Silicon excels at transporting electrons but struggles to generate light. This is a fundamental property of silicon's crystal structure—not something designers can engineer around. Ever since Richard Soref proposed the theoretical foundations of silicon photonics in the 1980s, researchers have worked within this constraint, fully aware of it from the outset. Now, a quarter century later, in 2026, a research team from Ghent University and imec is attempting to break through that barrier with a surprisingly simple idea: pasting materials on afterward.
The Materials CMOS Left Behind
The strength of silicon photonics lies in its ability to directly leverage the CMOS (complementary metal-oxide-semiconductor) manufacturing processes that dominate current semiconductor fabrication. Designers etch optical pathways (waveguides) and filters onto silicon wafers, mass-producing circuits that manipulate the optical signals traveling through data centers. Manufacturing costs have fallen to nearly the same scale as electronic ICs, and photonic integrated circuits (PICs) have become practical products in the telecommunications and datacom sectors.
However, the conservatism of CMOS lines severely limits this flexibility. Foundries impose strict controls on materials brought into the wafer fab, treating compounds containing lithium, phosphorus, or arsenic as "contamination sources." In other words, the very materials needed to generate light cannot enter these same CMOS lines. III-V semiconductors such as indium phosphide (InP) and gallium arsenide (GaAs) are, in this respect, fundamentally incompatible with silicon.
What's lacking isn't just light sources—modulator performance is also insufficient. Silicon modulators have limits in bandwidth and modulation efficiency, falling short of the standards required by next-generation transceivers like 800G and 1.6T. One candidate drawing attention as a solution is lithium niobate (, abbreviated LiNbO₃). It exhibits the Pockels effect—which converts electrical signals into optical signals—far more efficiently than silicon, but it too is a material that cannot be brought into CMOS lines.
The "Pasting" Solution
Micro-transfer printing (MTP) solves this problem by rearranging the order of the process steps.
The process begins with fabricating III-V devices and thin-film LiNbO₃ (TFLN) devices separately, each under optimal conditions, on a dedicated "source wafer" isolated from the standard CMOS line. Each device is created as a thin-film element called a "coupon," densely arrayed on the wafer, with a sacrificial layer sandwiched beneath it. Once the sacrificial layer is selectively etched away, rendering the coupons detachable, an elastomeric stamp made of polydimethylsiloxane (PDMS) comes into play. The stamp picks up multiple coupons at once and presses them onto precise locations on the target wafer—a silicon photonics wafer that has already completed its CMOS processing. Finally, the coupons are fixed in place using adhesive bonding or direct bonding, and the process is complete.
III-V devices and silicon photonics are "not made together." Each is fabricated separately using whatever process suits it best, and the two converge only at the final stage. This is what Ir. Ye Chen of Ghent University means when describing it as "a highly versatile technology that combines the advantages of die-level assembly with wafer-scale processing."
By adding materials forbidden from CMOS-compatible lines afterward, the approach draws out the performance of materials CMOS cannot handle, while leaving CMOS's design assets almost entirely intact. Different materials can coexist on the same wafer precisely because each has gone through an entirely separate process and does not interfere with the others.
Meeting Two Demands at Once: Precision and Diversity
The idea of "pasting things on afterward" has long been avoided in the photonics world due to concerns about precision. At coupling points where light passes between waveguides, losses spike sharply once design errors exceed several hundred nanometers. However, MTP's current positional accuracy has reached below ±0.5 µm—a level sufficient for evanescent-coupled photonic devices.
Moreover, the ability to select and transfer only "known good dies (KGD)" at the source wafer stage boosts overall system yield. In wafer bonding, defective devices cannot be sorted out after the fact, so defects discovered post-integration become outright losses. With MTP, defects can be excluded through pre-transfer inspection, and yield in demonstration tests has reportedly exceeded 95%.
What seems counterintuitive is the paradox that "integrating more materials may actually simplify the process." Because each material follows its own independent process flow without interfering with the others, parallel optimization becomes possible. InP is grown in III-V growth furnaces at several hundred degrees, while TFLN is fabricated through an entirely different low-temperature process. Since neither needs to enter the same process step, neither is forced into performance compromises. The greater the variety of materials involved, the more MTP's "divide-and-conquer" strategy demonstrates its strength.
Expanding Demonstrations
The breadth of demonstrations presented in the paper underscores the technology's versatility.
A photonic engine integrating InP lasers into silicon photonic circuits can process both optical and microwave signals on a single chip. An implementation combining GaAs lasers with Si₃N₄ waveguides has applications in virtual reality devices, quantum technology, and microwave photonics in view. A configuration integrating tunable-width, narrow-linewidth InP lasers with Si₃N₄ achieves the precise wavelength control required in both coherent communications and LiDAR.
Among these, the integration of thin-film LiNbO₃ (TFLN) shows the most dramatic numerical improvement. TFLN Mach-Zehnder modulators (MZMs) reach into territory that conventional silicon modulators could never access.
| Metric | Silicon Modulator | TFLN-MZM (Current Values) |
|---|---|---|
| Electro-optic bandwidth | ~30-50 GHz | 70-110 GHz |
| Half-wave voltage-length product (VπL) | ~10 V·cm | 1.15-1.9 V·cm |
| Propagation loss | ~1-3 dB/cm | <0.5 dB/cm |
| Extinction ratio | ~20 dB | >30 dB |
Bandwidth has improved 2-3x, and modulation efficiency (VπL) has improved more than 5-fold. Propagation loss falls below 0.5 dB/cm, and extinction ratio exceeds 30 dB—both far surpassing conventional silicon modulators. In June 2026, a preprint was published on arXiv demonstrating this TFLN-MTP technology on a 200 mm (8-inch) silicon wafer. A demonstration at the 200 mm wafer diameter standard for mass-production lines can be taken as an indicator that laboratory results are beginning to transition into manufacturing processes.
Why the Urgency Now
The fact that this technology is accelerating at this particular moment in 2026 is inseparable from the explosive growth of AI infrastructure. Data centers that train and run inference for large language models must move the enormous data traffic generated by GPUs and TPUs in real time—between chips, between racks, and between servers. Conventional copper wiring is approaching its limits in both bandwidth and power consumption, and the shift to optical interconnects is now moving forward as actual capital investment.
However, silicon-only optical devices cannot achieve the bandwidth or modulation efficiency demanded by next-generation standards such as 800G and 1.6T. The combination of III-V light sources and TFLN modulators provided by MTP is viewed as one of the few realistic solutions capable of meeting these demands.
The research team notes in the paper that they are also developing pilot lines aimed at mass production. While it is certain that MTP is beginning to step beyond research demonstrations into an industrial context, not all the elements necessary for commercial transition are yet in place.
Numbers Yet to Be Filled In
In the paper's conclusion, Ye Chen states that "consistent technical progress is bringing us closer to large-scale manufacturing," but this trajectory remains an optimistic outlook. Throughput at mass-production scale (number of coupons transferred per hour), long-term reliability data under high-temperature and high-humidity conditions, and the cost of adding MTP steps to CMOS processes—none of these figures are provided in the paper.
In particular, the maturity of the supply chain is explicitly noted as a challenge: building an ecosystem capable of stably supplying III-V and TFLN coupons in large volumes remains a problem independent of the technology's own level of completion. The reported 95% yield figure is also based on current laboratory conditions, and whether that level can be sustained in a mass-production environment is a question that will be answered by the track record of pilot lines.
The problem that silicon photonics researchers have grappled with for 30 years, knowing full well it "couldn't be done," is now being solved through a method of "separating and then bringing together later." Whether that solution will translate into real numbers on a manufacturing line is a question MTP has yet to answer.