Marvell Inks $5.5B Celestial AI Deal to Break Memory Wall
Marvell Technology has entered into a definitive agreement to acquire Santa Clara-based Celestial AI. The transaction is valued at $3.25 billion upfront, with approximately $1 billion in cash and $2.25 billion in stock. There is an additional $2.25 billion in potential earnouts based on revenue milestones.
The deal is expected to close in early 2026.
While Celestial AI is not yet a household name, the acquisition would give Marvell control over one of the most critical technologies in hyperscale architecture: the "Photonic Fabric." Celestial specializes in optical interconnects designed to separate memory from compute.
In current AI infrastructure, chips like NVIDIA’s Blackwell are often bottlenecked by how fast they can fetch data from local memory. Celestial AI uses light (photonics) rather than copper electricity to fetch data from remote memory pools at speeds and latencies that mimic being on-chip. If successful, this acquisition allows Marvell to sell the essential optical plumbing required for the next generation of massive AI clusters.
The Context
This acquisition arrives at a critical inflection point for AI infrastructure in late 2025:
- The "Memory Wall" is Here: Frontier LLMs have grown so large that they no longer fit on a single GPU—and for training, often not on a single rack. The bottleneck for training next-gen models isn't how fast the chip calculates (FLOPS), but how fast it can move data (Bandwidth).
- Copper is Dying: Traditional electrical interconnects (copper wires) are hitting physics limits. They get too hot and lose too much signal over distances longer than a few meters. To build the “million-chip clusters” planned by OpenAI and Microsoft, data centers are shifting to light.
- The Marvell vs. Broadcom War: This is a direct shot at Broadcom. While Broadcom dominates the standard switching market, Marvell is positioning itself as the king of custom AI plumbing. By owning Celestial, Marvell can offer a full optical stack—DSP, Switch, and now the Interconnect Fabric itself—to hyperscalers who want to build their own custom AI clusters.
TechArena Take: The Optical Moat Strategy
If the Palo Alto/CyberArk deals of 2025 were about building a "Moat of Trust" (software), this deal is Marvell attempting to dig a "Moat of Light" (hardware).
We view this acquisition as a defensive masterstroke and a high-risk integration challenge:
- For years, architects have dreamed of “disaggregated computing”—where you have racks of just GPUs connected to racks of just memory, all sharing data instantly. It seemed like a pipe dream because the latency was too high. Marvell spending $3.25B upfront on Celestial is the industry’s strongest signal yet that optical disaggregation works. They aren't buying a science experiment; they are buying the plumbing for GPT-6.
- NVIDIA’s "moat" isn't just the GPU; it’s NVLink—the proprietary cable that lets GPUs talk to each other. If you want top performance, you must buy the whole NVIDIA rack. Celestial AI offers a neutral alternative that challenges NVLink. It allows AMD, Intel, or custom silicon like Amazon’s Trainium chips to talk to each other (and to memory) just as fast as NVIDIA’s ecosystem. Marvell is effectively selling “NVIDIA-class performance for everyone else,” constructing a direct assault on NVIDIA’s dominance.
- Committing to a deal valued up to $5.5 billion for a startup with minimal revenue sounds like 2021 all over again. But look at the CapEx budgets. Microsoft and Google are spending $50B+ per quarter on infrastructure. If Celestial’s tech saves even 10% on power or increases efficient utilization of those clusters by 15%, the $5B price tag is a rounding error. Marvell knows that the only thing more expensive than buying Celestial is letting Broadcom buy it. The company’s press release states that it expects “meaningful revenue contributions” to begin in the second half of fiscal 2028, with a target of $500M annualized run rate by Q4 fiscal 2028.
This isn't just a chip deal; it’s an infrastructure bet. Marvell is betting that the future of AI isn't just bigger chips, but better wiring. If they are right, they just bought the nervous system of the 2026 data center.
Marvell Technology has entered into a definitive agreement to acquire Santa Clara-based Celestial AI. The transaction is valued at $3.25 billion upfront, with approximately $1 billion in cash and $2.25 billion in stock. There is an additional $2.25 billion in potential earnouts based on revenue milestones.
The deal is expected to close in early 2026.
While Celestial AI is not yet a household name, the acquisition would give Marvell control over one of the most critical technologies in hyperscale architecture: the "Photonic Fabric." Celestial specializes in optical interconnects designed to separate memory from compute.
In current AI infrastructure, chips like NVIDIA’s Blackwell are often bottlenecked by how fast they can fetch data from local memory. Celestial AI uses light (photonics) rather than copper electricity to fetch data from remote memory pools at speeds and latencies that mimic being on-chip. If successful, this acquisition allows Marvell to sell the essential optical plumbing required for the next generation of massive AI clusters.
The Context
This acquisition arrives at a critical inflection point for AI infrastructure in late 2025:
- The "Memory Wall" is Here: Frontier LLMs have grown so large that they no longer fit on a single GPU—and for training, often not on a single rack. The bottleneck for training next-gen models isn't how fast the chip calculates (FLOPS), but how fast it can move data (Bandwidth).
- Copper is Dying: Traditional electrical interconnects (copper wires) are hitting physics limits. They get too hot and lose too much signal over distances longer than a few meters. To build the “million-chip clusters” planned by OpenAI and Microsoft, data centers are shifting to light.
- The Marvell vs. Broadcom War: This is a direct shot at Broadcom. While Broadcom dominates the standard switching market, Marvell is positioning itself as the king of custom AI plumbing. By owning Celestial, Marvell can offer a full optical stack—DSP, Switch, and now the Interconnect Fabric itself—to hyperscalers who want to build their own custom AI clusters.
TechArena Take: The Optical Moat Strategy
If the Palo Alto/CyberArk deals of 2025 were about building a "Moat of Trust" (software), this deal is Marvell attempting to dig a "Moat of Light" (hardware).
We view this acquisition as a defensive masterstroke and a high-risk integration challenge:
- For years, architects have dreamed of “disaggregated computing”—where you have racks of just GPUs connected to racks of just memory, all sharing data instantly. It seemed like a pipe dream because the latency was too high. Marvell spending $3.25B upfront on Celestial is the industry’s strongest signal yet that optical disaggregation works. They aren't buying a science experiment; they are buying the plumbing for GPT-6.
- NVIDIA’s "moat" isn't just the GPU; it’s NVLink—the proprietary cable that lets GPUs talk to each other. If you want top performance, you must buy the whole NVIDIA rack. Celestial AI offers a neutral alternative that challenges NVLink. It allows AMD, Intel, or custom silicon like Amazon’s Trainium chips to talk to each other (and to memory) just as fast as NVIDIA’s ecosystem. Marvell is effectively selling “NVIDIA-class performance for everyone else,” constructing a direct assault on NVIDIA’s dominance.
- Committing to a deal valued up to $5.5 billion for a startup with minimal revenue sounds like 2021 all over again. But look at the CapEx budgets. Microsoft and Google are spending $50B+ per quarter on infrastructure. If Celestial’s tech saves even 10% on power or increases efficient utilization of those clusters by 15%, the $5B price tag is a rounding error. Marvell knows that the only thing more expensive than buying Celestial is letting Broadcom buy it. The company’s press release states that it expects “meaningful revenue contributions” to begin in the second half of fiscal 2028, with a target of $500M annualized run rate by Q4 fiscal 2028.
This isn't just a chip deal; it’s an infrastructure bet. Marvell is betting that the future of AI isn't just bigger chips, but better wiring. If they are right, they just bought the nervous system of the 2026 data center.
.png)
