Marvell’s $3.25 billion buyout of Celestial AI isn’t just another semiconductor acquisition. It’s a wager—big, loud, and physics-driven—on the idea that the future of artificial intelligence won’t run on copper wires at all. To bankroll the bet, Marvell unloaded its automotive Ethernet division to Infineon for $2.5 billion, then pivoted hard into the fast-forming world of silicon photonics, where bits don’t crawl along metal traces but zip through chiplets on beams of light.
This isn’t theoretical lab stuff anymore. AI compute loads have exploded so violently that the electrical interconnects inside datacenters—those boring copper pathways that have carried digital life for decades—are now hitting real physical limits. As bandwidth doubles, the usable length of a copper route effectively gets cut in half, and hyperscale operators are discovering they can’t cool or power traditional interconnects fast enough to link the next wave of super-sized AI accelerators. As recent market analysis notes, data movement itself now represents roughly a quarter of a datacenter’s total power draw, and optics offers one of the few viable paths to ease that energy burden while accelerating performance.
For years, silicon photonics was a niche technology, tucked away in telecom backhaul and government labs. Not anymore. Analysts expect the photonics market to rocket into double-digit billions by the early 2030s as AI superclusters scale into multi-terabit, multi-megawatt beasts. Co-packaged optics (CPO), moving the lasers and light guides directly next to or onto the processor package, are set to become a multibillion-dollar category on their own. Optical networking stocks are quietly making new all-time highs almost on a daily basis, driven by accelerating data center demand even outside the flashiest GPU narratives.
And the fight for CPO leadership is turning into a heavyweight contest. Intel, after years of quietly shipping millions of silicon photonics transceivers, is using that manufacturing base to fast-track its own CPO architectures. Intel’s foundry roadmap puts optical I/O near the center of its long-term datacenter strategy, and the company has been demonstrating early CPO prototypes that merge its packaging prowess with integrated lasers. Meanwhile, TSMC is pushing aggressively into optical packaging services. Its CoWoS and InFO platforms are slowly gaining photonic extensions, positioning TSMC as the manufacturing backbone for any company that wants CPO chiplets without having to build a photonics fab of its own. Between Intel’s deep IP and TSMC’s packaging ecosystem, CPO is moving from a research curiosity to a manufacturable product category.
That’s where Celestial AI comes in: the startup has been developing dense, on-package photonic fabrics designed to knit together accelerators and memory at bandwidths electrical traces simply can’t handle. Marvell is betting that owning this technology outright lets it ride the optical wave instead of being swept under it.
Marvell isn’t alone in sniffing opportunity. Lumentum, a longtime heavyweight in photonic components, became a starring player in 2025 when Nvidia tapped it as a key supplier of indium phosphide lasers for its next-generation Spectrum-X and Quantum-X optical switches, the very connective tissue of Nvidia’s AI networking empire. Nvidia’s recent GTC announcements made it clear that co-packaged silicon photonics networking switches are central to scaling AI factories, addressing both latency and power inefficiencies that plague conventional interconnects.
Meanwhile, POET Technologies has taken a more integrated, almost minimalist approach. Its Optical Interposer platform blends electronics and photonics into a single low-power module. By pairing this architecture with high-speed lasers from partners like Lumentum, POET is chasing the hyperscale transceiver market. Some analysts think the broader photonic integrated circuit market could surge toward $100 billion by 2034, especially as cloud and AI vendors hunt for cheaper, smaller, cooler optics that don’t require painstaking assembly under microscopes. POET is betting that its compact, manufacturable designs can undercut the sprawling photonic modules of old.
At the very bottom of this optical pyramid sits AXT, a company most consumers have never heard of but whose substrates quietly enable the entire photonics boom. AXT makes the compound semiconductor wafers—indium phosphide, gallium arsenide, germanium—that lasers, modulators, and detectors are built on. These aren’t optional materials; you simply can’t make high-speed optical devices on plain silicon. Recent moves, including China restoring AXT’s crucial indium phosphide export permits and the company’s progress toward 8-inch GaAs wafers, have restored momentum to a supply chain that looked shakier a year ago. The photonics revolution can’t happen without the right crystals, and AXT is one of the few companies that can reliably grow them.
Put together, this entire scene feels less like an industry shift and more like a platform reboot. Lumentum and other component powerhouses are pushing the frontier of laser physics. POET and Celestial AI are reorganizing how photonics and silicon live together on a chip. AXT is supplying the raw materials that make any of this possible. Intel and TSMC are turning CPO from a science project into a shipping product category. And now Marvell is stitching itself into the narrative, betting that buying Celestial AI puts it ahead of the wave as datacenters prepare to ditch electrons for photons at every critical juncture.
As silicon photonics and co-packaged optics graduate from experimental prototypes to full-scale datacenter deployments, the battle for AI infrastructure supremacy is expanding across every layer of the stack. From wafer fabrication labs to hyperscale server floors, legacy optics companies, advanced foundries, substrate makers, chiplet integrators, and ambitious startups are all racing toward the same goal: designing systems that can keep pace with AI’s ever-growing appetite for bandwidth and speed.
Marvell is betting boldly on one solution: if electrons can’t keep up, photons will carry the load.
