The announcement is short, and that is part of the story. On November 3, Tower Semiconductor and Innolight announced that their continued collaboration on near-packaged optics (NPO) for AI infrastructure has produced its first device in early production: Innolight’s first NPO device, built on Tower’s PH18M silicon photonics process flow in Tower’s California factory. In parallel, Innolight has taped out multiple further NPO products on Tower’s 300mm silicon photonics platform, aimed at future product generations. No volumes, no customers, no lane rates, and no device name are disclosed. What is on the record is a manufacturing fact of a specific kind — a foundry process, a factory, and a production status — in a part of the optical-interconnect story that has so far run mostly on demonstrations and roadmaps.
That distinction matters, because this desk’s co-packaged optics (CPO) arc has been dominated by announcements and demos. The question hanging over every CPO and NPO presentation has been the same: who is actually making these things, at what yield, on whose line? A first device entering early production does not answer the yield question or the volume question. But it moves the question from “can this be built?” to “can this be built repeatedly?” — and in photonics manufacturing, that is the crossing that counts.
What was actually announced
Precision first. Tower Semiconductor (NASDAQ/TASE: TSEM) and Innolight announced on November 3, 2026, from Migdal HaEmek, Israel and Suzhou, China, a continued collaboration on NPO applications for AI infrastructure. The release describes two milestones, and they should not be blurred together. The first is production: Innolight’s first NPO device has entered early production using Tower’s PH18M silicon photonics process flow, running in Tower’s factory in California. The second is development: Innolight has taped out multiple NPO products on Tower’s 300mm silicon photonics platform to support future product generations.
The release frames the collaboration as spanning Tower’s high-volume 200mm and 300mm silicon photonics platforms. It also gives the relationship some scale: Tower’s president, Dr. Marco Racanelli, said Innolight is now a manufacturing customer across four Tower facilities — two in the United States, in Newport Beach, California and San Antonio, Texas, and one each in Israel and Japan. Note what that sentence does and does not say. It describes the breadth of the customer relationship, not the production footprint of this NPO device, which the release places specifically on the PH18M flow in the California factory. Do not read the four-facility line as four factories building NPO devices; the release does not say that.
Innolight’s chief technology officer, Oliver Sun, called NPO “a key technology for next-generation AI and cloud infrastructure” and said the Tower collaboration gives Innolight the capabilities to accelerate NPO development for hyperscale and AI data center customers. That is a supplier’s characterization of its own market, and it should be read as such. The verifiable core of the announcement is narrower and more interesting than the framing: a named process, a named factory state, and a tape-out pipeline behind it.
Where NPO sits, and why the middle step matters
The optical interconnect has a geography problem. In a conventional pluggable transceiver, the electrical signal travels from the switch chip across the board to a module at the front panel — a run of ten centimeters or more over copper, cleaned up by a digital signal processor (DSP) in the module. In co-packaged optics, the optical engine is mounted on the switch chip’s own package substrate, millimeters from the silicon, and the long copper run disappears. NPO sits between the two: the optical engine moves off the front panel and onto the board beside the switch chip, close enough that the electrical path shrinks from centimeters to a few centimeters at most, but the engine remains a separate device from the switch package itself.
That middle position is the whole argument for NPO, and also the whole argument against it. Its advocates say it captures much of CPO’s power and bandwidth-density benefit without CPO’s hardest problems: the advanced packaging, the co-designed thermal management of optics and switch silicon in one package, and the serviceability penalty when a failed optical engine is soldered to a switch package that must then be discarded. Its skeptics say it is a compromise that inherits some of both worlds’ costs without fully solving either. That debate is unresolved, and this announcement does not settle it. What it does is give the NPO side something it previously lacked: a device in early production on a merchant foundry line, rather than a reference design or a lab result.
There is a second debate worth naming plainly. Vendors disagree even on what form NPO takes — some describe socketed, field-replaceable engines, others describe soldered-down assemblies — and Tower’s release does not specify the form factor of Innolight’s device. Until Innolight or a customer discloses packaging details, treat the serviceability question as open, not as a claimed advantage of this particular device.
