
21 Sep 2026
ams OSRAM unveils Thin-Film VCSEL for AI scale-up links
Photonics company ams OSRAM announced a next-generation 850 nm Thin-Film VCSEL platform for AI scale-up optical interconnects, citing error-free 32 Gb/s operation near 0.25 pJ/bit and a connectorized multi-core fiber demonstrator with BizLink at ECOC 2026 in Malaga.
HARDWARE desk — AI clusters are hitting a data-movement wall; cheaper, lower-power optical lanes matter as much as the next GPU.
What the company says it will show. At ECOC 2026 in Malaga, ams OSRAM says it will showcase a connectorized multi-core fiber solution with BizLink. ECOC is the European Conference on Optical Communication. Malaga is the host city in Spain. Connectorized means the fiber ends in a plug, so it can be connected, not only fused on a lab bench. Multi-core fiber is one strand with several light-carrying cores inside it. BizLink is the partner named for the fiber and connector side of that demonstrator. A demonstrator is a working show piece at a conference. It is not a shipment into a customer’s data center. File the showcase as ams OSRAM’s.
What the platform is, still that page. ams OSRAM describes 850 nm Thin-Film microVCSEL arrays with addressability, a dense 25 µm pitch, silicon TSV integration, and compatibility with standard multimode fiber. 850 nm is the wavelength, the color of the light, in the near infrared. That color is a common choice for short data links. An array is many of those tiny lasers on one piece. Addressability means each laser can be switched on its own. Pitch is the spacing. 25 µm is 25 millionths of a meter, far thinner than a human hair, so the lasers sit very close together. A TSV is a through-silicon via, a vertical electrical path through the silicon, so the lasers can be stacked on the chip that drives them. Multimode fiber is glass that carries many light paths at once, the usual cable for short runs inside a building. File that list as ams OSRAM’s. The page does not say how many lasers are in one array. This filing does not invent a count.
The speed and energy lines are the company’s measurements, not this desk’s. ams OSRAM reports error-free 32 Gb/s NRZ operation. 32 Gb/s is 32 billion bits a second on a lane. NRZ, short for non-return-to-zero, means a simple on-or-off signal: the light stands for a 1 or a 0, without a more complicated way of packing extra bits into each flash. Error-free is the company’s test result. The page also says the platform reaches about 0.25 pJ/bit. A joule is a unit of energy. A picojoule is one trillionth of a joule. Per bit means that energy for each 1 or 0. A quarter of a picojoule is a very small energy budget, which matters when a cluster uses many lanes at once. The company also reports more than 2,000 hours of reliability validation under elevated conditions. The page’s own words are no failures after more than 2,000 hours under elevated junction temperatures and current overstress. Junction temperature is the heat at the working part of the laser. Overstress means the test pushed current harder than a normal setting. File those figures as ams OSRAM’s. This desk did not rerun the test, and it does not turn 2,000 hours into a lifetime or a warranty.
The design name on the page is wide-and-slow. ams OSRAM means many lower-speed optical lanes running side by side, instead of a few ultra-complex high-speed channels. The bet, in the release, is that a wide bundle of simpler lanes can move a cluster’s data with less power and less complexity than squeezing the same traffic through a handful of very fast links. File that frame as the company’s. It is a design argument in a press release. It is not a network this desk has seen installed.
Two people are quoted in the release. Ashkan Seyedi, vice president and general manager of Optical Interconnects at ams OSRAM, says AI infrastructure needs a different way to move data, and that wide-and-slow optical architectures deliver a step-change in power efficiency, latency, and scalability. Latency is the wait before data arrives. Hendrik Coldenstrodt, chief technology officer of the Computing and Transportation business group at BizLink, says the demonstrator validates a different optical interconnect architecture aimed at ultra-low power and low latency, by pairing ams OSRAM’s addressable microVCSEL technology with BizLink’s work on fiber arrays. File both lines as the release’s. This desk did not interview them.
Where a visitor can look. ams OSRAM invites ECOC attendees to booth #1281 and to the company’s presentations at the Global Photonics Economic Forum, which the page shortens to GPEF 2026. The same page says the fiber system will be demonstrated in a private setting for selected customers and industry partners. It does not name those customers. An invitation to a booth is not a price list, not a shipping volume, and not a roster of hyperscale buyers. Do not invent those. A conference demonstrator is not a product already installed in AI clusters.
Plain English for the rest of the card: VCSEL = a tiny laser that sends light straight out of a chip. thin-film = that laser built as a thin layer. 850 nm = the near-infrared color of the light. array = many lasers on one piece. addressable = each laser can be switched on its own. 25 µm pitch = how close those lasers sit. TSV = a vertical electrical path through the silicon. multimode fiber = short-reach glass that carries many light paths. 32 Gb/s = 32 billion bits a second on a lane. NRZ = a simple on-or-off light signal. pJ/bit = energy used per bit; lower is cheaper to run. wide-and-slow = many simpler lanes side by side, instead of a few very fast ones. scale-up = links inside one AI cluster. ECOC = the European optical-communications conference, in Malaga. This filing is the 21 Sep announcement and the planned BizLink demonstrator. It is not a shipment.
PRIMARY here: ams OSRAM’s 21 Sep 2026 press release — Tier A PRIMARY, the company’s own record. The Thin-Film VCSEL platform for AI scale-up, the ECOC 2026 Malaga showcase with BizLink, the connectorized multi-core fiber demonstrator, the 850 nm addressable microVCSEL arrays, the 25 µm pitch, the silicon TSV integration, the multimode-fiber compatibility, the error-free 32 Gb/s NRZ line, the about 0.25 pJ/bit line, the more-than-2,000-hour reliability line under elevated conditions, the wide-and-slow framing, the Seyedi and Coldenstrodt quotes, booth #1281, and the GPEF 2026 invitation are ams OSRAM’s. The private-setting line for selected customers, with no names printed, is also the page’s. NOT claimed: a customer name, a shipping volume, a price, that the part is already mass-deployed in hyperscale AI clusters, an array laser count, that this desk measured 32 Gb/s or 0.25 pJ/bit, that 2,000 hours is a warranty, a stock tip, or investment advice. Distinct from the already-filed mediatek-dimensity-cx-c10-max, cxmt-g5-dram-mass-production, and amd-1-trillion-market-cap.
RELATED
On 21 Sep 2026, ams OSRAM announced a Thin-Film VCSEL platform aimed at AI scale-up optical interconnects. The record is the company’s own press release, datelined Premstaetten, Austria, and Munich, Germany, 21 September 2026. A VCSEL is a vertical-cavity surface-emitting laser: a tiny laser that sends light straight out of the face of a chip, the kind already used to push data down short fiber. Thin-film, here, means that laser is built as a thin layer that can sit on other electronics. Scale-up, in this release, means the short links inside one AI cluster — the connections that let many chips share work — not the long cables between cities. These lines are ams OSRAM’s. This desk did not measure a link.
Sources
- ams OSRAM — Thin-Film VCSEL for AI scale-up
ams-osram.com