Marvell Technology just completed its acquisition of Celestial AI for $3.25 billion in cash and stock, with an additional $2.25 billion in potential earnouts. The total deal could reach $5.5 billion - a staggering bet on photonics as the solution to AI’s bandwidth problem.
The Problem Marvell Is Solving
Modern AI training runs into a wall: data can’t move fast enough between chips. As AI models grow larger and training clusters expand to thousands of GPUs, the electrical connections between chips become the bottleneck. Data centers are hitting the limits of what copper interconnects can deliver.
This is the “memory wall” problem. AI chips sit idle waiting for data. Training takes longer. Costs go up. Power consumption rises.
Celestial AI’s answer: replace electrical connections with light.
What Celestial Brings
Celestial AI built what it calls a Photonic Fabric platform. The key product is a chiplet that integrates all the electrical and optical components needed for high-speed chip-to-chip communication into a single compact package.
The numbers are striking:
- 16 terabits per second of bandwidth in a single chiplet
- 10x the capacity of current 1.6T ports used in scale-out networking
- Bandwidth can scale up to 25x compared to existing off-package solutions
- Substantial power reduction compared to electrical alternatives
The technology works by converting electrical signals to light, transmitting data optically, then converting back to electrical. Light doesn’t generate as much heat as electrical signals and can carry more data over longer distances with less power.
The Financial Math
Marvell paid $1 billion in cash and $2.25 billion in stock at closing. The earnouts add another $2.25 billion tied to revenue milestones.
Marvell expects Celestial to start contributing revenue in fiscal 2028 (roughly two years out), hitting a $500 million annual run rate by Q4 of that year and doubling to $1 billion by Q4 fiscal 2029.
At the base deal price of $3.25 billion, Marvell is paying roughly 3.25x the expected revenue at full ramp. If the earnouts trigger (requiring Celestial to hit aggressive targets), the multiple rises but so does the revenue.
Why This Matters for AI
The AI infrastructure buildout has created massive demand for faster interconnects. Training runs now use thousands of GPUs that need to share data constantly. Inference clusters for frontier models face similar challenges at smaller scale.
Companies like Nvidia have their own interconnect solutions (NVLink, for example), but photonic interconnects could leapfrog electrical alternatives in raw performance. If Celestial’s technology works as advertised, it could become essential infrastructure for next-generation AI clusters.
This acquisition positions Marvell as a key supplier for whoever is building those clusters - whether that’s OpenAI with its new $100 billion war chest, Microsoft, Google, or the hyperscalers racing to build AI infrastructure.
The Competitive Picture
Marvell isn’t alone in pursuing optical interconnects. Nvidia acquired Mellanox in 2020 for networking technology. Broadcom, Intel, and others are developing their own solutions. Startups like Lightmatter and Ayar Labs are also chasing photonic interconnects.
The acquisition takes Celestial’s technology off the open market. Other AI infrastructure buyers will need to either work with Marvell or find alternative suppliers. That could give Marvell leverage in a market where everyone is scrambling to build AI compute capacity.
What Could Go Wrong
Photonic technology has been “almost ready” for decades. Integration with existing chip manufacturing remains challenging. Celestial’s products need to work reliably in production environments at scale - a bar that startup technology doesn’t always clear.
The earnout structure suggests Marvell isn’t certain about the timeline either. Paying a significant portion of the deal in performance-based payments hedges the risk.
If photonic interconnects don’t achieve cost parity with electrical alternatives, or if competing solutions prove good enough, the $5.5 billion bet could look expensive.