Why channel counts are the manufacturing story
Tower’s release contains one technical claim that deserves more attention than the marketing around it, because it explains why a foundry announcement is the right kind of milestone for this architecture. NPO architectures, the release says, typically require substantially higher optical channel counts than conventional pluggable modules, often integrating 16 or more channels on a single device. As channel counts rise, the release argues, manufacturing expertise, process control, and yield performance become increasingly critical — and Tower’s pitch is that its silicon photonics platforms supply the process uniformity, high-yield performance, and device library those designs need.
This is the valley-of-death logic this desk has covered before, applied to AI optics. A 16-channel device is not sixteen times as hard to build as a single-channel device; it is harder in a compounding way, because the yield of the whole device is the product of the yields of its channels and functions. A process adequate for a pluggable module with a handful of channels can fail economically at NPO channel counts even if its per-channel performance looks fine. That is why the platform detail matters more than the usual press-release superlatives.
What the release does not do is quantify any of this. There are no yield figures, no channel count for Innolight’s specific device beyond the general “16 or more” characterization of NPO architectures, no bandwidth per lane, no aggregate throughput, and no power figure. Those omissions are not a criticism of the release — early production announcements rarely carry that data — but they set the limit of what can responsibly be claimed here. This article asserts no lane rate and no throughput for Innolight’s device, because none is on the record. Any coverage that attaches a number to this device is inventing it.
The PH18M detail is worth placing in context, carefully. Tower’s PH18 silicon photonics platform is an established merchant process that other photonics companies used for modulator development during 2026, including multi-project wafer runs reported in March. That history does not make Innolight’s device proven — a platform’s track record with other customers’ designs is not data on this device’s yield — but it does mean the device is entering production on a line with existing silicon photonics volume behind it, rather than on a bespoke or pilot-only flow.
The day after the $115 billion deal
The timing is impossible to ignore. This announcement landed on November 3, the day after Marvell and Lumentum announced a definitive agreement for Marvell to acquire Lumentum in an all-stock transaction — and the contrast between the two stories is the real news of the week in optical interconnects.
Get the Marvell terms right, because some fast coverage did not. The deal is a proposed acquisition, not a closed one. Under the definitive agreement announced November 2, Lumentum stockholders will receive a fixed exchange ratio of 4.500 Marvell shares for each Lumentum share, giving Lumentum an implied enterprise value of $115 billion, or approximately $1,132 per share — a premium of about 23.9% to the companies’ unaffected closing prices on October 28, the last trading day before market speculation about a transaction began. Post-closing, current Marvell stockholders would own approximately 67% of the combined company on a fully diluted basis and Lumentum stockholders about 33%. The transaction is expected to close in the first half of calendar 2028, subject to stockholder approvals, regulatory approvals, and other customary closing conditions. Both boards approved the agreement unanimously. None of that is a done deal: shareholder votes and regulatory review stand between announcement and closing, and the fixed exchange ratio means the dollar value Lumentum holders ultimately receive will move with Marvell’s share price until then.
The two announcements answer different questions. Marvell–Lumentum is a consolidation story: it says the optical layer is now valuable enough, and scarce enough, that a $115 billion enterprise value can be attached to the company that makes the lasers. Tower–Innolight is a manufacturing story: it says the intermediate architecture between pluggables and CPO can be built on a foundry floor today. Consolidation answers who owns the supply chain; production answers whether the supply chain can make the next thing. Readers should keep the tenses straight: the deal is announced, the device is in production, and neither fact implies the other.
There is also a narrower supply-chain read. Lumentum’s constraint, as this publication covered in October, is factory capacity for indium phosphide devices, with AI optical components sold out through nearly 2029. Tower’s announcement is about a different material system — silicon photonics, not indium phosphide lasers — and it does not relieve that constraint. What it does is show a second supply chain, the silicon photonics foundry chain, moving a next-generation architecture into production at the same moment the laser chain is consolidating.
