ECTC 2026 | Furukawa | Technical Analysis | VCSELs Can Go 2 km: Furukawa Fits a 1060nm Single-Mode VCSEL CPO into a Standard MT Ferrule
VCSELs have long been typecast as "cheap and power-efficient, but short-reach only" light sources: 850nm multimode VCSELs are killed by modal dispersion at around 100 meters. The 8-channel 1060nm single-mode VCSEL CPO transceiver that Furukawa Electric presented at ECTC 2026 turns that stereotype on its head: it runs 2 km over off-the-shelf 980nm single-mode fiber at 106-Gb/s PAM4 per lane with a TDECQ of just 2.53 dB, and with all 8 channels operating simultaneously the link energy comes in at 4.1 pJ/bit. More important than the numbers themselves is that it replaces the previous generation's multicore fiber (MCF) with a standard 24-lane MT ferrule plus off-the-shelf single-mode fiber. That step is the real watershed that moves a lab demo toward something the industry can actually adopt.
1. Background: A Veteran Team Moves VCSEL CPO from Showpiece to Deployment
The paper is led by Furukawa Electric together with FUJIFILM Business Innovation and Professor Fumio Koyama of the Institute of Science Tokyo, and was presented at the 2026 IEEE 76th Electronic Components and Technology Conference (ECTC 2026). Koyama is a VCSEL veteran, and the line-up itself signals the direction: not another faster VCSEL, but getting a single-mode VCSEL "into a package the industry is willing to use."
To appreciate the weight of this paper, you need to know the team's previous step. At ECOC 2025 / OFC 2026 they showed a 16-channel, multicore-fiber (MCF) 1060nm single-mode VCSEL CPO with link energy as low as 3.95 pJ/bit, which we broke down in full in OFC 2026 - 800G CPO Breakthrough: 1060-nm Single-Mode VCSEL Achieves 2 km Transmission and 4.1 pJ/bit. The problem is that MCF is not something data centers buy off the shelf. What this ECTC paper does is "downshift" that technology from MCF back to standard interfaces, giving up a sliver of efficiency (3.95 → 4.1 pJ/bit) in exchange for fiber and connectors the whole supply chain already has.
The real innovation here is not lower power, but that "it finally plugs into existing fiber infrastructure."
2. The Core Question: Can the VCSEL's Power Advantage Survive 2 km?
The data-center pain point is straightforward. OIF's 2024 System Vendor Requirements document (front-end interconnect, FEI) nails down the spec: reach must cover 0 to 500 meters, extending to 2 km in some scenarios, while link energy must be below 10 pJ/bit.
The problem sits between two extremes. On one side are VCSELs: low drive current and inherently good efficiency, the workhorse of short-reach interconnect in AI/ML clusters, but 850nm multimode solutions are limited by modal dispersion and cannot carry 100-Gb/s PAM4 beyond 100 meters. On the other side are single-mode solutions such as LPO (linear-drive pluggable optics): they reach far, but link energy is around 5 pJ/bit, and the bulky pluggable form factor cannot be packed densely next to the switch ASIC, capping total bandwidth.
Furukawa's bet is to use a 1060nm single-mode (SM) top-emitting VCSEL to get both the VCSEL's low power and single-mode reach. The paper sets out to answer one question: can a single-mode VCSEL deliver a clean eye at 2 km and 106-Gb/s PAM4 while holding the FEI efficiency line?
3. Key Figures, One by One
What this figure shows: how the whole transceiver fits into 7.7 × 15.9 × 7.95 mm
The cross-section in Figure 1 is the foundation for the whole paper. Two 4-channel SM VCSEL and PD arrays are precisely mounted on an organic substrate and connected to two 4-channel VCSEL drivers (VD) and trans-impedance amplifiers (TIA). The bottom is a 0.3mm-pitch LGA (land grid array), a very tight pitch that is a prerequisite for high-density mounting. Both the VD and TIA are commercial 53-Gbaud linear-drive parts, deliberately omitting clock-data recovery (CDR) to cut power: essentially LPO's power-saving philosophy moved into CPO. The VCSEL array's 3-dB bandwidth exceeds 28 GHz, and the PD has a 16µm aperture with bandwidth above 26.5 GHz.
What this figure shows: two-lens single-mode coupling holds misalignment tolerance to ±1.0µm
Figure 2 is the most hardcore engineering detail of this transceiver. Single-mode coupling is unforgiving of misalignment, and matching the mode-field diameter (MFD) between a VCSEL and single-mode fiber is especially sensitive. Furukawa mounts two microlens arrays on a precision-machined lens spacer, and measurement matches calculation: both TX and RX coupling loss can be held below 1 dB, and with the assembly process keeping optical-axis offset within ±1.0µm, actual TX/RX coupling loss still stays below 2 dB. Notably, RX has generous alignment tolerance because the PD aperture (16µm) is far larger than the single-mode fiber MFD; TX is the narrow-tolerance bottleneck, so active alignment prioritizes TX.
What this figure shows: ample bandwidth, no inherent bottleneck for 53-Gbaud operation
The E/O and E/E responses in Figure 4 confirm the right components were chosen. The E/O 3-dB bandwidth reaches 29 GHz, well above the 26.5 GHz Nyquist frequency needed for 53-Gbaud operation; the loopback optical link's 6-dB bandwidth is 28 GHz as well. In other words, the transceiver's bandwidth is not scraping by; it has headroom for 53-Gbaud PAM4. The E/O response is deliberately lifted around 25 GHz by the VD's built-in emphasis, a typical linear-drive technique that relies on equalization rather than CDR to restore the signal.
