Last week the optical communications industry gathered at the ECOC conference in Málaga, Spain. Four different companies presented four versions of exactly the same idea: stop cabling the optical fiber to the compute chip, and start putting it inside the same package.
And the reason is not speed. It is electricity.
Errors or inaccuracies are possible. Spotted something that looks wrong? Write to me and I will correct it.
In brief:
- The components that convert light to electricity and back consume roughly half a modern switch's power - per Broadcom.
- The answer the whole industry is showing is to shorten the distance between the optics and the chip.
- Coherent launched a full platform for it - and the stock fell the next day.
- And Israel has a seat at this table: Tower and NewPhotonics have started producing exactly this component in volume.
Why Now
When one AI accelerator sends data to another, the signal travels a chain: it leaves the chip as electricity, crosses the circuit board, reaches a component that converts it to light, travels down a fiber, and reverses the process at the other end.
Every step in that chain costs power, and the most expensive step is the electrical journey across the board.
Per a paper published in June in the peer-reviewed Journal of Lightwave Technology, by Alan Evans of MIT's Microphotonics Center:
| 2024 | 2029 (projected) | |
|---|---|---|
| Standard pluggable transceiver | 20.6 picojoules per bit | 11.25 |
| Low-power transceiver | 9 | 5 |
| Optics co-packaged with the chip | 6.9 | 1 |
And for a full switch at 102.4 terabits per second: 6,400 watts with the traditional approach, 4,983 with low power, 4,093 with co-packaged optics - a saving of about 36%.
And per Broadcom, pluggable transceivers consume roughly half a switch's power and more than half its cost.
When a data center is measured in megawatts rather than square metres, 36% of the networking layer's power is not an engineering improvement - it is a limit on how many accelerators fit in the same building.
What Coherent Presented
On 21 September Coherent launched PhotonLink, a platform bundling its whole optical chain - indium phosphide lasers, VCSEL arrays, silicon photonics, specialty fibers and detectors - into one offering matched to three architectures: optics inside the chip package, optics placed next to the chip, and direct chip-to-chip links.
What the company stated in its release:
- More than ten customer engagements in co-packaged optics, more than ten in near-packaged, and more than five in chip-to-chip.
- Anchor customers and long-term agreements signed in the first two architectures.
- "Revenue from the PhotonLink platform is expected to begin ramping in Q4 of CY2026" - the release's own wording.
- The industrial backdrop: the company said it has shipped more than 300 million indium phosphide lasers and more than a billion photodetectors to date.
And CEO Jim Anderson, in the release: "Customers need complete optical solutions, not just individual components."
And the figures from the last quarterly report explain why this is the centre of the business: in the quarter ended June, Coherent reported revenue of $2.05 billion, up 33.8% year on year. The datacenter and communications segment delivered $1.615 billion - 79% of revenue - and grew 58.6%. For the full year: $7.12 billion, up 22.5%.
Its constraint is not demand but production. The company said it produced about 80% more indium phosphide lasers in the June quarter than a year earlier, and that internal output is on track to double by the end of the September quarter - a quarter ahead of plan. Capital spending in a single quarter: $556 million.
And the Market? It Sold the Launch
Closing prices for Coherent (COHR, listed on the New York Stock Exchange):
| Date | Close | Change |
|---|---|---|
| 16 September | 289.93 | +6.92% |
| 17 September | 295.98 | +2.09% |
| 18 September | 317.36 | +7.22% |
| 21 September (launch) | 321.52 | +1.31% |
| 22 September | 310.18 | -3.53% |
The week as a whole: +7.06%. The day after the launch: -3.53% - while the Philadelphia semiconductor index rose 2.06%.
So the stock rose 9.5% in the three sessions before the launch, and fell on the day after. The obvious reading is that the market priced the product in advance - and that is my interpretation, not a fact: one session is not evidence, and other optics stocks also weakened that day while semiconductor stocks rose.
What can be said with certainty: the investment bank Stifel published an update the day after the launch in which it reiterated its rating and left its price target unchanged. The wave of target increases that surfaces in searches on the stock came from the 12 August quarterly report, not from the launch.
