The next constraint on artificial intelligence may not be how quickly a chip can calculate. It may be how efficiently thousands of chips can communicate.
That is the strategic idea behind one of the largest proposed deals in the AI infrastructure market. On November 2, 2026, Marvell Technology announced a definitive agreement to acquire optical technology company Lumentum in an all-stock transaction with an implied enterprise value of $115 billion. The deal brings together Marvell’s custom silicon, switching and interconnect technologies with Lumentum’s lasers, optical transceivers, optical circuit switching and compound-semiconductor manufacturing capabilities.
The headline value is striking. The industrial logic is more important. As AI systems expand from individual servers into enormous clusters, moving data between processors, memory, switches and racks is becoming a central engineering challenge. The next phase of AI infrastructure will be shaped not only by compute, but by the networks that keep compute working.
AI’s next bottleneck is the movement of data
Modern AI workloads depend on large numbers of accelerators working together. During training and inference, those processors constantly exchange parameters, activations and other data. If communication slows, expensive compute capacity can sit underused while workloads wait for information to arrive.
This is why data-centre networking has become more than a supporting layer. It is part of the performance architecture. Bandwidth, latency, distance, reliability and power consumption all influence how effectively an AI cluster operates. Microsoft Research has examined the networking wall in AI infrastructure, describing how memory and network bottlenecks can constrain GPU utilisation and overall system efficiency.
For years, the industry could increase compute capacity by adding more processors. But adding processors also increases the volume of communication required between them. The result is a system-level problem: more compute creates more pressure on the connections linking that compute.
Why copper cannot solve every connectivity problem
Copper remains useful. It is cost-effective, reliable and power-efficient over short distances, which is why it continues to play an important role inside racks and in other short-reach connections. Optical links are not a universal replacement, and the right choice depends on distance, speed, power, cost and system design.
But as data rates rise and connections need to travel farther, electrical signals over copper become harder to manage. Optical technologies transmit information as light through fibre, making them attractive for higher-bandwidth connections across longer distances and increasingly dense systems.
There is a second shift underway: optics is moving closer to the compute and switching silicon. Co-packaged optics, for example, places optical engines alongside networking chips rather than relying only on separate pluggable transceivers. NVIDIA’s technical overview of optical interconnects describes how these systems can increase bandwidth density and address power and signal-loss challenges in high-performance AI networks. The technology is still evolving, and copper and optical links will coexist across different parts of the data centre.
The architecture is becoming more layered. Short-reach electrical connections remain valuable. Optical transceivers connect equipment across racks and rooms. Silicon photonics and co-packaged optics seek to bring optical communication closer to the chips that process and route data.
What Marvell is buying with Lumentum
The acquisition is designed to combine capabilities across that signal chain. Marvell brings custom compute silicon, switching, electrical and optical interconnect technologies. Lumentum adds high-performance lasers, optical transceivers, optical circuit switching and specialised compound-semiconductor manufacturing, including indium phosphide capabilities.
In practical terms, the strategic ambition is to design more of the connectivity system together: the electrical interface, the optical engine, the laser, the switching layer and the manufacturing process. The companies argue that closer integration can help customers deploy pre-validated solutions faster, while improving bandwidth, performance and power efficiency.
The agreement is an all-stock transaction, as detailed in Reuters’ deal coverage. Lumentum shareholders are set to receive 4.5 shares of Marvell for each Lumentum share. The companies said Marvell shareholders would own approximately 67% of the combined company and Lumentum shareholders approximately 33% after closing. The SEC filing and merger agreement set out the transaction terms. The expected completion window is the first half of 2028, subject to shareholder and regulatory approvals and other customary conditions.
Marvell and Lumentum have also presented a combined addressable-market estimate of $550 billion and said the transaction offers a line of sight to more than $1 billion in annual run-rate cost synergies. Those are management projections, not guaranteed outcomes. The scale of the deal makes execution, integration, customer retention and the economics of future AI infrastructure spending important tests.
Connectivity is becoming an AI infrastructure moat
The deal points to a broader change in where competitive advantage may be built. The AI infrastructure stack is no longer simply a contest between accelerator manufacturers. It includes high-bandwidth memory, advanced packaging, networking silicon, optical components, power delivery, cooling, data-centre capacity and the software that coordinates workloads.
