Imagine trying to run the world’s biggest AI models with extension cords and duct tape. Sounds ridiculous, right? Yet that’s pretty much what the industry has been doing with old-school copper wiring inside supercomputers. Last night Marvell basically said “enough” and wrote a check that could reach $5.5 billion to fix it.
Marvell Makes Its Biggest AI Bet Yet
I’ve been watching Marvell for years, and honestly, 2025 has been rough for them. Down 18% while Broadcom doubled and Nvidia became worth more than most countries. The market was punishing them for being late to the custom AI chip party.
Then Tuesday happened.
After the bell, Marvell didn’t just beat earnings (76 cents adjusted EPS on $2.08B revenue versus 73 cents expected). They dropped the kind of acquisition that makes you sit up straight: they’re buying Celestial AI, the red-hot optical interconnect startup, for a base $3.25 billion in cash and stock — with another $2.25 billion in earnouts that kick in if Celestial hits aggressive revenue targets.
Shares immediately jumped 6% in extended trading. For a $70-billion company, that’s a $4-billion swing on one announcement. The market clearly thinks this is different.
Why Optical Interconnects Suddenly Matter So Much
Let me try to explain this without getting too nerdy — though it’s hard, because this stuff is legitimately fascinating.
Today’s monster AI models (think GPT-5 level and beyond) don’t run on one chip. They run on tens of thousands of GPUs or custom accelerators spread across hundreds of server racks. The bottleneck isn’t compute anymore — Nvidia solved that. The bottleneck is moving data between all those chips fast enough.
Copper wires hit its limit. At the distances and speeds required inside these mega-clusters, electrical signals degrade, consume insane power, and generate heat that could cook dinner.
Light doesn’t have those problems.
Enter photonic interconnects — basically fiber optics on steroids, integrated directly onto the chip package. Celestial AI calls their version a “photonic fabric.” Think of it as replacing the crowded highway system of copper traces with unlimited-lane fiber-optic superhighways that also use 10x less power and run cooler.
“This builds on our technology leadership, broadens our addressable market in scale-up connectivity, and accelerates our roadmap to deliver the industry’s most complete connectivity platform for AI and cloud customers.”
Matt Murphy, Marvell CEO
The Numbers Behind the Deal
Here’s what caught my attention: the earnout structure is brutal in the best way.
- Base price: $3.25 billion
- Maximum price: $5.5 billion
- Trigger for max payout: Celestial must generate $2 billion cumulative revenue by end of fiscal 2029
That’s roughly $500 million a year starting essentially from zero. In startup land that’s an insane growth trajectory — but in AI infrastructure right now? Not impossible. Some of these hyperscalers are spending $50–100 billion annually on AI capex. A few big design wins and Celestial could actually hit those numbers.
Reportedly, Celestial was valued at $2.5 billion in March. Marvell is paying a premium, but not a crazy one given the strategic fit.
How This Changes Marvell’s Position Overnight
Marvell was already strong in data-center networking switches and custom ASICs (they do a lot of work with Amazon, for example). But they were missing the “scale-up” piece — the direct chip-to-chip optical links at rack and even pod level.
Now they own it.
Suddenly Marvell can walk into Google, Microsoft, Meta, or Amazon and say: “You want to build 100,000-XPU clusters that train in days instead of months? We have every piece — custom silicon, switching, and the optical fabric that ties it all together at light speed.”
That’s a conversation Broadcom and Nvidia can have today. Marvell just bought their ticket to the same table.
The Amazon Endorsement Says Everything
Perhaps the most telling part of the announcement was this line buried in the press release:
“Marvell’s acquisition of Celestial will help further accelerate optical scale-up innovation for next-generation AI deployments.”
Dave Brown, Vice President, Amazon Web Services
When the world’s largest cloud provider publicly blesses your acquisition the same day it’s announced, you’re doing something right.
What Happens Next (My Take)
Three things I’m watching closely:
- Integration speed – Marvell promised Celestial tech in custom XPUs “soon.” If they ship silicon with integrated photonics by late 2026, that’s a massive win.
- Competitive response – Broadcom has its own optical efforts (they bought Luxtera years ago), but this lights a fire. Nvidia already has NVLink optics coming. Expect announcements.
- Valuation re-rating – If Marvell starts winning big custom designs because of this, the stock multiple deserves to expand. They’re trading at roughly 30x forward earnings while Broadcom is north of 40x.
In my experience, the market is slow to price in connectivity advantages, then suddenly wakes up. Marvell’s 2022–2023 run happened when investors realized they were quietly taking Amazon’s custom silicon business. This feels bigger.
Look, I’m not saying buy the stock tomorrow morning — do your own homework. But this move dramatically improves Marvell’s seat at the AI infrastructure table. And in this market, that matters more than ever.
The deal is expected to close early 2026. By then we’ll know if those earnouts are in reach.
Either way, the age of optical AI clusters just got a lot closer. And Marvell just positioned itself front and center.
