Nvidia acquires $2 billion stake in Marvell Technology in strategic AI partnership

Nvidia just paid $2 billion for a stake in Marvell Technology, and the market noticed immediately. Marvell's stock jumped nearly 10% on the news. That kind of reaction tells you something: this isn't just a financial investment. It's Nvidia planting a flag in the custom chip space, where it has until now relied almost entirely on its own silicon.

The deal centers on Nvidia's NVLink Fusion platform, a rack-scale architecture designed to let companies build AI infrastructure using a mix of Nvidia and third-party components. Marvell will contribute custom XPUs and NVLink Fusion-compatible scale-up networking. Nvidia brings the Vera CPU, ConnectX NICs, Bluefield DPUs, NVLink interconnect, and Spectrum-X switches. Together, the two companies are positioning this as a full-stack AI factory build, not just a chip swap.

Why Marvell, and why now

Marvell has been quietly building custom silicon for hyperscalers for years. Google, Amazon, and Microsoft have all worked with Marvell on application-specific chips that sit alongside general-purpose GPUs. That makes Marvell a natural fit for what Nvidia is trying to do with NVLink Fusion. The platform is specifically built for customers who want to use Nvidia's networking and compute fabric but pair it with their own or third-party processors.

Nvidia has been watching the custom silicon trend with obvious interest. Big cloud providers are designing their own AI chips to reduce dependence on Nvidia GPUs. Rather than fight that trend, Nvidia seems to be threading through it. If your custom chip runs inside an Nvidia-connected rack, Nvidia still controls the interconnect, the networking, and a significant portion of the value chain.

What NVLink Fusion actually changes

Before NVLink Fusion, the NVLink interconnect was essentially proprietary to Nvidia's own GPU clusters. If you wanted NVLink's bandwidth, you bought Nvidia's full stack. The Fusion variant opens that up. Third-party processors can now connect into an NVLink fabric, which means Marvell's XPUs can communicate with Nvidia hardware at speeds that weren't possible over standard PCIe.

For AI workloads, interconnect speed matters as much as raw compute. Training large models across hundreds or thousands of chips requires constant, fast data movement between those chips. Bottlenecks in that communication layer can negate gains from faster processors entirely. By letting Marvell's silicon plug into NVLink, Nvidia is making it easier for customers to build heterogeneous clusters without sacrificing the high-bandwidth fabric that makes Nvidia's own clusters fast.

Silicon photonics and the longer play

The two companies also plan to work together on silicon photonics, which uses light instead of electrical signals to move data between chips. This is where the deal gets more interesting long-term. Silicon photonics can dramatically reduce power consumption and latency in data center interconnects. At the scale modern AI infrastructure operates, those savings are worth billions annually in energy costs alone.

Marvell already has a silicon photonics business. Nvidia, for all its dominance in AI compute, has been less prominent in optical interconnects. Combining Marvell's photonics work with Nvidia's NVLink architecture could give both companies a stronger position in next-generation data center design, where the connection between compute nodes is increasingly the limiting factor.

The competitive context

Broadcom is Marvell's closest competitor in the custom ASIC space, and it has been doing well. Broadcom's custom chip work for Google's TPU program and Meta's MTIA accelerator has made it a go-to for hyperscalers that want an alternative to Nvidia's H100 and B200 GPUs. The Nvidia-Marvell tie-up directly complicates that dynamic. Broadcom now has to consider that its main rival in custom silicon is partially owned by, and technically integrated with, the company whose networking gear everyone is already using.

Goldman Sachs reiterated a buy rating on Nvidia the same day the Marvell deal was announced, citing the company's expanding presence in healthcare AI. Nvidia's share price was up 1.4% in premarket trading on the news. Marvell's 10% surge, though, was the more telling reaction. It reflects how much weight the market puts on being inside Nvidia's ecosystem, even partially.

What this means for AI infrastructure buyers

For companies building out AI infrastructure, this deal offers something genuinely useful: a clearer path to hybrid compute setups. Rather than choosing between an all-Nvidia cluster and a fully custom design, companies can now mix Marvell's XPUs into an NVLink Fusion rack and maintain compatibility with Nvidia's software stack. That reduces the integration risk that has historically made custom silicon less attractive than just buying more H100s.

The Marvell stake also signals that Nvidia sees the AI infrastructure market maturing. In the early phase of the AI boom, customers bought whatever compute they could get. Now, with supply more stable and costs under scrutiny, buyers are asking harder questions about price-per-FLOP and total cost of ownership. A rack that blends Nvidia networking with Marvell custom compute could hit a better number on both metrics than a pure Nvidia configuration at some workloads.

Both companies are expected to share more technical details about the NVLink Fusion platform at Nvidia's GTC conference. The silicon photonics collaboration timeline has not been disclosed.