# NVIDIA NVLink Fusion: What the $3.5B MediaTek Deal Really Changes

> Source: <https://byteiota.com/nvidia-nvlink-fusion-what-the-3-5b-mediatek-deal-really-changes/>
> Published: 2026-09-08 08:10:54+00:00

NVIDIA put $3.5 billion into MediaTek last week. Most headlines called it a chip deal. It isn’t — it’s NVIDIA’s most explicit move yet to become the connective tissue of every AI factory in the world, including the ones running someone else’s chips. The mechanism is called **NVLink Fusion**, and if you’re building or deploying AI infrastructure, this is what the next five years of compute architecture looks like.

## What NVLink Fusion Actually Does

Standard NVLink connects NVIDIA GPU to NVIDIA GPU. NVLink Fusion extends that same sixth-generation NVLink protocol to chips NVIDIA didn’t design. Custom ASICs and XPUs connect via a UCIe bridge chiplet that translates between the Universal Chiplet Interconnect Express standard and the NVLink protocol. Custom CPUs connect via NVLink-C2C, a direct chip-to-chip interface that’s up to 6x more energy efficient than PCIe.

The performance numbers are serious. A single NVL72 domain — 72 XPUs, all connected all-to-all — delivers 3.6 TB/s per XPU and 260 TB/s of total scale-up bandwidth. Packet rates are 10x higher than off-the-shelf Ethernet alternatives. NVIDIA’s roadmap extends this to 1,152 accelerators per domain.

The business implication is bigger than the bandwidth number. Building a custom AI chip used to mean rebuilding the entire surrounding infrastructure: rack design, cooling, power delivery, networking, validation. NVLink Fusion removes that cost. Custom silicon teams focus on differentiated compute; NVIDIA provides everything else. As Dion Harris, NVIDIA’s Senior Director of Product Marketing, put it: “NVIDIA is an AI infrastructure company. We expanded beyond pure computing chips years ago.”

## AWS Already Committed — And So Have You

The clearest proof that NVLink Fusion is production-ready, not vaporware, is AWS. [Trainium4](https://developer.nvidia.com/blog/aws-integrates-ai-infrastructure-with-nvidia-nvlink-fusion-for-trainium4-deployment/) — Amazon’s next-generation custom AI accelerator — will share the same rack footprint, cooling system, and power distribution as NVIDIA GPU deployments, connected via NVLink 6 Switches with 400G custom SerDes. Trainium4 chips, Graviton CPUs, Elastic Fabric Adapters, and the Nitro virtualization layer all live in the same MGX rack design NVIDIA uses for its own hardware.

This matters for AWS developers in a concrete way: you’re increasingly going to run AI workloads on custom silicon without knowing it. AWS abstracts this through the Neuron SDK. The API surface looks identical; the chip underneath is optimized for specific workloads and, eventually, cheaper per token. NVIDIA’s interconnect runs the whole thing regardless.

AWS isn’t the only one. [Marvell joined the NVLink Fusion ecosystem in March 2026](https://investor.nvidia.com/news/press-release-details/2026/NVIDIA-AI-Ecosystem-Expands-as-Marvell-Joins-Forces-Through-NVLink-Fusion/default.aspx) and is building custom XPUs plus NVLink-compatible scale-up networking. MediaTek’s $3.5B deal gives hyperscalers a production path for custom XPU design with advanced packaging included. Intel Xeon, Fujitsu’s MONAKA Arm CPU, Qualcomm, and SiFive’s RISC-V chips all connect via NVLink-C2C. This is a functioning ecosystem, not a roadmap.

## The Memory Innovation That Changes XPU Economics

Bundled with NVLink Fusion is NVHBM — [NVIDIA’s custom HBM base-die technology](https://developer.nvidia.com/blog/nvidia-nvlink-fusion-brings-nvhbm-to-next-generation-ai-infrastructure) developed with leading memory vendors. The numbers compound in a useful way.

| Metric | NVHBM | Standard HBM4e | 
|---|---|---|
| Memory bandwidth per stack | +30% | Baseline | 
| PHY footprint | −67% | Baseline | 
| Power per stack | −15% | Baseline | 
| Additional compute silicon per die | +30% | Baseline | 
| End-to-end XPU performance | +30% | Baseline | 

Moving the memory controller into the HBM base die rather than the XPU die is the key design decision. It frees up roughly 30% more silicon for compute logic — meaningful real estate at advanced process nodes. At data center scale, that 15% power reduction means 15,000 additional XPUs per gigawatt-scale facility.

## The Alternative Is Real

AMD’s Helios platform makes the opposing case: 72 MI455X accelerators, Epyc CPUs, Pensando networking, connected via [UALink](https://www.amd.com/en/developer/ualink.html) — an open standard that doesn’t require NVIDIA licensing. If your organization has reasons to keep NVIDIA out of the stack, UALink is a credible path.

The honest comparison: UALink is open and vendor-neutral. NVLink Fusion has more validated ecosystem partners, better-documented integration paths, and the largest install base of any scale-up interconnect. Matt Kimball of Moor Insights summed up the broader dynamic: “Heterogeneity is the future of AI.” The question is which heterogeneity — NVIDIA-adjacent or genuinely open.

## What Developers Should Take From This

If you deploy workloads on AWS, Google Cloud, or Azure, the chip your model runs on is going to change. It may already be a custom XPU rather than an H100. NVLink Fusion is the reason that change won’t break your existing toolchain — CUDA still works, the networking stack is familiar, and the abstraction layers your team uses are unaffected.

If you’re on an ML infrastructure team, the heterogeneous cluster is the planning assumption now, not a future scenario. Trainium4’s architecture is the template: GPUs for flexible workloads, XPUs for cost-optimized inference, all in the same rack. [NVIDIA just made sure its interconnect runs both.](https://nvidianews.nvidia.com/news/nvidia-and-mediatek-deepen-long-standing-partnership-to-build-ai-edge-to-cloud-computing-platforms)

The $3.5B MediaTek investment is a number. The real news is that NVIDIA has positioned itself to collect infrastructure rent on custom AI chips — including the ones being built specifically to reduce dependency on NVIDIA GPUs. Whether that’s brilliant strategy or concerning lock-in depends on whether you’re a shareholder or an infrastructure engineer trying to reduce vendor dependency.