This is not breaking news. On September 17, 2026, Tower Semiconductor and NewPhotonics announced high-volume shipments of laser-integrated, serviceable optical engine photonic integrated circuits (PICs) for AI interconnect, built on Tower’s PH18DA silicon photonics platform. The announcement surfaced in trade coverage in mid-September, then slipped out of the daily conversation. It is worth returning to now: it is the foundry-level proof point this desk’s co-packaged-optics arc has been waiting for.
The claim, stated plainly: NewPhotonics’ NPG102 PIC products, for 800G and 1.6T optical engines at 200G per lane, are in volume shipment, with heterogeneously integrated InP lasers on a monolithic silicon photonics chip. Tower calls NewPhotonics the first supplier to bring such engines on PH18DA into high-volume production, and a CPX-MSA-compliant 6.4T near-packaged optics (NPO) chipset, the NPC505, is slated for volume shipment in the first half of 2027. That is a production milestone about the hardest part of co-packaged optics. Whether it is also a deployment milestone, the announcement does not say.
What was actually announced
Precision first. Tower Semiconductor (NASDAQ/TASE: TSEM) and NewPhotonics, a Tel Aviv-based fabless photonics company, announced on September 17, 2026, that high-volume shipment had begun of laser-integrated serviceable optical engines for AI interconnect. The PIC optical engine chips support OEM solutions spanning FRO, LRO, and LPO pluggable transceiver modules, Extra-Dense Packaged Optics (XPO) pluggables, and NPO socket pluggable applications. Shipping volume-production PICs support data rates from 800G to 1.6T, with future volume manufacturing planned at 6.4T NPO solutions.
The technical core is the platform: Tower’s PH18DA silicon photonics (SiPho) platform, with heterogeneously integrated InP components enabling lasers, semiconductor optical amplifiers (SOAs), modulators, and photodetectors on a single PIC. The release argues this integration cuts complexity and variability, improving yield, reliability, and time-to-market. Doron Tal, senior vice president and general manager of NewPhotonics, said the companies are “translating laser-integrated photonics into a practical choice and volume availability.” Dr. Ed Preisler, vice president and general manager of Tower’s RF business unit, said NewPhotonics “is a valued partner and the first to bring heterogenous laser-integrated optical engines on the PH18DA platform into high-volume production.”
The announcement was timed to ECOC 2026, September 21–23 in Malaga, where both companies planned to show the products. The 6.4T volume date is a forward-looking statement: a plan, not a shipment.
Why the laser is the hard part
In a conventional pluggable, the laser is a discrete component, often a separate InP package, attached to the silicon photonics chip during module assembly — and a bad laser costs a bad module, a survivable failure. In co-packaged and near-packaged optics, the optical engine sits on the switch package substrate or on the board beside the switch, committed to an expensive assembly — a switch ASIC and its package — before its performance is fully known. A laser that fails late in that flow does not cost a module. It costs the assembly it was bonded into.
The laser is also the component silicon photonics cannot make for itself: silicon is an indirect-bandgap material and does not lase efficiently, so any silicon photonics platform that wants an on-chip light source must borrow one from a III-V material — indium phosphide is the standard choice. The older way attaches a discrete laser after the fact: tested before assembly, but one more step, one more alignment, one more failure mode. The harder way is heterogeneous integration — bonding InP material onto the silicon wafer and processing lasers, SOAs, modulators, and detectors together on one chip, in the foundry. That is what PH18DA does.
Heterogeneous integration is hard in the specific way that matters for volume: bonding III-V material to silicon at wafer scale, at the uniformity and defect density a production line requires, is a process-control problem first. A lab laser and a foundry-yield laser are different achievements. When Tower says NewPhotonics is the first to bring heterogeneous laser-integrated engines on PH18DA into high-volume production, the word doing the work is “production” — not “demonstrated,” not “sampled.” The engines are also described as serviceable — socket-pluggable NPO engines that can be field-replaced — though the release details no socket mechanism and no field-service result.
