Nvidia opens NVLink to custom chips with NVLink Fusion

For the past two years, hyperscalers faced an expensive dilemma: pay Nvidia’s massive premiums for end-to-end systems, or bet the company farm on fully homegrown silicon.

At COMPUTEX on May 18, Jensen Huang decided to erase that either-or choice altogether.

Nvidia officially unveiled NVLink Fusion, a new silicon interconnect architecture that essentially lets outside custom chips tap directly into Nvidia’s gold-standard computing fabric and rack-scale infrastructure.

Custom silicon joins the club

Simply put, NVLink Fusion is an open invitation for third-party chips to speak native NVLink.

Historically, Nvidia’s proprietary interconnect has been its crown jewel. In systems like the GB200 NVL72 and GB300 NVL72, NVLink delivers a massive 1.8 TB/s of bandwidth per GPU—roughly 14x faster than standard PCIe Gen5. Replicating that bandwidth at scale has been the single biggest headache for anyone attempting custom silicon.

Now, a heavy-hitting coalition of ASIC design and silicon partners—including MediaTek, Marvell, Alchip, Astera Labs, Synopsys, and Cadence—is already building custom silicon compatible with NVLink Fusion.

It is not just for accelerator chips, either. Qualcomm and Fujitsu are plugging their own custom processors into the mix. Fujitsu confirmed that its upcoming 2nm Arm-based CPU, FUJITSU-MONAKA, will connect directly to Nvidia GPUs via NVLink Fusion to target extreme power efficiency.

To tie it all together, hyperscalers can scale out to millions of mixed chips using Nvidia’s 800Gb/s networking stack—spanning ConnectX-8 SuperNICs, Spectrum-X Ethernet, and Quantum-X800 InfiniBand switches, with co-packaged optics (CPO) coming down the line. A management software platform called NVIDIA Mission Control orchestrates the whole hybrid setup.

Why it matters

We think this is one of the smartest tactical moves Nvidia has pulled in years.

By opening up NVLink, Nvidia is voluntarily giving up full exclusivity over compute silicon. But in exchange, it guarantees that every custom accelerator or CPU built by Big Tech remains permanently anchored to Nvidia racks, Nvidia switches, and Nvidia software.

For cloud giants, this eliminates an agonizing compromise. They can keep pouring billions into specialized internal chips for inference or proprietary models, without losing access to the world’s fastest interconnect fabric or tearing down their existing server architectures.

At the same time, this puts immense pressure on competing open-fabric initiatives. When integrating with NVLink is this frictionless, the business case for building an alternative scale-up ecosystem becomes a lot harder to justify.

Basically, Nvidia saw hyperscalers rushing to design their own silicon—and decided to become the highway everyone has to drive on anyway.