What this figure shows: after 2 km, the 106G PAM4 eye is still open
The real report card is in Figures 10 and 11. Single-lane 106-Gb/s PAM4 (PRBSQ 2^13−1) shows clearly open eyes at back-to-back, 500m, 1km and 2km; the worst TDECQ at 2 km is only 2.53 dB, far below the 3.4 dB limit in IEEE802.3df. BER measures 1.9 × 10⁻⁶, more than two orders of magnitude below the KP4-FEC threshold of 2.4 × 10⁻⁴. With all 8 channels running over 2 km (the others acting as aggressors), the monitor channel's BER also holds at 1.9 × 10⁻⁶. Even more impressive, 53-Gb/s NRZ keeps the eye open out to 3 km, meaning the link not only meets the 2 km FEI target but has room to extend further.
The paper's two most innovative points are not about chasing record numbers, but about "spec convergence":
First, a standardized optical interface. The previous generation's MCF is replaced with a standard 24-lane MT ferrule plus ribbonized off-the-shelf 980nm single-mode fiber. The VCSEL/PD array aperture pitch is set to 250µm, matching the fiber layout of a standard MT ferrule exactly. That means data centers don't need to lay a new fiber network for it: existing single-mode fiber and existing MT connectors just plug in. 980nm single-mode fiber is a mature product originally made for 980nm pump lasers, with a cutoff wavelength below 970nm and about 1.6 dB/km loss at 1060nm, and it is readily available.
Second, the efficiency combination of linear drive plus single-mode VCSEL. With no CDR and commercial linear-drive VD/TIA, total power with all 8 channels running is just 3.5W, which works out to a link energy of 4.1 pJ/bit, comfortably inside OIF FEI's 10 pJ/bit line, while packing 800G-class bandwidth (8 × 106G) into a tiny 7.7 × 15.9 mm package. This path is worth comparing with Intel's circuit-level approach that pushes VCSEL CPO to sub-1 pJ/bit, which we analyze in full in Intel Takes VCSEL CPO to sub-1 pJ/b. Both prove the same point: VCSELs should not be locked into short reach.
5. Industry Implications: How Far from Volume Production, and Who Benefits?
First, the reality check: this is still a demo-grade paper using a test bench and PPG/DCA measurements, not volume-production yield data. But every engineering choice points toward "manufacturable": commercial drivers/TIAs, off-the-shelf fiber, a standard ferrule, and active alignment controllable to ±1.0µm. That is production-line language, not lab language.
Looking at the supply chain, the beneficiaries are clear. On the VCSEL epitaxy and die side, demand for 1060nm single-mode VCSELs will directly pull related foundry and epi houses (WIN Semiconductors, Lumentum and others are positioning in 1060nm VCSELs; see our earlier OFC coverage for Lumentum's 1060nm VCSEL interconnect solution). On the optical transceiver assembly side, Furukawa and FUJIFILM turn microlens coupling and precision mounting know-how into a barrier to entry. This VCSEL path is not a zero-sum fight with DFB/EML or silicon photonics (SiPh) with external lasers; the market tiers by reach and power: VCSELs take the short-to-mid reach of scale-out, while EML/SiPh hold longer reach and higher per-lane rates. This tiering logic becomes clearer in the context of CPO's overall evolution; we recommend reading it alongside The Great Shift in Optical Packaging (Part 2): CPO's Three-Stage Evolution.
Some cold water is in order too: the single-mode VCSEL's TX alignment tolerance is only ±1.0µm, so alignment yield and consistency in volume production will decide whether the costs work. The 2 km reach relies on the low-dispersion window of 980nm fiber at 1060nm; this wavelength/fiber pairing leans toward a "custom coincidence," and whether it can slot painlessly into the data center's existing 1310/1550nm ecosystem remains an open question.
6. Conclusion
Placed on the timeline of technology history, this paper is not a point breakthrough of "VCSELs running farther," but the step that pulls single-mode VCSEL CPO from "needing specialty fiber like MCF" back to "plugging into a standard MT ferrule plus off-the-shelf SMF." For STT readers, the signal to remember is this: when VCSELs simultaneously achieve 2 km reach and 4.1 pJ/bit efficiency while accommodating existing fiber infrastructure, the CPO scale-out battlefield gains a player with a completely different cost structure. The EML and silicon photonics camps should start treating single-mode VCSELs as a real competitor rather than a supporting act confined to in-rack links. The next thing to watch is whether this alignment process can deliver volume-production yield; that will decide whether it stays on paper or makes it onto the production line.
References
W. Yoshida, Y. Iwane, K. Nagashima, K. Takeda, S. Yoneyama, H. Nasu, F. Koyama, "A 106-Gb/s × 8-Channel 1060-nm Single-Mode VCSEL-Based Ultra-Compact CPO Transceiver enabling 2-km Parallel-Optical Links," 2026 IEEE 76th Electronic Components and Technology Conference (ECTC), DOI: 10.1109/ECTC51846.2026.00020. Affiliations: Furukawa Electric Co., Ltd. / FUJIFILM Business Innovation Corp. / Institute of Science Tokyo.
W. Yoshida et al., "An Ultra-Compact 50-Gbaud × 16-Channel CPO Transceiver employing a 1060-nm Single-Mode VCSEL array and Multicore Fibres," ECOC 2025, Tu.01. (previous-generation 16-channel MCF version)
Optical Internetworking Forum, "System Vendor Requirements Document for Energy Efficient Interfaces," OIF-EEI-Requirements-RD-01.0, Mar. 2024.
IEEE 802.3df standard (basis for the 3.4 dB TDECQ spec).
Comments