NVIDIA Just Invested $2 Billion in Marvell — And Silicon Photonics Is the Reason Why

NVIDIA (NASDAQ: NVDA) doesn't make $2 billion bets quietly. When the company announced this morning that it has made a $2 billion strategic investment in Marvell Technology (NASDAQ: MRVL) — alongside a formal partnership integrating Marvell into the NVIDIA AI factory ecosystem via NVLink Fusion — the market reacted immediately, sending MRVL shares up roughly 11% intraday.

But the headline number, as striking as it is, may not be the most important part of this announcement.

Buried in the press release — and reinforced by Marvell's own CEO — is a phrase that should be getting more attention: silicon photonics.

What the Deal Actually Says

The partnership covers several pillars: custom XPUs, NVLink Fusion-compatible scale-up networking, AI-RAN infrastructure for 5G/6G, and — critically — "advanced optical interconnect solutions and silicon photonics technology."

Marvell Chairman and CEO Matt Murphy put it plainly:

"By connecting Marvell's leadership in high-performance analog, optical DSP, silicon photonics and custom silicon to NVIDIA's expanding AI ecosystem through NVLink Fusion, we are enabling customers to build scalable, efficient AI infrastructure."

That's not boilerplate. When the CEO of a company just handed a $2 billion check leads his quote with optical DSP and silicon photonics, he's telling you exactly where the strategic value is being created.

Samsung Just Entered Silicon Photonics — Here's What That Te

Jensen Huang framed the macro: "The inference inflection has arrived. Token generation demand is surging, and the world is racing to build AI factories."

Those AI factories need interconnect. A lot of it. And copper doesn't scale at the bandwidth densities that modern AI workloads demand.

This Didn't Come Out of Nowhere for Marvell

Today's announcement is the culmination of an aggressive strategic build-out that has been unfolding over the past several months — one that the market is only now beginning to fully appreciate.

In early March, Marvell reported fiscal Q4 and full-year 2026 results that left little room for debate about the company's trajectory. Q4 revenue came in at $2.219 billion — a new record, up 22% year-over-year. Full fiscal year 2026 revenue hit $8.195 billion, also a record, driven overwhelmingly by AI data center demand. Management guided Q1 FY2027 revenue to approximately $2.4 billion, another step-up.

But the financial results were arguably secondary to what Marvell was doing on the M&A front.

The company completed two acquisitions that directly shape today's NVIDIA partnership. The first was XConn Technologies, which added PCIe and CXL switching silicon and bolstered Marvell's UALink scale-up switching roadmap. The second — and more strategically significant — was Celestial AI.

Celestial AI brought what it calls Photonic Fabric technology into Marvell's Data Center Group: an optical interconnect architecture designed to move data between compute nodes at the bandwidth and power efficiency levels that AI clusters demand. Marvell expects Celestial AI to contribute initial revenue in the second half of fiscal 2028, ramping to a $500 million annualized run rate by Q4 FY2028 and $1 billion by Q4 FY2029.

Read that timeline again. Marvell just told the market that it expects photonics to be a billion-dollar annual business within three years — and then NVIDIA wrote them a $2 billion check.

The Optical Interconnect Layer Is No Longer a Roadmap Item

The photonics conversation has shifted meaningfully in the past 90 days. Coming out of OFC 2026 in March, every major player in the AI infrastructure stack — from hyperscalers to foundries to chip designers — confirmed that optical interconnect is not a future consideration. It is a present requirement.

Co-packaged optics, optical DSPs, and silicon photonics platforms are now being designed into rack-scale architecture at the same time as the compute layer, not after it. The NVLink Fusion framework is a concrete expression of that — it's a rack-scale platform designed to let customers compose semi-custom AI infrastructure, and the optical layer is explicitly part of what gets composed.

Poet Technologies Live Demo’s Three Products at OFC 2026

What today's announcement does is formalize the organizational structure around that transition. NVIDIA anchors the compute. Marvell handles the custom silicon, networking, and now — through Celestial AI — the photonics fabric. The message is unambiguous: you cannot build a competitive AI factory without solving the optical interconnect problem.

The Companies Most Investors Are Watching — And One They May Want To Start Looking At

The names drawing the most investor attention in optical interconnect are the usual suspects: Coherent (NASDAQ: COHR), Lumentum (NASDAQ: LITE), Applied Optoelectronics (NASDAQ: AAOI). These are legitimate businesses doing real volume in the space.

What's interesting about Marvell's Celestial AI acquisition, however, is what it reveals about the architecture of the problem. Marvell didn't buy a transceiver company. It bought a photonics platform — a chip-scale approach to moving data optically across an AI cluster. That distinction matters, because it points toward a different tier of the stack: not the module, but the underlying integration architecture.

POET Technologies (NASDAQ: POET) has been developing exactly that kind of platform — and doing so without an acquisition price tag attached.

POET's Optical Interposer is a chip-scale integration architecture that embeds optical and electronic components on a single substrate, eliminating discrete packaging steps and reducing power consumption and cost per port at scale. Critically, the platform is foundry-agnostic by design — it can be manufactured across multiple fabs, a structural advantage at a moment when supply chain flexibility is a procurement priority, not just a preference.

Where Marvell needed to spend roughly $1 billion to bring photonics integration in-house via Celestial AI, POET has been building that capability organically. The company isn't a household name. It doesn't have a $2 billion check or a record earnings quarter attached to its ticker. But the architecture it's building addresses the same bottleneck that just caused the world's most valuable semiconductor company to make one of its largest-ever strategic investments.

The market has a history of pricing the picks-and-shovels giants first, and discovering the enabling platform companies later.

What This Means Going Forward

The NVIDIA-Marvell partnership does several things at once:

• Validates silicon photonics as a rack-scale requirement, not just a data center curiosity

• Accelerates the timeline for optical integration by anchoring it to the world's dominant AI compute platform

• Raises the competitive bar for anyone selling optical interconnect solutions — the customer expectation is now NVIDIA-grade integration

• Expands the addressable market by tying optical interconnect directly to AI-RAN, which extends the photonics opportunity into carrier infrastructure

The market is, understandably, focused on MRVL's 11% move today. That's the obvious trade. But the more durable question — the one that tends to generate outsized returns when answered early — is: who builds the optical layer that makes NVLink Fusion actually work at scale?

Marvell answered that question for itself. The broader market hasn't fully answered it yet.