Tech giants unite against Nvidia with new UALink chip standard

If you want to stitch a thousand GPUs into a single AI brain, the bottleneck isn't raw computing power—it's how fast those chips can talk to each other.

Right now, Nvidia controls the roads, the toll booths, and the speed limits. And Big Tech is finally trying to build a bypass.

Intel, Google, Microsoft, Meta, AMD, Broadcom, Cisco, and HPE just formed the Ultra Accelerator Link (UALink) Promoter Group. Their goal is straightforward: create an open industry standard to connect AI accelerators inside data centers.

Stitching 1,024 GPUs together

The group's first proposal, UALink 1.0, is designed to link up to 1,024 AI accelerators—specifically GPUs for now—across a single server pod.

Built on open specs including AMD's Infinity Fabric, UALink enables direct memory reads and writes between accelerators. The coalition promises this will slash latency and dramatically boost data transfer speeds compared to existing generic standards.

A formal UALink Consortium will launch in Q3 to hand out the 1.0 spec to member companies, followed quickly by a higher-bandwidth UALink 1.1 in Q4. However, don't expect hardware tomorrow: AMD data center head Forrest Norrod noted the first commercial products won't arrive for a couple of years.

The two elephants not in the room

The most telling detail of the announcement is who didn't sign up: Nvidia and AWS.

Nvidia controls an estimated 80% to 95% of the AI accelerator market. Its data center revenue surged over 400% last quarter, and it already has its own proprietary interconnect technology. Put simply, Nvidia is winning by a mile and has zero incentive to play nice with rivals.

Amazon Web Services, the lone absent cloud titan, seems content playing both sides while it quietly iterates on its own custom silicon.

For Microsoft, Meta, and Google, however, the math is urgent. They are spending tens of billions buying Nvidia GPUs while simultaneously designing in-house chips like Google TPUs and Microsoft Maia. Microsoft and OpenAI are even reportedly weighing a $100 billion supercomputer. They desperately need a universal cable to wire up whatever custom hardware comes next.

Why it matters

For enterprise buyers and cloud customers, there is nothing to buy today. But UALink could reshape the bill you pay for compute three years down the road.

By taking aim at the interconnect layer, the coalition is attacking Nvidia's real moat. It's easy to build a fast chip, but it's brutally hard to make a thousand of them act like one machine without Nvidia's proprietary networking.

We think rallying behind an open standard backed by AMD and Broadcom is the right strategic move, yet it carries real risks. Hardware consortiums move notoriously slow. By the time UALink chips hit production in a couple of years, Nvidia will already be two generations ahead in proprietary interconnect speed.