Marvell Technology is buying Celestial AI for $3.25 billion, a cash-and-stock push into photonic interconnects that jolted the stock 9% higher in premarket trading. The deal, split between $1 billion in cash and about 27.2 million Marvell shares, is slated to close in the first quarter of 2026 pending approvals. The target is clear: move Marvell deeper into the AI data center stack and narrow the gap with Nvidia and Broadcom as the biggest chipmakers race to unclog AI bottlenecks.
Investors like the direction even if the payoff takes time. In an AI market obsessed with near-term design wins, Marvell’s timeline is longer. The company expects revenue from Celestial AI’s photonic technology to reach an annualized $500 million by late fiscal 2028 and scale to about $1 billion by the end of fiscal 2029. That puts the deal well inside the investment window for hyperscaler capex plans, as Amazon, Microsoft, and Google stretch their 400G-to-800G networking upgrades and prep for the next wave of accelerator racks.
The equity component will dilute existing holders, but the cash outlay is manageable relative to Marvell’s balance sheet and cash generation. What changed the stock’s tone was not just the asset but the narrative: this is a doubling down on data center plumbing—exactly where profit pools are accumulating as AI inference and training costs shift from compute alone to end-to-end system throughput.
Celestial AI’s core pitch is light as the new wire. Its photonic technology aims to move data between chips and memory using optical interconnects instead of electrons, improving bandwidth and slashing power. As AI models scale and memory climbs with high-bandwidth memory stacks, the data path between accelerators and memory is turning into the constraint. Faster, lower-power links are now as critical as the GPUs themselves.
Marvell knows this terrain. The company’s $10 billion acquisition of Inphi in 2021 put it at the center of cloud optics with PAM4 DSPs and coherent solutions that sit inside modern data center links. Celestial AI extends that franchise from rack-to-rack links toward chip-to-memory and in-package or near-package optical I/O. If Marvell can integrate this stack—optical DSPs, custom silicon, and now photonic fabrics—it can sell not just components, but reference architectures tuned for AI clusters. That is where margin durability lives, even as the industry’s cost per bit keeps falling.
The move lands squarely in a heated race. Broadcom is pushing co-packaged optics and merchant silicon that anchor many hyperscaler networks. Nvidia is fusing compute with NVLink, NVSwitch, and networking silicon to keep traffic inside its walled garden. Marvell’s counter is a more open, build-to-spec model: custom accelerators, cloud-optimized silicon, and optics that plug into whatever topologies the customer wants. For hyperscalers wary of vendor lock-in, that is attractive—if the performance and power match Nvidia’s vertically integrated systems.
Analysts see it as a necessary swing. Barclays trimmed its Marvell price target to $130 from $150, keeping an Overweight call while noting Marvell hasn’t hit the loftiest expectations tied to the Amazon supply chain. That critique highlights the stakes. Winning sockets in next-gen AI racks is a design-win game measured in quarters and years, not days. Celestial AI gives Marvell a differentiated story to bring to procurement teams that are now modeling entire rack-level architectures, not just line cards or NICs.
The regulatory clock—closing expected in early 2026—signals a long integration runway. There’s execution risk on two fronts. First, translating a cutting-edge photonic platform into production-grade, high-yield components at cloud scale is tough. Optical I/O at or near the package forces a rethink of thermal design, packaging, testing, and supply chain. Second, aligning roadmaps with hyperscaler build cycles is unforgiving. Miss a cycle, and your revenue curve slides a year.
There’s also competition from other optical I/O efforts and in-house engineering at the largest buyers. Several players are pursuing similar concepts to bring light closer to the die to ease bandwidth pressure and power draw. Broadcom and networking OEMs are advancing co-packaged optics; GPU vendors are tightening proprietary interconnects; chipmakers are experimenting with CXL-based memory pooling that could shift traffic patterns again. Marvell has to prove Celestial AI’s technology offers a clear, defensible cost-per-bit and watts-per-bit edge that holds up in the field.
Marvell’s data center business is already the growth engine. The company recently swung to profit on rising data center demand, and it has leaned into custom silicon programs that line up with hyperscaler strategies. Celestial AI plugs into that playbook, boosting differentiation where it’s most scarce: the memory-to-accelerator pathway and the aggregation layers that bog down when models and datasets grow.
The medium-term revenue targets—$500 million then $1 billion—are meaningful in the context of Marvell’s scale. Hitting them would imply real adoption inside top-tier data centers by 2028–2029. It also builds a hedge against cyclicality in other segments by staking more of Marvell’s future to AI infrastructure rather than legacy carrier or enterprise spending.
AI capex is shifting from a GPU-only land grab to end-to-end performance buys. Bottlenecks have moved to memory bandwidth, interconnect latency, and power. Power budgets are forcing operators to hunt for efficiency as they add racks; anything that lowers watts per bit and increases usable bandwidth per rack wins. Photonics isn’t new, but its role is expanding from long-haul and data center interconnect into shorter-reach, chip-adjacent links. If that shift holds, the value accrues to companies that can supply and support optical stacks at volume.
Marvell is betting that value will accrue horizontally, not just to the compute vendor with the best GPU. That’s a different stance than Nvidia’s vertical control. If hyperscalers push for modular, multi-vendor designs to keep leverage and reduce single-supplier risk, companies like Marvell and Broadcom stand to benefit. The prize is a durable slot in the AI rack bill of materials that is far less cyclical than end-device demand.
Investors should track three proof points. First, design-win disclosures with named hyperscalers or networking OEMs that put Celestial AI tech into next-generation racks. Second, productization milestones that move optical I/O from lab demos to qualified, shippable parts with clear power and cost metrics. Third, alignment with HBM suppliers and packaging partners, since memory bandwidth and thermals will dictate how aggressive optical paths can be.
Marvell’s stock will trade on those signals more than on the distant 2028–2029 revenue targets. Today’s pop reflects a market that wants credible AI infrastructure roadmaps. Photonics is the right battle. Now Marvell has to win it at scale.
