PASADENA, 23 AUG 2026 — Researchers at Caltech have fabricated optical waveguides on a semiconductor chip that approach the loss performance of optical fibre, using germanium-doped silica — the material fibre itself is made from — deposited with standard lithography and surface reflow.
The devices reached optical quality factors above 180 million in ring resonators, corresponding to waveguide losses below 0.1 dB per metre in the telecom band. At visible wavelengths the improvement over silicon nitride is up to twentyfold, with more than a hundredfold gain in laser coherence. The work was published in Nature.
The result
The stated applications span optical clocks, atomic sensors, data-centre communications, and quantum and precision measurement systems.
Why loss is the number that governs everything else
Optical loss is a measure of how much light disappears as it travels. On a chip, it has historically been the main constraint on performance.
Fibre is extraordinary at this. Light travels kilometres through glass with a small fraction lost, which is why intercontinental communication runs on it rather than on copper. Integrated photonics has never come close, because the materials and geometries that work at chip scale scatter and absorb far more, and every device built on them inherits that penalty.
Loss compounds, and it does so through quality factor — the number of times light circulates in a resonator before fading. That figure then propagates outward. It sets the linewidth a laser can reach. It bounds the smallest signal a sensor can pick out. It caps how many components you can chain together before there is nothing left to work with. So a jump to 180 million does not improve one specification. It relaxes the constraint that had been bounding an entire family of designs.
Using the fibre material is the obvious idea, and the hard part is fabrication
Germanium-doped silica is not a new discovery. It is what optical fibre has been made from for decades, precisely because of its very low absorption.
The difficulty has always been getting it onto a chip. Fibre is drawn from a preform at high temperature into a long uniform strand; a chip waveguide is patterned lithographically in a planar process with entirely different thermal and geometric constraints. Depositing the fibre material and shaping it without introducing the surface roughness that causes scattering is the problem, and roughness is what has limited planar waveguides.
The reported approach uses standard semiconductor lithography followed by surface reflow, briefly heating the patterned structure so that surface tension pulls it smooth. The roughness problem was solved in the process rather than in the material, which is the reason the CMOS-compatibility claim deserves more attention than the loss figure does.
Demonstrate something on bespoke laboratory equipment and you have physics. Demonstrate it on the tooling foundries already run and you have something that could plausibly be manufactured without anyone building a new plant for it.
Where this would show up first
The visible-wavelength improvement points at the near-term application more clearly than the telecom number does.
Silicon nitride is the incumbent platform for visible-light photonics. That part of the spectrum matters because it is where atomic physics happens — the transitions used by optical clocks, atomic sensors, and many quantum systems. A twentyfold loss improvement and hundredfold better laser coherence in that band is aimed squarely at instruments that currently occupy an optical bench.
Optical clocks are the clearest case. They are the most precise instruments humans have built, they are laboratory-scale, and shrinking one onto a chip would change what timing infrastructure is possible — including for navigation systems that must work where satellite signals do not reach.
Data-centre communication is also a listed application. It is the largest potential market, but likely the slowest to adopt the technology.
The data centre case is real and slower than it sounds
Interconnect is a hard constraint on large AI systems. Moving data between accelerators consumes power and adds latency, and the industry has been shifting toward optical links precisely because electrical ones stop scaling.
Lower-loss on-chip optics would help. But a data centre link is judged on cost per bit, yield across millions of units, packaging, temperature stability and integration with electronics, and a laboratory quality factor is silent on all of those.
The history of silicon photonics is a long sequence of impressive laboratory results followed by a decade of manufacturing work before anything shipped in volume. Nothing here suggests that pattern has changed, and the honest expectation is that precision instruments arrive first because they tolerate cost and volume constraints that data centres do not.
Why this is worth watching from this region specifically
Photonics is unusual among semiconductor technologies because it does not require the most advanced process nodes.
Leading-edge logic demands fabrication capability concentrated in very few places. Photonic devices are patterned at far coarser dimensions, which means they can be manufactured in mature-node facilities of the kind that exist across Singapore, Malaysia and increasingly Vietnam and the Philippines.
That makes photonics one of the few frontier-adjacent categories where regional capacity is not structurally excluded by node access. If chip-scale optics becomes a significant component class, there is no obvious reason the fabrication ends up only where leading-edge logic already is.
The qualifier is that photonics packaging is hard. Aligning fibres to chips with sub-micron tolerance is a specialised discipline, and regional opportunity depends more on developing that capability than on lithography.
Aligning an optical fibre to a waveguide means positioning it accurate to a fraction of a micron and holding it there through years of thermal cycling. The assembly and test industry already established across Malaysia, the Philippines and Singapore sits closer to that discipline than to lithography, which makes existing packaging capability the asset here rather than any ambition to fabricate.
What remains unconfirmed
The material reviewed is secondary coverage of a paper rather than the paper itself. Device dimensions, the number fabricated, yield and reproducibility across wafers are not described, and a quality factor from a best device is a different claim from one that manufactures repeatably.
The paper is silent on several manufacturing questions: whether the reflow process is compatible with other on-chip structures, how the waveguides couple to external fibre, and what their thermal and long-term stability is. It gives no cost comparison against silicon nitride and names no commercial partners.
What to watch for
Independent reproduction is the first test, and it matters who does it. A foundry repeating this on a production line would be establishing something a research group repeating it cannot, because CMOS compatibility asserted in a paper is a weaker claim than CMOS compatibility survived at volume.
The second is whether anyone reports yield. One resonator hitting 180 million tells you the physics works; what a manufacturer needs to know is how the whole wafer distributes around it, and nobody has published that.
The third is which application attracts funding first. Investment in chip-scale optical clocks and atomic sensors would fit the reading above. Money going instead into data-centre interconnect would mean somebody with more information than us thinks the manufacturing gap has closed faster than the field's history suggests.
View author profile → · Editorial policy
