Nvidia’s Marvell deal shows the AI race is shifting from chips to control

Nvidia (NVDA) just made a move that tells you where the AI race is heading next.

The AI chip giant expanded its partnership with Marvell Technology, plugging Marvell’s custom silicon, networking, optical interconnects, and photonics directly into Nvidia’s NVLink Fusion platform. On the surface, it looks like another bullish signal. Big Tech is gearing up to pour roughly $630 billion into AI infrastructure and data centers by 2026, and Nvidia wants a bigger piece.

The real signal is about control. For years, the scarce asset in AI was the accelerator. Now, it’s the system layer that ties everything together. If data can’t move fast enough, if power use gets out of hand, or if systems can’t scale cleanly, the whole buildout starts to break.

That’s the bottleneck Nvidia is targeting.

AI data centers aren’t getting simpler. They’re getting more mixed, with different types of silicon working together. And Nvidia and Marvell didn’t frame this deal around chip volume. They focused on semi-custom infrastructure, bandwidth, power efficiency, optical links, and photonics.

When companies start talking more about how chips connect than the chips themselves, they’re pointing straight at the friction.

NVLink Fusion is Nvidia’s way into that problem. It lets customers bring in specialized silicon but keeps those systems tied into Nvidia’s interconnect and software stack. That’s not full openness. It’s controlled interoperability. Customers get flexibility where they want it, but Nvidia still sets the rules that matter, a direction that also fits with some of the biggest strategic signals from Nvidia’s 2026 GTC.

That’s why Marvell stands out here. It sits right at the pressure points that are tightening across AI infrastructure, including custom compute, high-speed connectivity, optical DSP, and silicon photonics. And with Marvell projecting nearly 40% revenue growth toward about $15 billion by fiscal 2028, Nvidia is getting closer to a company positioned exactly where the next constraints are forming.

Put simply, Nvidia isn’t just trying to sell more AI chips. It’s trying to own the system that makes those chips work together. That’s a much bigger prize for the stock.

Control is shifting from the chip to the compatibility layer

Value is moving up a layer.

In the first phase of the AI boom, the winner was the company that could ship the most scarce accelerators. In the next phase, the edge may belong to the company that makes mixed systems run like one machine. Once buyers are combining merchant GPUs, semi-custom accelerators, CPUs, switches, and optical links, performance stops being a single-chip question. It becomes a system question: how efficiently data moves, how much power the rack pulls, and whether the whole thing can be deployed at scale without bespoke complexity. Nvidia and Marvell put those constraints at the center of the announcement for a reason.

That shift protects Nvidia even in the scenario investors worry about most: large customers designing around it. If a hyperscaler insists on custom silicon, Nvidia can still hold a durable position if that silicon has to live inside an Nvidia-defined fabric and software stack to work at scale. In that world, Nvidia does not need to own every accelerator socket to keep control of the architecture. It needs to own the standard that makes heterogeneous compute deployable. Selling the chip is powerful. Setting the operating terms for mixed-chip systems is better.

That is also why this partnership is more than a read-through on hot demand. Demand was already obvious. The harder question was where the next choke point would show up as spending expanded. Nvidia’s answer seems to be: in the links between chips, not just the chips themselves. Marvell gives Nvidia more reach into the exact components that matter if that view is right. And Nvidia’s recent pattern of investing across the AI stack suggests this is not a one-off endorsement. It looks more like a broader push to shape the infrastructure around its accelerators, not just the accelerators themselves.

The strategic inversion is the part most people will miss. “Open” AI infrastructure is usually framed as a weakening of vendor control. Nvidia is trying to turn it the other way: more silicon variety underneath, tighter architectural control above. That is a smarter place to defend. In a market moving toward semi-custom and mixed compute, the company that sets the compatibility layer may have more staying power than the one that simply sells the most processors in any single generation.

What to watch next

The first test is partner breadth. If NVLink Fusion starts pulling in more semi-custom or non-Nvidia compute partners, Nvidia’s strategy is real. That would show the company is willing to give up some share at the chip level to gain leverage at the architecture level. If the roster stays narrow, this looks more like branding than a new control point.

The second test is buyer language. Watch whether hyperscalers and suppliers start talking more about optical interconnects, power efficiency, rack design, and networking constraints as core deployment decisions rather than supporting details. If those topics keep moving toward the center of capex discussions, that would confirm the bottleneck is shifting upward from silicon alone to system movement and coordination.

The third test is where custom silicon actually lands. If large customers keep building proprietary accelerators but deploy them inside Nvidia-compatible systems, Nvidia will have preserved the most important layer of control. If they push those chips outside Nvidia’s fabric and software boundary, the strategy is weaker than it looks. That is the real contest now. Not who sells one more processor into an overcrowded rack, but who decides how the rack works when every buyer wants a different mix of silicon. If Nvidia wins that layer, this Marvell deal will look less like a supplier investment and more like a bid to write the rules of the next AI data center.

P.S: And if Nvidia extends that same systems-first logic into entirely new environments, the long-term opportunity could reach beyond terrestrial buildouts, especially as it moves toward making AI chips for data centers in space.

Nvidia (NVDA) and Marvell (MRVL) have a Disruption Score of 4 and 5, respectively. Click here to learn how we calculate the Disruption Score.  

Nvidia is also part of the Disruption Aristocrats, our quarterly list of the world’s top disruptive stocks.