Not long ago, NVIDIA's Jensen Huang stated that next-generation AI infrastructure will require massive amounts of optical connectivity, as copper cabling can no longer meet the demands. This is no exaggeration.
We Are Stepping Into a World of Light
With the rapid development of information technology, global data traffic is growing exponentially, and the demand for information capacity and processing power continues to rise. Driven by emerging technologies such as 5G communications, the Internet of Things, cloud computing, big data, and artificial intelligence, traditional electronic communication systems are increasingly facing bandwidth bottlenecks and high energy consumption. Optical communication technology, with its significant advantages of high bandwidth, low loss, and immunity to electromagnetic interference, has become a key solution to these challenges. The core reason next-generation AI infrastructure must rely heavily on optical connections is that the "interconnection wall" has replaced computing power as the biggest bottleneck. As GPU clusters scale up to tens of thousands of cards and single-channel rates head toward 224G, copper cabling hits physical limits due to skin effect and dielectric loss, compressing effective transmission distances to less than 2 meters-insufficient for cross-rack scale-out needs. At the same time, all-optical interconnects can reduce per-bit bandwidth power consumption by over 40%, making them the only path to solving the energy crisis in AI factories.
Lithium Niobate: Decades on the Cold Bench
As a key component of optical communication systems, the electro-optic modulator (EOM) converts electrical signals into optical signals and performs modulation. Its performance directly affects the transmission rate, energy consumption, quality, and stability of the entire communication system.
Lithium niobate (LiNbO₃, LN) is a crucial electro-optic material. With its excellent Pockels effect, high refractive index (~2.2), broad transparency window (350 nm–5 μm), and good chemical stability, it is respected in the photonics community as "optical silicon." Since the 1960s, it has been widely used in electro-optic modulators.
However, while it was indispensable at the system level, it was left out of the wave of chip-scale integration for three decades. This is because conventional bulk lithium niobate modulators rely on electric fields to control optical phase or intensity. Limited by the material's physical properties and processing techniques, the waveguide dimensions of bulk LN are on the order of millimeters to centimeters, resulting in a short interaction length between the optical and electric fields. To achieve effective modulation, high driving voltages (several to tens of volts) are required. The large device size makes it difficult to integrate with silicon-based photonic platforms, limiting its use in chip-scale integrated optoelectronic systems. Additionally, conventional fabrication processes suffer from high waveguide propagation loss, further restricting energy efficiency and long-distance transmission. As a result, platforms such as silicon photonics, InP, and SiN rose to prominence, and LN was once considered "great performance, but can't be made small or dense."
The Breakthrough of Thin-Film Technology, Arriving Just as Demand Calls for It
The turning point came with the maturation of thin-film lithium niobate (TFLN) technology. TFLN is based on a heterostructure of "lithium niobate–insulator–substrate." Using advanced fabrication techniques such as crystal ion slicing and chemical mechanical polishing, a single-crystal LN thin film is peeled from the bulk material and transferred onto a substrate (silicon, sapphire, or silicon dioxide). Compared to bulk material, TFLN's sub-micron waveguides enable much stronger optical field confinement, increasing the light–electric field interaction efficiency by tens of times, thus significantly lowering the driving voltage and reducing device size. Additionally, TFLN's low propagation loss gives it a unique advantage in long-haul photonic integrated circuits, and its compatibility with silicon-based platforms opens new paths for heterogeneous integrated photonics.
Let's look at a few key metrics to understand why it's being "suddenly" snapped up in the 1.6T/3.2T era:
① Bandwidth: easily exceeds 100 GHz, heading toward 200 GHz.
② Power consumption: only around tens of femtojoules per bit (fJ/bit).
③ Signal quality: low insertion loss, minimal chirp, excellent linearity.
④ Versatility: a single platform that handles electro-optic, nonlinear, and quantum applications.
On the industry demand side, with AI computing power exploding, data center optical interconnects are moving from 400G to 800G/1.6T/3.2T, exactly the era that needs TFLN. Take the current hot topic of co-packaged optics (CPO): it moves the optical engine from the front-panel pluggable module onto the same package substrate as the switch chip/ASIC. After NVIDIA mass-produced CPO solutions on its Spectrum-X and Quantum series, measured data showed stunning results-insertion loss dropped from about 22 dB to ~4 dB, signal integrity improved by a factor of ~63, and system optical power efficiency increased by up to 5×.
But CPO is not simply about "relocating" existing optical modules. The package volume shrinks drastically, power budgets are cut to the bone, heat dissipation conditions worsen, and the electrical environment becomes extremely harsh-every device inside the optical engine is pushed to its physical limits. Under this new set of constraints, TFLN has arrived at the perfect moment, evolving from a "performance benchmark" to an "engineering necessity."
In short, the reason thin-film lithium niobate has become so hot is not just that it has been made thinner-but because the edifice of computing power has finally risen to the floor where TFLN must serve as the load-bearing wall.