These layers are interdependent. A faster accelerator does not deliver its full value if memory cannot feed it or the network cannot move data quickly enough. Likewise, a high-performance optical link is only useful when it fits the system’s topology, reliability requirements, thermal envelope and cost structure.
This is also why the market is attracting investment beyond a single acquisition. NVIDIA and Coherent’s strategic partnership focuses on advanced optics, manufacturing capacity and research and development for next-generation data-centre architectures. The Optical Scale-up Consortium, involving companies including AMD, Broadcom, Meta, Microsoft, NVIDIA and OpenAI, is another sign that the ecosystem is preparing for new ways to connect AI systems.
The strategic question is shifting from “Who can supply the most compute?” to “Who can deliver the most useful compute per watt, per dollar and per unit of time?” Connectivity is central to that equation.
What this means for India
India’s opportunity is real, but it needs to be described accurately. The country is expanding its data-centre footprint and building capabilities in AI infrastructure, software, systems integration and semiconductor design. That does not mean India is already a major global source of advanced optical components or compound-semiconductor manufacturing. Those are specialised capabilities that require deep process expertise, capital, supplier ecosystems and long qualification cycles.
The near-term opportunity may lie across several adjacent layers: data-centre design and operations; network architecture and deployment; power and cooling optimisation; optical-fibre infrastructure; testing and reliability engineering; and software that improves utilisation across expensive compute clusters. Indian engineering teams can also contribute to system design, silicon development, firmware, network management and the tools used to operate AI factories.
There is a useful domestic reference point in the build-out of GPU infrastructure hosted in Indian data centres. FutureIsNow has examined how STT GDC and Neysa’s AI compute infrastructure deployment in Noida reflects the emergence of local capacity for AI workloads. The next question is how that capacity is connected, powered and operated efficiently as deployments scale.
India’s broader AI opportunity also depends on whether enterprises can turn infrastructure into measurable productivity. Our analysis of the AI productivity paradox in Indian IT services explores why infrastructure investment alone does not automatically translate into economic gains. Better connectivity is an enabler; the commercial value still depends on how well organisations use the systems they build.
For Indian policymakers and investors, the lesson is to look beyond headline announcements about data-centre capacity. The deeper value chain includes optical networking, advanced packaging, semiconductor materials, testing, power electronics, cooling and specialised manufacturing. Building meaningful participation in these areas will require sustained investment and collaboration between industry, research institutions and global suppliers.
The risks behind the optics opportunity
There are reasons to be cautious. The $115 billion enterprise value places a high bar on the strategic and financial benefits that Marvell expects to capture. The companies must integrate businesses with different product cycles, customers, manufacturing requirements and operating models. They must also prove that demand for advanced optical connectivity will grow quickly enough to justify the investment.
Technology transitions rarely happen uniformly. Copper will remain competitive in short-reach applications, while optical architectures must continue to improve on cost, power, reliability and manufacturability. New approaches such as co-packaged optics also introduce design and serviceability questions. The market will not move in a straight line from electrical to optical connections.
There is a wider capital-allocation risk. AI infrastructure spending is rising, but customers will eventually demand evidence that the capacity being built produces sustainable returns. If deployment growth slows, or if customers change architectures, suppliers across the stack could face pressure. The acquisition is therefore a bet on the durability and direction of AI infrastructure demand—not proof that every part of the optics market will win.
The next AI advantage may be built between chips
Marvell’s proposed acquisition of Lumentum is significant because it recognises that the performance of an AI system is determined by more than the processors inside it. The connections between those processors increasingly shape how far a cluster can scale, how much energy it consumes and how efficiently its compute is used.
The deal does not mean copper is finished, nor does it guarantee that optical technologies will dominate every connection. It does signal that the companies building the next generation of AI infrastructure see connectivity as a strategic layer worth owning more deeply.
For the AI industry, the next competitive advantage may not come only from making chips faster. It may come from making the entire system communicate better.
FutureIsNow perspective: The next AI bottleneck may not be computing. It may be connectivity.
