack of chips as one pool of compute, rather than stitching them together with slower, general-purpose networking.

The standard is developed and published by the UALink Consortium, a member-run industry body, not by any single chipmaker. That matters because the entire pitch of UALink is that no one company controls the roadmap or the licensing terms — the opposite of how Nvidia's NVLink works today.

UALink at a glance
  • What it is: Open scale-up interconnect spec for AI accelerator-to-accelerator links
  • Who runs it: UALink Consortium, incorporated October 29, 2024
  • Backers: AMD, Intel, Google, Microsoft, Meta, AWS, Cisco, HPE, Astera Labs, Alibaba, Apple, Synopsys and others
  • Scale: Up to 1,024 accelerators in one pod
  • Latest spec: UALink 2.0 (Common Spec + 200G Data Link/Physical Layer), published April 7, 2026
  • Rival: Nvidia's proprietary NVLink
  • Shipping hardware: None confirmed in volume as of November 2026, per trade-press reporting

The problem it solves: GPU-to-GPU bandwidth inside AI pods

Modern AI training runs don't fit on one GPU, or even one server. A large model gets split across dozens or hundreds of accelerators that constantly need to exchange activations, gradients and parameters. The connection between those chips — not the chips themselves — is often what limits how fast a cluster can actually train or serve a model, because AI data centers are already straining power and infrastructure budgets trying to keep ever-larger GPU pods fed with data at low latency.

Inside a single server, accelerators have long used proprietary short-reach links — Nvidia's NVLink, AMD's Infinity Fabric — to get that job done. The trouble, from the perspective of UALink's backers, is that none of those links are open across vendors: an AMD GPU can't scale up over NVLink, and a non-Nvidia switch can't natively join an NVLink domain. As AI pods grow from 8 GPUs to dozens or hundreds, cloud providers and chipmakers that don't want to standardize on one company's silicon need a shared, multi-vendor way to wire accelerators together. That's the gap UALink is designed to fill: an interconnect any member can implement in silicon, so a rack can mix accelerators, switches and NICs from different suppliers and still scale as one coherent pod.

UALink defines a layered protocol stack — a protocol layer (UALink Protocol Layer Interface), a transaction layer, a data link layer and a physical layer — purpose-built for accelerator memory semantics rather than general file or packet networking. The first public specification, UALink 200G 1.0, was based on the IEEE P802.3dj 200 Gb/s PHY layer and supported pod sizes up to 1,024 accelerators addressed through dedicated UALink switches, according to the consortium's own published specification summary. Hosts still reach accelerators through existing links like PCIe or CXL; UALink's job starts at the accelerator-to-accelerator hop, where it offers direct loads, stores and atomics with a thinner protocol stack than general Ethernet or InfiniBand fabrics.

InterconnectPrimary jobTopologyGovernance
UALinkAccelerator-to-accelerator scale-up inside a podSwitched, up to 1,024 acceleratorsOpen, multi-vendor (UALink Consortium)
Nvidia NVLink / NVSwitchAccelerator-to-accelerator scale-up inside a podSwitched, vendor-defined pod sizes (e.g. NVL72)Proprietary, single-vendor (Nvidia)
PCIeGeneral host-to-device I/O (CPU to GPU, storage, NICs)Point-to-point / tree, broad ecosystemOpen, multi-vendor (PCI-SIG)
CXLMemory expansion and pooling, cache-coherent host-device linksPoint-to-point, switchable with CXL 3.xOpen, multi-vendor (CXL Consortium)

NVLink is Nvidia's own high-speed interconnect for linking GPUs within and across servers, paired with its NVSwitch chips to build large coherent domains like the NVL72 rack. It's mature, it ships today in Nvidia's own systems, and it's deeply tied to Nvidia's CUDA software stack — which is exactly what UALink's backers object to. If you want the scale-up bandwidth NVLink provides, Nvidia's ecosystem has historically been the only practical route, reinforcing the GPU lock-in that has fueled Nvidia's dominant position in the AI accelerator market.

