Nvidia invests $3.5bn in MediaTek to lock in NVLink Fusion
On Monday, MediaTek became the latest chip designer to adopt Nvidia's NVLink Fusion high-speed interconnect for its own datacenter XPU (custom AI chip) line. In exchange, Nvidia is investing $3.5 billion in convertible bonds issued by MediaTek. Nvidia will also keep licensing MediaTek's own chip designs for use in future DGX and RTX Spark systems.
NVLink Fusion started as a commercial version of Nvidia's inter-GPU interconnect, introduced early last year in two forms: a chip-to-chip variant for linking CPUs to GPUs or XPUs in a memory-coherent fabric, and a switched fabric for stitching multiple accelerators into one large logical rack-scale system. It has since grown into Nvidia's general vehicle for licensing its semiconductor IP, and MediaTek is joining a list that already includes Amazon, Fujitsu, Qualcomm, Arm and Marvell. Because Nvidia's MGX rack design is part of the Open Compute Project, licensees can reportedly take an existing NVL72 rack and slot in their own compute blades, cutting the system design and mechanical engineering work needed to go from silicon to a finished AI rack. Amazon is already doing this: it used Nvidia's MGX NVL72 reference design for its Trainium3-based racks, and Trainium4, expected late this year, is expected to drop Amazon's in-house NeuronLink interconnect for NVLink Fusion.
The arrangement also keeps rivals tied to other Nvidia products. Under NVLink Fusion's original terms, a customer could pair a partner XPU with an Nvidia CPU, or a partner CPU with an Nvidia GPU, and needed Nvidia NVSwitches to get the tech's scale-up networking. On that basis, Nvidia had the potential to sell roughly two GPUs per partner CPU, one Grace or Vera CPU for every two partner XPUs, and up to nine NVSwitches for every rack of about 72 accelerators. Nvidia has since eased that restriction, which is what let Amazon pair its own Graviton CPUs with Trainium. Dion Harris, Nvidia's senior director of HPC and AI Hyperscale Infrastructure, told press on Monday that the platform lets customers mix and match across "the CPU to the GPU to the scale-up fabric to the scale-out networking across our Spectrum-X and InfiniBand, as well as our DPU," adding that Nvidia offers five different networking technologies and topologies through NVLink Fusion. Nvidia has not disclosed licensing pricing, but Harris said the company will not give the technology away for free.
Nvidia's licensing push is not limited to networking. Its quarterly networking revenue already exceeds $15 billion. Last week Nvidia detailed a new IP offering called NVHBM, a custom HBM base die with an integrated memory controller that Nvidia says frees up 25 percent of die area for compute, cuts HBM power draw by 15 percent, and increases effective bandwidth by 30 percent. Amazon's Annapurna Labs will be among the first to use NVHBM, in its Trainium chip family. Broadcom and Marvell are both reported to be developing similar base-die technology of their own.
The overall picture is a set of relationships that could turn into rivalries: Nvidia is licensing interconnect and memory IP to companies such as MediaTek and Marvell that could eventually compete with it directly on custom chip design. For now, the article frames the arrangement as a net benefit for XPU designers, since a licensed interconnect and rack design let them focus on the chip itself rather than reinventing scale-up networking from scratch, even if that costs Nvidia some switch sales.
Key facts
- MediaTek is adopting Nvidia's NVLink Fusion interconnect for its datacenter XPU line; Nvidia is investing $3.5 billion in MediaTek convertible bonds in exchange and will keep licensing MediaTek's own chip designs for future DGX and RTX Spark systems.
- Amazon, Fujitsu, Qualcomm, Arm and Marvell are also licensing NVLink Fusion. Amazon already used Nvidia's MGX NVL72 rack design for Trainium3, and Trainium4, expected late this year, is expected to drop Amazon's in-house NeuronLink interconnect for NVLink Fusion.
- Nvidia's quarterly networking revenue exceeds $15 billion. Last week it detailed NVHBM, a custom HBM base die that Nvidia says frees up 25 percent of die area for compute, cuts HBM power draw by 15 percent and boosts effective bandwidth by 30 percent; Amazon's Annapurna Labs will be among the first to use it, in Trainium chips.
- Nvidia senior director Dion Harris says Nvidia is not ready to discuss the NVLink Fusion licensing structure but confirmed there will be licensing elements in place, meaning the technology will not be free.
- Broadcom and Marvell are reported to be developing base-die technology similar to NVHBM, and the deals raise the prospect of Nvidia, Broadcom, Marvell and MediaTek becoming rivals on custom chip design even as they license Nvidia's interconnect and memory IP.
Why it matters
Nvidia's core business has been GPU sales, but a wave of custom AI chips (XPUs) from OpenAI, Meta, Microsoft, Amazon and others has raised the question of who still needs Nvidia's GPUs. Rather than losing ground to that trend, Nvidia is turning its networking IP, NVLink Fusion, and now its memory IP, NVHBM, into a licensing business. That keeps Nvidia embedded in AI datacenter builds even when a competitor's own chip sits at the center of the rack, and extends its revenue beyond GPU sales into IP licensing, NVSwitch hardware and rack design.
Who it affects
MediaTek is the new NVLink Fusion licensee, receiving a $3.5 billion Nvidia investment in its convertible bonds while continuing to have its own chip designs licensed into future Nvidia DGX and RTX Spark systems. Amazon, through its Annapurna Labs unit and Trainium chip line, is both an NVLink Fusion licensee and the first named adopter of NVHBM. Fujitsu, Qualcomm, Arm and Marvell are also named as NVLink Fusion licensees. Broadcom and Marvell appear again as rivals reportedly building competing base-die memory technology. Nvidia's Dion Harris, senior director of HPC and AI Hyperscale Infrastructure, is the company's on-record voice defending the approach.
How to use it
This is datacenter chip-design news rather than a consumer product. For teams building custom AI accelerators, licensing NVLink Fusion and using Nvidia's Open Compute Project MGX rack designs means a custom compute blade can reportedly be slotted into an existing NVL72 rack, cutting the system design and mechanical engineering work needed to go from silicon to a working AI rack. Nvidia has not published pricing or licensing terms for NVLink Fusion or for licensing MediaTek's designs into DGX and RTX Spark systems; Harris said only that the company is not yet ready to discuss the licensing structure and that it will not be free.
How solid is it
The MediaTek deal and the quoted figures come from Nvidia's own senior director, Dion Harris, speaking to press on the record on Monday, alongside The Register's reporting on prior NVLink Fusion and NVHBM announcements. Some claims are the outlet's own analysis rather than statements attributed to a named source: that Trainium4 will drop NeuronLink for NVLink Fusion, and that Broadcom and Marvell are developing base-die technology similar to NVHBM, are both presented as the article's own understanding.
Risks and caveats
Nvidia has not disclosed a licensing fee structure for NVLink Fusion, pricing for licensing MediaTek's own designs into DGX and RTX Spark systems, a launch date for MediaTek's datacenter XPU, or a commercial availability date for NVHBM beyond Annapurna Labs being an early adopter. The arrangement also carries a structural tension: Nvidia is licensing core interconnect and memory IP to companies, including MediaTek and Marvell, that could become direct competitors in custom chip design, a dynamic the article itself describes as more like frenemies than a clean partnership.
“It is safe to assume that there will be some licensing elements in place in that Nvidia NVLink Fusion is an Nvidia technology.”
— Dion Harris, Nvidia senior director of HPC and AI Hyperscale Infrastructure