Marvell Bets $115 Billion on Light: Inside the Photonic Interconnect Race

That is the newest and largest move in a sector now consolidating around one problem. As AI clusters outgrow rack-scale electrical links, the wires between chips and switches become the bottleneck. Marvell says it wants to own the fix.

The price is notable. EE Times reported on 3 November that the deal is valued at $115 billion in stock, and that both companies are Nvidia-backed. Marvell supplies network processors, DSPs, switches and custom accelerators that sit inside data center optical transceivers. Lumentum brings the lasers that turn electrical signals into light, plus optical circuit switches, indium phosphide manufacturing and transceiver know-how. The stated aim is to co-design the electrical and optical signal chain rather than buying the two halves from separate suppliers. That is a different proposition from stitching together off-the-shelf parts. It is also why the two companies frame this as a single design effort rather than a portfolio purchase.

“AI infrastructure is undergoing a fundamental architectural shift toward optical connectivity, creating an opportunity that comes along once in a generation,” Marvell chairman and CEO Matt Murphy said, according to EE Times. Lumentum CEO Michael Hurlston put the addressable market at $550 billion, the same report says.

Not everyone is convinced by the arithmetic. Axel Rudolph, chief technical analyst at IG Group, told Reuters the deal is an enormous bet. He said investors are right to ask whether Marvell is paying too much and taking on too much at once, EE Times reported. The same article notes the hand-off between silicon-based network interface cards and laser-driven transceivers as a major connectivity bottleneck in cloud computing. Tighter co-design there could cut both latency and power draw.

Why the copper wall matters

The physics is not new, but the scale is. Electrical links degrade over distance and burn power pushing bits through copper. Optical links carry more data further with less loss. As AI training and inference clusters grow, the interconnect between accelerators, switches and memory has become as important as the compute itself. That is why a chip company is buying a photonics company. It is also why analysts have been calling for this kind of combination since the early 2020s, according to EE Times.

The research pipeline shows the same pressure from the other direction. Semiconductor Engineering reported on 2 November that IBM Research published work demonstrating four long-distance couplers. They operate between two 120-qubit IBM Nighthawk R1 superconducting processors. The paper, “Meter-scale quantum interconnects between large quantum processors,” is on arXiv as 2610.29986. It argues that modular quantum architectures need coherent links over meter-scale distances to stitch processor modules together. It is a quantum result, not an AI networking product, but it makes the same point: once systems get big enough, the links decide what is possible.

Packaging faces its own version of the problem. Also on 2 November, Semiconductor Engineering flagged an open-access paper by researchers at Materials Center Leoben, PVA TePla and Advanced Semiconductor Engineering. The paper covers using AI image enhancement to catch cracks in the extreme low-k layers of flip-chips. These are the so-called white bumps that traditional failure analysis struggles to detect in advanced packaging. Better inspection does not make interconnects faster. It does make the advanced packages that host optical engines more manufacturable.

The money follows the light

Capital is moving at several layers of the stack. In Germany, Quint Computing said on 3 November that it raised €9.1 million in an initial round led by Vsquared. Quantonation, Nucleus Capital, Faber, Vanagon Ventures and ZOHO.VC participated, according to Tech.eu and HPCwire. The round comes with a separate €2.2 million Pathfinder grant from the European Innovation Council. The Erlangen-based startup was founded in October 2025 by Michael Hartmann, Christopher Eichler and Jona Stinner. It is developing a superconducting chip architecture that it says removes a large share of errors at the circuit level. The goal is to cut complexity, hardware overhead and energy use by a factor of 100 once fully implemented.

“The race to build commercially viable quantum computers is gaining momentum,” Eichler, co-founder and CEO, said in the HPCwire report. “Quint has a clear roadmap to accelerate the development of the required technology.” The company emerged from Friedrich-Alexander-Universität Erlangen-Nürnberg and sits inside the Munich Quantum Valley ecosystem.

On the classical side, the demand signal is already in earnings. ZTE shares rose more than 7 per cent in Hong Kong on 27 October. The Chinese equipment maker reported third-quarter net profit of 1.31 billion yuan, up 395 per cent year on year, on revenue of 38.81 billion yuan, up 34 per cent, according to the South China Morning Post, citing the exchange filing. For the first nine months, computing-related revenue grew more than 70 per cent and accounted for 38 per cent of total sales. The stock closed at HK$19.94, a 4.4 per cent gain.

What all of this has in common is that the constraint has moved. For a decade the question was how fast a chip could compute. Now the question is how fast data can get in and out of it, and how much power that trip costs. Optical links are the answer the industry has settled on. That is why a $115 billion all-stock deal for a laser maker reads as strategic rather than eccentric.

Optics is not the only bottleneck being renamed

The political weather around the infrastructure is shifting too. The Verge reported on 3 November that OpenAI has begun using the term “compute campus” instead of data center. The wording debuted in a roughly 900-word announcement for its Grid Protectors Program, which starts with a site in Effingham County, Georgia. The phrase appeared more often than “data center,” which showed up twice. Axios had reported on Monday that OpenAI was planning the rebrand as part of a policy and PR push after the midterms, according to The Verge.

Tom's Hardware reported on 2 November that OpenAI's chief global affairs officer, Chris Lehane, described the midterms as “a bit of a movie trailer for what you're going to see in 2028.” Lehane said the company expects AI to be a major presidential issue. Critics quoted by The Verge were unimpressed. “OpenAI is clearly responding to the unpopularity of data centers,” said Holly Elmore, founder and executive director of PauseAI US. “Rebranding them 'compute campuses' may fool listeners once or twice but it's not the name 'data center' that is unpopular.”

The point for the optics trade is that the buildout is now a political object, not just an engineering one. Every extra campus means more power, more cooling and more optical links between buildings. The interconnect vendors get the orders, the utilities get the bill, and the locals get the noise.

What to watch next

Three things will decide whether the optical bet pays off. The first is whether co-designed electrical and optical parts actually ship in volume. The integration work could slip the way large semiconductor mergers often do. The second is whether the demand holds. ZTE's numbers and the general AI capex cycle suggest it will for now, but a single quarter is not a trend. The third is standards and supply. Lumentum's indium phosphide manufacturing is a scarce asset, and scarce assets invite both premiums and regulators.

There is also the quieter research layer. The arXiv paper by Nastaran Okati, Ricardo Olmedo and Moritz Hardt, submitted on 29 October, formalises “computational arbitrage” in AI model marketplaces. It proves the allocation problem is generally NP-hard, then shows it can be solved efficiently under a diminishing-returns condition. It is about pricing inference, not photons. Still, it is a reminder that the economics of AI infrastructure are still being worked out in public, paper by paper.

And security has not gone away. The Hacker News reported on 30 October that TeamViewer patched a flaw tracked as CVE-2026-111585, scored 8.8 out of 10. A malicious host machine could crash or, on Windows, potentially run code on the viewer's computer. The fix is in version 15.83. Researcher Omri Dar of Remedio published a technical account on 29 October. TeamViewer said it was unaware of any attacks when it published the fix.

None of that changes the direction of travel. The industry has decided that the next generation of AI data will mostly run on light, as EE Times put it, and the cheque has been written. The open question is who ends up holding the lasers.