UALink is explicitly designed as an open, cross-vendor substitute for that role: any member can build UALink-compliant accelerators, switches or retimers without paying Nvidia or depending on its roadmap. This piece is not a full spec-by-spec NVLink comparison — the short version is that UALink is chasing NVLink's job, not yet matching its market position. As the trade press covering the 2.0 release has noted, UALink remains, in the words of connectivity analyst David Harold cited by Network World, "quite a way behind NVLink today," even as it offers something NVLink structurally cannot: a shared standard multiple competing chipmakers can build to.

PCIe and CXL solve adjacent but different problems. PCIe is the general-purpose bus that connects a CPU host to GPUs, storage and network cards — it's ubiquitous, but it wasn't designed for the all-to-all, low-latency memory semantics that dense GPU-to-GPU training traffic needs. CXL extends that world with cache coherency and memory pooling between hosts and devices, which is valuable for disaggregating and sharing memory across a server, but it's still fundamentally a host-centric link rather than a dedicated accelerator-to-accelerator scale-up fabric. UALink is scoped narrowly to that last piece: fast, direct accelerator-to-accelerator communication inside a pod, working alongside PCIe and CXL rather than replacing them.

The Ultra Accelerator Link Consortium formally incorporated on October 29, 2024, after AMD, Broadcom, Cisco, Google, HPE, Intel, Meta and Microsoft first announced the effort as a promoter group earlier that year. Its board has since grown to include AWS and Astera Labs among the founding directors, plus Alibaba, Apple and Synopsys, who joined the board in January 2025. The consortium describes its mission as uniting the industry around "an open standard... for AI scale-up networking," according to its own About UALink page, and it has reported dozens of additional contributor members joining since launch, spanning chipmakers, cloud providers, IP vendors and test-equipment companies.

That breadth is the whole point: UALink's backers are companies that either compete directly with Nvidia in AI silicon (AMD, Intel's various accelerator efforts, custom cloud chips from Google, Microsoft, Amazon and Meta) or sell infrastructure into AI data centers and don't want to be boxed into one proprietary fabric. The consortium publishes its specifications, white papers and member news through its own press room.

On April 7, 2026, the UALink Consortium published its next round of specifications, which the industry has generally shorthanded as "UALink 2.0." Per the consortium's own release, reported in detail by HPCwire, the update actually ships as four documents:

  • UALink Common Specification 2.0 — adds in-network compute, letting the fabric perform some computation alongside data movement to cut latency and bandwidth use for distributed training and inference.
  • UALink 200G Data Link and Physical Layers (DL/PL) Specification 2.0 — splits the physical and data-link layers out of the Common Specification, so the consortium can update speeds and PHYs independently of the rest of the protocol stack.
  • UALink Manageability Specification 1.0 — defines centralized control and management for UALink fabrics using standard protocols: gNMI, YANG, SAI and Redfish.
  • UALink Chiplet Specification 1.0 — specifies how to integrate UALink into chiplet-based system-on-chip designs, aligned with the UCIe 3.0 chiplet interconnect standard so UALink links can be built as die-to-die chiplets rather than only board-level connections.

UALink Consortium board chair Kurtis Bowman said the release would "enable the industry to quickly and efficiently integrate UALink solutions" into upcoming architectures, per the consortium's statement reported by HPCwire. Member companies, including AMD, filed their own statements of support for the release, published in the consortium's 2.0 Statement of Support.

DateMilestone
May 2024AMD, Broadcom, Cisco, Google, HPE, Intel, Meta and Microsoft announce the UALink promoter group
October 29, 2024UALink Consortium formally incorporates; opens contributor membership
January 2025Alibaba, Apple and Synopsys join the board of directors
April 2025UALink 200G 1.0 specification published publicly; up to 1,024 accelerators per pod
April 7, 2026UALink 2.0 Common Spec, 200G DL/PL 2.0, Manageability 1.0 and Chiplet 1.0 published

Real-world status: a spec ahead of silicon

It's worth being direct about where UALink actually stands: as of this writing, no UALink-based accelerator or switch has reached broad commercial availability. Trade press covering the 2.0 release pointed out that the consortium published its second major specification update before hardware built on the first one had even shipped — reporting pegged UALink 1.0-based silicon as unlikely to appear in meaningful volume until the second half of 2026, with commercial systems more plausibly arriving in 2027. Nvidia, meanwhile, ships NVLink-based systems today, and its lead extends beyond open alternatives like UALink to other proprietary and Ethernet-based scale-up approaches as well, according to analyst commentary cited by Network World.