The PH18DA platform, and what “first” means
The “first” claim needs care, because PH18DA has more than one tenant. Trade coverage noted an earlier milestone: OpenLight reported in March 2026 that it had received what it called the first-ever volume production orders by a customer on PH18DA, developed with Tower. The two claims are not contradictory, but they are not the same: OpenLight’s was about volume production orders on the platform; Tower’s is about shipment of a specific product class into high-volume production.
What the September announcement adds is a shipped product — the NPG102 — with a named lane rate and rate range, on a platform that already had a production order behind it. That sequence — order first, shipment second — is the normal order of things in semiconductors. It is not, however, evidence about yield, and neither announcement supplies a yield figure.
PH18DA is a merchant foundry platform — a process any fabless photonics company can design to, not a captive line inside a vertically integrated vendor. When a merchant platform reaches volume shipment with a laser-integrated product, the process is available to the next fabless company with a similar design: a foundry milestone, not just a product launch. The supply chain gains a production-proven route to laser integration.
What “volume” does and does not establish
“High-volume shipment” is the release’s phrase, and it is doing a great deal of work. What it establishes is that NewPhotonics is shipping NPG102 PICs in production quantities — not sampling, not prototyping — to customers building 800G and 1.6T engines. What it does not establish is how many, to whom, at what yield, or at what price. No customer is named, no volume, ramp schedule, or revenue figure is given, and no yield, power-per-bit, or reach figure is disclosed.
Each omission has an innocent explanation: customers in AI infrastructure do not generally permit suppliers to name them, and volumes, pricing, and yield are commercially sensitive. But the pattern sets the limit of what can be claimed. This article asserts no customer, no shipment quantity, no yield, and no power figure for the NPG102, because none is on the record. Any coverage attaching a hyperscaler name or a unit volume is inventing it.
The safe arithmetic is the release’s own: a 200G-per-lane design reaches 800G in four lanes and 1.6T in eight. The 6.4T figure for the NPC505 is a product-family target tied to a volume date in the first half of 2027, not a shipped rate, and should be read in the future tense. A caution: “volume production” for a PIC product and “volume” for a hyperscale deployment are different scales, and the release does not say which it means. The milestone is real; its scale is undisclosed. It is also seven weeks old, and in seven weeks no customer, volume, or performance disclosure has followed it — the announcement is the whole of the public record.
The 6.4T roadmap, and the CPX socket
The forward-looking half of the announcement is the NPC505. The 6.4T chipsets are described as CPX-MSA-compliant, with volume shipment to begin in the first half of 2027. CPX — the Co-Packaged Optics Multi-Source Agreement — is the industry effort to define a common socket and interface for co-packaged and near-packaged optical engines, so one vendor’s engine can plug into a switch system designed for another’s. Compliance is a claim about interchangeability.
Trade coverage added one detail the release does not carry: NewPhotonics had recently introduced the 6.4T NPC50506 DR32 optical engine, combining an integrated laser with 224 Gbps PAM4 operation in a socketed, field-serviceable NPO architecture. That detail comes from Converge Digest, not from Tower’s release, and should be read as trade-press reporting rather than a company disclosure. At face value, the 6.4T step is not only faster lanes but a defined socket, a defined modulation format, and a serviceability story to match.
“First half of 2027” is a window, not a date, and Tower’s own safe-harbor notice applies to it. The credible reading: the NPC505 is the product the NPG102’s volume shipment is meant to de-risk — if the laser-integrated PIC ships at volume today, the higher-density chipset on the same integration approach is a scaling step rather than a new invention. Whether that step lands in the first half of 2027 is the question the next two quarters will answer, visible first in a shipment announcement.
Where this sits in the desk’s arc
This announcement does not arrive in isolation. On November 3, this publication covered Tower and Innolight’s first near-packaged optics device entering early production on Tower’s PH18M process in Tower’s California factory. On November 4, it covered ficonTEC and HTSI’s cooperation to industrialize double-sided wafer test for co-packaged optics in Taiwan.
Read together, the three stories describe a supply chain assembling itself: ficonTEC–HTSI is the test step, finding bad photonic dies at wafer level before they are bonded into expensive assemblies; Tower–Innolight is the device step, a first NPO device in early production on a foundry line; and Tower–NewPhotonics, seven weeks older and only now getting its full hearing, is the component step that makes the other two possible at scale — the laser, integrated on the chip, shipping in volume.