Who Else Presented at the Same Conference
| Company | What was shown |
|---|---|
| Lumentum, Qualcomm and Corning | A joint demonstration of direct optical chip-to-chip connectivity built on 1060-nanometre VCSEL arrays - about 10 terabits per second each way, 32 gigabits per lane, and density of about 1 terabit per millimetre with a path to four |
| Marvell | First optical components on a 2-nanometre process - 400 gigabits per lane and 800 gigabits for site-to-site links - and a co-packaged platform for 102.4 terabits |
| Nokia | A low-power 1.6-terabit transceiver, with a claim of about 80% less power than a traditional solution |
| Coherent (the day before the launch) | A 3.2-terabit transceiver with eight 425-gigabit lanes, and a 6.4-terabit near-packaged engine |
| Credo · MACOM | A 1.6-terabit module · components for 448 gigabits per lane |
Qualcomm rose 9.29% on 21 September. (The semiconductor index rose 4.29% that day, and Qualcomm also had its own product summit that week - so the demonstration explains at most the excess over the index, not the whole move.)
And what was conspicuously missing: almost none of the announcements cited a power-per-bit figure. In a category whose entire justification is saving power, the industry is selling bandwidth and density - and leaving efficiency to be measured by third parties.
The Israeli Link
On 17 September, four days before Coherent's launch, Tower Semiconductor and Israel's NewPhotonics announced the start of high-volume shipments of laser-integrated optical engines - precisely the component the whole conference was about.
Per the companies' release:
- The shipping product: NPG102, a photonic integrated circuit for 800-gigabit and 1.6-terabit optical engines, designed at 200 gigabits per lane.
- It is built on PH18DA, Tower's silicon photonics platform, which integrates laser, amplifier, modulator and detector on the same circuit.
- The next generation, NPC505 at 6.4 terabits, is planned to ship in volume in the first half of 2027.
- Doron Tal, SVP and GM at NewPhotonics: "The next phase of AI infrastructure demands optical connectivity that scales in both production and performance. With Tower, we are translating laser-integrated photonics into a practical choice and volume availability."
- Dr Ed Preisler, VP and GM of Tower's RF business unit: "We are excited to see our vision of integrating lasers with high-performance photonic integrated circuits come to fruition."
And the market reaction arrived late, for a technical reason: the Tel Aviv Stock Exchange was closed from 18 to 21 September for Yom Kippur. On the first session afterwards, 22 September, Tower shares rose 8.83% and led the Tel Aviv 35 index, on turnover of about 181 million shekels.
(Tower is also listed on Nasdaq, so part of the repricing happened while Tel Aviv was shut.)
הזווית שלי
דעה אישית של אילן אברמוב - לא ייעוץ ולא המלצה
What I take from this week is that AI's bottleneck has moved, and most of the public conversation has not moved with it.
The AI conversation runs on who builds the most powerful accelerator. But when tens of thousands of accelerators are wired into one cluster, the interesting question is how much power is burned simply moving data between them - and there, per the research, sits between a third and a half of the networking layer's budget.
And what interests me most is the gap between the announcements and the numbers. Every player showed bandwidth and density; almost nobody published power per bit. In a category whose entire economic logic is power, that is exactly the number I would want to see - and anyone selling an energy revolution without citing energy consumption is asking for trust.
As for the share reaction - Coherent rose almost 10% in the three days before the launch and fell the day after. In my view that says nothing about the product; it says something about buying ahead of a known event. An event whose date was announced in advance, as this one was, is priced before it happens.
And what I find most interesting from the Israeli angle: the link Tower and NewPhotonics have entered is not "another supplier". It is the link all four large players presented at the conference as their flagship - and the Israeli companies announced volume production in that same week.
(It is important to stress: this is my personal opinion only, and nothing herein constitutes a recommendation to take any action.)
Ilan Abramov, the author, or anyone on his behalf may hold some of the securities or funds mentioned in this article.
