Marvell的硅光互连布局与商业化前景
Marvell的硅光互连布局与商业化前景
Conclusion
Marvell’s photonics push could make its custom AI chips and switches more useful in larger, tightly connected AI systems—but the strategic prize is an integrated optical interconnect platform, not simply owning a laser. Lasers can carry data as light where copper connections face bandwidth, distance, or power constraints. Marvell’s acquired Photonic Fabric is aimed at bringing that optical connectivity closer to accelerators, including between chips and across racks.
The opportunity is credible, but commercial proof is still early: Marvell’s latest filing says Celestial AI’s revenue and earnings since acquisition were not material through the quarter ended August 1, 2026. Investors should look for customer qualification turning into production orders, reliable high-volume manufacturing, and revenue arriving on the company’s stated timetable.
Data as of October 30, 2026. The latest Marvell quarterly filing reviewed covers the period ended August 1, 2026 and was published August 28, 2026.
What lasers could contribute
AI accelerators exchange large volumes of data with one another and with memory. Copper is effective over short distances, but as systems demand more bandwidth over longer links, the electrical connection itself can consume power and constrain how far apart components can be placed. In an optical link, a laser supplies the light; components such as modulators encode data onto it, waveguides route it, and photodetectors convert it back into electrical signals. That moves data—it does not make the AI computation faster by itself.
The potential payoff is architectural: optical links could let systems connect more accelerators with high bandwidth and low latency while reducing interconnect power, and could make it easier to build larger scale-up domains. Marvell’s acquisition announcement describes Celestial AI’s Photonic Fabric as targeting this kind of scale-up connectivity, including chiplet-based links co-packaged with custom XPUs and switches. The company said its first-generation chiplet offers of bandwidth and claimed more than twice the power efficiency of copper. Those are company claims, not yet evidence of results in broad production deployments.
One important qualification to the “laser technology” framing: Marvell is adding photonic interconnect capability through Celestial AI. Its annual filing lists among its existing optical products, but that does not establish that Marvell is manufacturing laser emitters itself. The value to watch is the complete link—optical engine, packaging, switching, control electronics, reliability and customer integration—not the laser in isolation.
How it could strengthen Marvell’s position
-
More content around Marvell’s custom silicon. If optical links are integrated with Marvell-designed XPUs and scale-up switches, the company could sell a broader system-level set of components to customers building their own AI accelerators. That could deepen customer relationships and differentiate Marvell from a supplier focused on a single chip category.
-
A way to address an emerging bottleneck. Celestial’s proposed Photonic Fabric is aimed at moving scale-up connectivity closer to the package and accelerator, rather than only improving conventional network links between servers or racks. The acquisition announcement also described possible later applications such as pooled memory and optical die-to-die connections. Those are opportunities, not yet established revenue streams.
-
A wider connectivity portfolio. Marvell already sells optical and interconnect products, including laser drivers, DSPs and silicon-photonics-related solutions, and is developing scale-up switching. Celestial potentially adds a differentiated photonic-fabric layer. The strategic case is stronger if Marvell can sell these pieces together—and across more than one customer’s architecture—rather than relying on a single product or design win.
Competition: a real market, but not a greenfield one
Broadcom has an established CPO effort and has described progress from earlier generations to newer products. Its announced Tomahawk 6 platform claims of switching capacity, supports scale-up clusters of , and uses replaceable external laser modules. Broadcom’s releases also describe partner activity around manufacturing, fiber connections and system integration. This is a reminder that performance alone will not decide the contest: packaging, manufacturing yield, reliability and ecosystem readiness matter too.
NVIDIA is both competitor and partner. Its published Spectrum-X Ethernet Photonics material describes a co-packaged optical networking platform and claims a 5× power reduction per 1.6 Tb/s port versus pluggable interconnects. Separately, Reuters reported that NVIDIA invested in Marvell as part of a partnership. So the relationship is better viewed as coopetition: Marvell may supply useful custom silicon and optical technology within an ecosystem in which NVIDIA also sells a tightly integrated compute-and-networking platform.
The distinction investors should test is whether Marvell’s approach becomes a compelling, customer-adopted scale-up optical fabric that complements its custom ASICs and switches—not just whether it can demonstrate photonics. Broadcom and NVIDIA have their own CPO roadmaps, platforms and partner ecosystems.
What investors should watch
-
Customer commitments and conversion. Marvell’s acquisition announcement referred to engagement with multiple hyperscalers and partners, but the public description does not identify production commitments. Look for customer-named design wins where disclosed, qualification milestones, volume orders and evidence that customers are integrating the technology into actual accelerator and switch systems—not just evaluating it.
-
Execution against the revenue timetable. Marvell said meaningful Celestial revenue would begin in the second half of fiscal 2028, with a US$500 million annualized run rate by the fourth quarter of fiscal 2028 and a by the fourth quarter of fiscal 2029. These are forward-looking company targets, not achieved results. An annualized run rate is also not the same as revenue already recognized for a full fiscal year.
-
Proof of manufacturability. Track product sampling and qualification, production availability, manufacturing yield, fiber-attach and packaging capacity, thermal performance, and reliability in customer systems. Optical components can look compelling in a lab and still be difficult to manufacture and service at hyperscale.
-
Measured system-level economics. Seek customer or independently corroborated evidence on power per bit, bandwidth, latency, link stability and total system cost versus copper and competing optical solutions. Company comparisons are useful claims to test, not substitutes for deployed results.
-
Revenue quality and concentration. Watch whether Photonic Fabric adds recurring product revenue and improves Marvell’s position across multiple customers, or remains dependent on a small number of programs. Also examine margins and cash generation: a larger technical footprint only strengthens the business if it translates into attractive economics.
-
Integration with Marvell’s own roadmap. The strategic logic depends on photonics working alongside Marvell’s custom accelerators, switching and other interconnect products. Evidence of working end-to-end systems and interoperability will matter more than a standalone chip specification.
Bottom line
The laser’s role is to provide the light that carries data; the investment thesis is that Marvell can package that capability into a reliable, efficient interconnect product that customers adopt alongside its AI silicon and switches. That could expand Marvell’s relevance in AI infrastructure, especially as systems grow beyond the limits of short-reach electrical links. But the latest filing still describes Celestial’s acquired revenue and earnings as through August 1, 2026, and the company’s announced revenue milestones lie ahead. For now, treat photonics as a potentially valuable strategic option—not a proven earnings engine. The clearest evidence of a stronger competitive position will be production deployments, diversified customer adoption, reliable yields and revenue that meets or exceeds the stated ramp.