The sequence clarifies the difference between the two Tower announcements. Innolight’s device is in early production — a defined, limited status that is explicitly not volume production. NewPhotonics’ NPG102 is in high-volume shipment, a stronger claim about a different product on a different process flow (PH18DA with heterogeneous InP integration, versus Innolight’s PH18M). The two should not be conflated, and neither should be upgraded into the other.
A further contrast from that same week: on November 2, Marvell and Lumentum announced a proposed all-stock combination that would consolidate the laser supply chain under one roof, subject to shareholder and regulatory approvals, closing well in the future. That is a consolidation story: the optical layer is valuable enough to acquire. Tower–NewPhotonics is a manufacturing story: the laser can be integrated on a merchant foundry platform and shipped at volume today. Consolidation answers who owns the supply chain; production answers whether it can make the next thing.
What to watch
The signals that move this story from milestone to evidence are specific. First, a customer: an OEM or switch vendor naming the NPG102 in a system, or NewPhotonics naming a customer itself. Second, a volume marker: any disclosure of ramp, wafer starts, or shipment timing that converts “high-volume” into a scale claim. Third, a performance disclosure: power per bit, reach, or yield, ideally from a customer system. Fourth, the NPC505: whether the 6.4T chipset reaches volume shipment in the first half of 2027 on schedule. Fifth, the platform signal: whether another fabless photonics company announces a laser-integrated product on PH18DA.
A competitive signal, too: if socketed NPO engines in volume shipment start appearing in switch vendors’ roadmaps as a named option alongside pluggables and CPO, the serviceability argument for NPO gains procurement weight. If they do not, the laser-integration milestone stays a supply-side fact waiting for a demand-side confirmation. Either outcome is informative.
The fair summary is narrow. NewPhotonics’ NPG102 PICs, with heterogeneously integrated InP lasers on Tower’s PH18DA platform, are in high-volume shipment for 800G and 1.6T optical engines at 200G per lane, and a CPX-MSA-compliant 6.4T NPC505 chipset is planned for volume shipment in the first half of 2027. No customer, volume, yield, or power figure has been disclosed. Against a season of consolidation, tape-outs, and industrialized test capacity, that is a modest claim — and because it concerns the hardest step in the stack, it is the one worth believing first. The laser is on the chip, and the chip is shipping. Whether it is shipping at the scale the AI buildout needs is the next story.
Sources and further reading
Tower Semiconductor and NewPhotonics, “NewPhotonics® and Tower Semiconductor Begin High-Volume Shipments of Laser-Integrated, Serviceable Optical Engine PICs for Scale-Out and Scale Up AI Interconnect,” September 17, 2026 — the announcement itself: NPG102 PICs for 800G and 1.6T at 200G per lane; PH18DA platform with heterogeneously integrated InP components; CPX-MSA-compliant NPC505 chipsets for 6.4T to begin volume shipment in the first half of 2027; ECOC 2026, September 21–23, Malaga. Converge Digest, “NewPhotonics and Tower Move 800G/1.6T Laser-Integrated PICs into High-Volume,” September 2026 — trade coverage noting the NPC50506 DR32 optical engine with 224 Gbps PAM4 operation; trade-press reporting, not a company disclosure. Unite.ai, “Tower and NewPhotonics Ship Laser-Integrated PICs for AI Interconnect,” September 2026 — trade coverage noting OpenLight’s March 5, 2026 report of first-ever volume production orders on the PH18DA platform; reported context. Excitonic, “Near-Package Optics Gets Its First Production Device,” November 3, 2026 — Innolight’s first NPO device in early production on Tower’s PH18M process in Tower’s California factory. Excitonic, “Testing Is the CPO Bottleneck,” November 4, 2026 — ficonTEC–HTSI cooperation on the INS-2 double-sided wafer test platform. No customer, volume, yield, power, or price figure for the NPG102 is asserted here. This article is analysis, not investment advice.