What is undisclosed, and why it matters
An honest accounting of what this announcement does not contain is part of the story, not a footnote to it. No customer is named. No volume, ramp schedule, or revenue figure is given. No lane rate, aggregate bandwidth, power consumption, or reach is disclosed. No product name or part number is provided. The device is described as in “early production,” a term the release does not define — it is not volume production, and it is not a shipment announcement, and readers should not upgrade it into either. The tape-outs on the 300mm platform are described as supporting “future product generations,” with no dates attached.
Each omission has a plausible innocent explanation. Early production volumes are often small and customer-specific; lane rates may be covered by customer confidentiality; product names may await a customer launch. But the pattern of omissions also means the claims that matter most to a buyer — does it work at yield, at what cost, in whose system — are entirely untested in public.
That posture also disciplines how to read the forward-looking language. Tower’s release carries a standard safe-harbor notice, and its expectation that NPO architectures will drive demand beyond today’s pluggable market is the companies’ projection, not observed demand.
What to watch
The signals that will move this story from milestone to evidence are specific. First, a customer: a hyperscaler or switch vendor naming Innolight’s NPO device in a system, or Innolight naming a customer itself. Second, a volume marker: any disclosure of production ramp, wafer starts, or shipment timing that converts “early production” into a scale claim. Third, a performance disclosure: lane rate, aggregate bandwidth, or power per bit, ideally from a customer system rather than a datasheet. Fourth, the 300mm tape-outs: whether those future-generation products reach production on a similar cadence, which would suggest the PH18M device was a first step rather than a one-off. Fifth, the standards question: whether NPO form factors converge enough for second sourcing, or whether each vendor’s near-package engine remains a custom assembly.
There is also a competitive signal worth watching that has nothing to do with Tower or Innolight. If NPO devices in production start appearing in switch vendors’ roadmaps as a named option alongside pluggables and CPO, the architecture debate shifts from theory to procurement. If they do not, then the interim-step thesis weakens, and CPO’s longer, harder path looks more like the only path. Either outcome is informative.
The fair summary is a narrow one. Innolight’s first NPO device has entered early production on Tower’s PH18M silicon photonics process in Tower’s California factory, and multiple further NPO products have been taped out on Tower’s 300mm platform. No customer, volume, or performance figure has been disclosed. Set against a week in which the optical industry’s biggest consolidation was proposed but not closed, that is a modest claim — and precisely because it is modest, it is believable. The supply chain answered consolidation with a manufacturing fact. Whether that fact becomes a product line is the next story.
Sources and further reading
Tower Semiconductor and Innolight, “Tower Semiconductor and Innolight Advance High-Performance Near-Package Optics for AI Infrastructure,” November 3, 2026 — continued collaboration on NPO; first NPO device in early production on Tower’s PH18M silicon photonics process flow in Tower’s California factory; multiple NPO products taped out on Tower’s 300mm silicon photonics platform; NPO architectures typically integrating 16 or more channels per device with monolithic wavelength-division multiplexing/demultiplexing; Innolight a manufacturing customer across four Tower facilities (Newport Beach, California; San Antonio, Texas; Israel; Japan). Marvell Technology and Lumentum, “Marvell and Lumentum to Combine, Creating a Category Leader in AI Connectivity,” November 2, 2026 — definitive all-stock agreement; fixed exchange ratio of 4.500 Marvell shares per Lumentum share; implied enterprise value of $115 billion (approximately $1,132 per share, approximately 23.9% premium to October 28 unaffected closes); pro forma ownership approximately 67% Marvell / 33% Lumentum, fully diluted; expected close in the first half of calendar 2028, subject to stockholder and regulatory approvals and other customary conditions. Background on NPO versus pluggable and co-packaged optics, and on the unresolved debate over NPO form factors and serviceability, is drawn from general industry explainers; no lane rate, volume, customer, or power figure for Innolight’s device is asserted here, because none is disclosed in the announcement. This article is analysis, not investment advice.