Marvell Technology is paying 3.25 billion for Celestial AI and telling Wall Street its AI engine is just getting started. Shares jumped as much as 10% premarket after the chipmaker paired the deal with a bullish outlook for next fiscal year, a show of confidence that its custom silicon franchise with hyperscalers can power through a crowded field. The bet is squarely on silicon photonics, the next battleground for data center performance and a front where Marvell is racing Broadcom and Nvidia.
The move landed on a volatile tape. Marvell’s stock initially slipped in late trading Tuesday, then reversed higher before the open as investors digested a guidance framework that implies steadier growth into 2026. Third-quarter revenue rose 36.8% to 2.07 billion, in line with estimates, and the company guided fourth-quarter sales to about 2.2 billion, plus or minus 5%, a notch above consensus. Management expects roughly 10 billion in total revenue in its next fiscal year, including a 25% jump in data center revenue and about 20% growth from its custom chip business. Crucially, Marvell said it does not expect big quarter-to-quarter swings in custom silicon next year, a nod to smoother ramps with cloud customers.
For a stock that was down more than 15% year to date heading into results, the setup is cleaner: a credible growth trajectory anchored by hyperscaler programs and a transformative asset in Celestial AI. The combined message—bigger pipeline, fewer air pockets—helped flip sentiment from bubble fears back toward execution and share gain potential.
Celestial AI brings photonic interconnects that use light, not electricity, to shuttle data between AI accelerators and memory. That matters because training frontier models is increasingly constrained by bandwidth, latency, and power at the cluster level, not just the speed of any single chip. Optical I/O can move more bits per joule and per dollar than copper at scale, and it points to how hyperscalers will wire racks and pods when GPUs are memory-bound. Marvell says Celestial’s technology will feed a new 10 billion dollar market in photonics-related infrastructure, folding into next-generation products that sit alongside its networking and custom silicon lines.
CEO Matt Murphy put it bluntly: “We’re going to have a silicon photonics powerhouse at Marvell when this is all done.” The timing is long-dated but specific. Large cloud deployments of these photonic fabrics begin around 2027 or 2028, with Marvell modeling meaningful revenue contributions in the second half of fiscal 2028, a 500 million annualized run rate by the fourth quarter of fiscal 2028, and a 1 billion run rate by the fourth quarter of fiscal 2029. That trajectory gives investors an anchor for the out-years while the core custom and networking businesses carry the near term.
The competitive framing is clear. Broadcom already dominates custom accelerators and high-speed optics inside data centers, and Nvidia—now the world’s most valuable company—has built an ecosystem around NVLink, NVSwitch, and proprietary networking that keeps customers inside its stack. Hyperscalers want alternatives and control, which is why custom AI silicon is a priority for Amazon and Microsoft. Marvell is one of the few with the IP blocks, packaging know-how, and customer access to stitch compute, memory, and interconnect together at hyperscale economics.
JP Morgan’s Harlan Sur has previously flagged Marvell’s role in helping Amazon and Microsoft build in-house AI chips. The Celestial AI acquisition tightens that story by attacking the system bottleneck that follows compute: moving data off chip and between nodes efficiently. If Marvell can deliver photonic fabrics that slot into hyperscaler roadmaps without rewriting the entire software stack, it stands to pry open sockets that today default to Nvidia or Broadcom.
Under the agreement, Celestial AI holders receive 1 billion in cash and about 27.2 million Marvell shares valued at 2.25 billion. There are also performance-based earnouts that could lift the total consideration to as much as 5.5 billion, tying more of the price to revenue milestones as adoption ramps. Closing is targeted for the first quarter of calendar 2026, subject to customary approvals.
Yes, issuing 27.2 million shares carries dilution. Marvell is offsetting that with a long runway of incremental revenue tied to hyperscaler spend—and with a structure that pushes the highest deal values into future periods only if the technology lands. Investors will focus on how quickly Celestial’s photonics IP can be productized inside Marvell’s portfolio and what that means for gross margins. Optical modules and co-packaged optics can be margin-accretive if the company controls enough of the stack, but early generations often carry higher costs and yield learning curves.
Alongside the purchase, Marvell issued a warrant to Amazon. The agreement allows Amazon to buy up to 90 million dollars of Marvell stock—roughly one million shares—at an exercise price near 87 dollars, based on Amazon’s purchases of Marvell’s photonic fabric products through the end of 2030. In other words, the more AWS deploys Marvell’s optics, the more equity it earns the right to buy.
That mechanic matters. It signals a real commercial path with a marquee cloud buyer and aligns incentives around volume. Combined with Marvell’s existing programs in custom silicon for Amazon and Microsoft, the warrant reads less like a headline sweetener and more like a breadcrumb trail to production orders. If those orders materialize, the warrant’s dilution is paired with revenue and operating leverage that should overwhelm the share count impact.
Silicon photonics is not a weekend project. Integrating Celestial AI’s designs, qualifying modules, and hitting hyperscaler reliability targets will take years, not quarters. Manufacturing yields, supply chain for laser components, and co-packaging with advanced nodes are real execution risks. There is also the calendar: management’s timeline points to 2027–2028 for large-scale installations, leaving a long gap for competitors to respond or for standards like CXL to shift system architectures.
Regulatory risk around the deal looks manageable given market structure, but integration risk is real. Watch for early design wins, proof points on power per bit and latency at cluster scale, and signs that cloud customers are willing to standardize on Marvell fabric around 2027. Any slip in those markers would soften the out-year revenue bridge that’s supporting today’s optimism.
Near term, guidance puts a floor under expectations. The call for 10 billion in fiscal revenue with 25% data center growth and steadier custom chip shipments telegraphs better visibility with cloud customers. It also sets a higher bar. If demand normalizes or competitors undercut, the no-whipsaw promise will be stress-tested. Gross margin trajectory will be the tell on mix and pricing power across custom silicon, networking, and nascent optics.
The setup for the stock is deceptively simple. If Marvell turns Celestial AI into a shipping photonics platform on the timeline outlined, and if Amazon’s warrant signals broader hyperscaler adoption, the market will reward the company with a premium custom-silicon multiple. If photonics slips or stays niche while Nvidia tightens its grip with proprietary interconnects, today’s bounce will fade. For now, the combination of a cleaner 2026 guide and a credible plan to own the next bottleneck is the first real narrative reset Marvell has had all year.