That gap hasn't stopped momentum on paper: member companies including AMD have outlined roadmaps pairing UALink with upcoming accelerator lines, and infrastructure partners have begun building switch and retimer silicon aimed at the spec. But a standard is only as useful as the chips and switches that implement it, and UALink's backers are explicitly racing to close a lead Nvidia built over nearly two decades of NVLink generations. Reaching the kind of switching fabric Nvidia already sells in volume — the same scale-up demand that's pushing data centers toward technologies like co-packaged optics to keep up with rack-scale bandwidth growth — will likely take at least another product cycle.

What's next for UALink

The UALink Consortium has said interoperability and compliance testing programs are coming later in 2026, which would let multiple vendors' chips be certified to actually work together — a necessary step before UALink gear can be deployed with confidence at data-center scale. Watch for three signals over the next year: whether named UALink-compliant switch silicon actually ships from vendors beyond prototypes and demos, whether major cloud providers commit specific next-generation custom accelerators to UALink rather than proprietary or Ethernet-based scale-up links, and whether the consortium's compliance program produces real multi-vendor interoperability rather than specs that stay largely academic. Until accelerators, switches and software built on UALink 2.0 are actually running in production racks, it remains an ambitious open standard chasing a well-entrenched incumbent rather than a proven alternative to it.

Frequently asked questions

What does UALink stand for?

UALink stands for Ultra Accelerator Link. It is an open specification, developed by the UALink Consortium, for connecting AI accelerator chips like GPUs directly to one another inside a server rack.

Who is behind UALink?

UALink is developed by the UALink Consortium, which formally incorporated on October 29, 2024. Its board includes AMD, AWS, Astera Labs, Cisco, Google, HPE, Intel, Meta and Microsoft, with Alibaba, Apple and Synopsys joining in January 2025, plus dozens of additional contributor members.

How is UALink different from Nvidia's NVLink?

NVLink is Nvidia's proprietary interconnect for linking its own GPUs, tied to Nvidia's hardware and software stack. UALink is an open, multi-vendor alternative meant to do a similar job, accelerator-to-accelerator scale-up links, but across chips from different companies. As of the 2.0 spec release, UALink is still well behind NVLink's real-world deployment, per trade-press analyst commentary.

Is UALink the same as PCIe or CXL?

No. PCIe is a general-purpose bus connecting hosts to devices like GPUs, storage and NICs. CXL adds cache coherency and memory pooling between hosts and devices. UALink is scoped specifically to fast, direct accelerator-to-accelerator communication inside a rack-scale pod, and works alongside PCIe and CXL rather than replacing them.

What is new in the UALink 2.0 specification released in April 2026?

Published April 7, 2026, the update covers four documents: UALink Common Specification 2.0 (adding in-network compute), the UALink 200G Data Link and Physical Layers Specification 2.0 (now split from the Common Spec), a Manageability Specification 1.0 (centralized management via gNMI, YANG, SAI and Redfish), and a Chiplet Specification 1.0 aligned with UCIe 3.0.

Is UALink hardware available to buy yet?

No. As of this writing, no UALink-based accelerators or switches have reached broad commercial availability. Trade-press reporting on the 2.0 release noted that silicon based on the original UALink 1.0 specification was not expected in meaningful volume until the second half of 2026, with commercial systems more likely arriving in 2027.

Sources

More on UALink →UALinkNVLinkAI acceleratorsinterconnectsAMDNvidia
Mara Lindqvist
Written byMara Lindqvist

Mara Lindqvist edits the hardware desk. She covers graphics cards, processors, memory and storage, the foundries and chip designers behind them, and what the numbers on a spec sheet mean for people choosing a PC. Specifications in her stories come from manufacturer spec pages and datasheets.

More from Hardware

See all