mxx1's avatar

PsiQuantum

Silicon nitride photonics has become a well-funded path to large-scale quantum computing. PsiQuantum’s are using this approach to build their Omega platform quantum computer where every qubit is a photon. The high-level benefit is that decoherence is not the issue it is in matter-based quantum computing because photons are quite robust at maintaining their quantum state so long as you keep them in the right environment. Silicon nitride is low loss, has useful Kerr nonlinearity, and supports polarization control, although it is not inherently polarization maintained.

However, I think all they seem to have done is transferred an intractable decoherence problem (usually solved with near-zero-Kelvin engineering) to an intractable optical design and integration problem.

This is how it works:

  1. An edge-coupled laser provides the light source which is pumped into a SiN/silica waveguide.
  2. A process known as four-wave mixing in a ring resonator generates pairs of single photons whose frequencies are symmetrically offset around half the pump frequency. The ring cavity enhances the process by trapping and recirculating the pump light.
  3. The photon pair is separated from the residual pump photons with directional couplers. One photon is kept in a buffer loop while the other goes to a detector (so they know there’s one in the buffer loop).
  4. When the detector fires, a modulator releases the buffered photon into the guts of the machine.
  5. The guts of the machine are a collection of modulators and Mach–Zehnder interferometers that impose superpositions and control phase—the quantum logic operations. In fact, they apply unitary transformations equivalent to gate operations.
  6. Then entanglement is produced by forcing two photons to interfere at a 50:50 beam splitter. When the photons become entangled, their measurement outcomes start depending on each other. That pattern of dependencies is critical to running the chosen algorithm.
  7. The entangled photons are sent to another Mach–Zehnder interferometer or beam splitter, and each output path is sent to a single-photon detector where it is measured as a 0 or 1.
  8. The electronics watch each detector, update the measurement plan in nanoseconds, and feed new settings back into the modulators for the next batch of photons. Quantum computation works by building up statistical results from millions or billions of photon detection events.
  9. The photon sources and detectors operate in parallel—thousands of identical generation paths all firing at once to keep the pipeline full. The optical logic inside the chip runs in series, where photons pass sequentially through multiple interferometers and beam splitters that each apply part of the computation. The architecture is parallel in production and detection but serial in how each photon’s state is processed and fused into the growing entangled network.

It’s the mousetrap to end all mousetraps. Here is a list of all the pain points that I see:

They need less than 3 dB total optical loss per photon journey to scale. That’s ❤ dB across a single photon’s full journey: multiple waveguide crossings, layer transfers, fibers, tens of interferometers, couplers, and splitters, plus centimeters of waveguide attenuation—all in a SiN core with silica cladding on a silicon substrate.

A bragg grating in SiN isn’t as simple as laser writing as for Ge-doped silica – it requires lithography and etch. Ouch.

Because the bonded BTO modulators, single-photon detectors, and drive electronics must operate at cryogenic temperatures, much of the surrounding optics must be cooled too. That introduces new issues: thermal contraction, refractive index drift, and local heating from active components. And you either have the optics at the same temp or suffer a really large thermal profile, which won’t work. If you instead cool the whole assembly to cryogenic temperature, you at least get a uniform baseline, but now everything else (fiber arrays, wire bonds, adhesives) has to survive 2-4 K without delamination or cracking, and the cryostat must sink the static load from heaters and electronics.

What amazes me is that the electronics keep up with the photons in real time. They have to, or none of it works.

To get entangled photons, they must arrive at the same beam splitter within roughly a picosecond (not femtoseconds). You can’t just pile up a million photons in one spot, so the system builds small entangled groups (pairs, triplets, clusters) and then fuses those clusters step by step using more beam splitters and detectors. Each successful “fusion” extends the entangled network.

So a million entangled photons means billions of precisely timed events across thousands of optical paths, all synchronized to a single laser clock. The photons are never physically co-located—they’re coordinated in time and interference space. That’s why “scaling” really means running an enormous optical assembly line of microscopic interference events that somehow all stay phase-aligned.

Every photon must be identical (same wavelength, polarization, and temporal profile) to interfere correctly. Any drift in the pump laser, resonator temperature, or fabrication tolerance breaks interference visibility and kills entanglement fidelity.

All interferometers must stay phase-locked to within a small fraction of a wavelength. Thermal noise, mechanical vibration, or electrical cross-talk changes phase by milliradians and destroys coherence.

Every photon emission, detection, and fusion event must align with a global clock at gigahertz rates. Even tens of femtoseconds of jitter between optical and electronic domains can ruin the timing. Phase drift and timing jitter are not simply noise; they accumulate irreversibly over many fusion steps.

Ring resonators must have identical resonance frequencies. SiN process variation shifts refractive index or waveguide width by a few nanometers, detuning the cavity.

Superconducting nanowire detectors (SNSPDs) have nonzero dark counts and recovery times on the order of tens of nanoseconds. With millions of detectors firing billions of times per second, false triggers and dead time become a major issue.

The electronics must read, process, and respond to billions of detector events per second while keeping latency below one nanosecond, an aggregate data bandwidth on the order of terabits per second, distributed across cryogenic and warm tiers.

Each chip layer (SiN core, BTO modulators, detector layer, fiber array) has its own yield curve. When you hybrid-bond them, the effective yield is the product of all those probabilities. To reach a meaningful qubit count, they’ll also need to tile multiple chips and link them optically through inter-chip waveguides or fiber vias, adding yet another layer of alignment, coupling loss, and yield risk. Even within a single chip, multiple stacked waveguide layers are required to achieve sufficient device density, introducing lossy vertical vias and additional fabrication complexity.

Cryogenic systems at 2-4 K have limited cooling power, typically a few milliwatts per stage. Active modulators and readout lines dump more heat than that, forcing delicate balance. The cooling system also introduces vibration that destabilizes the optics.

Even if the chip works, coupling light efficiently in and out, especially at cryogenic temperatures, requires submicron alignment between hundreds or thousands of fibers or waveguide facets. Thermal cycling shifts everything.

Every MZI, modulator, and detector must be calibrated and actively tuned. That’s millions of control voltages and feedback loops, all running in parallel. Managing calibration drift alone is its own full-time problem.

And finally, production yield: every fabrication error concatenates because no virtually errors in production can be fixed. There are thousands of individual fabrication steps involved. A rough estimate is that they will need better than 99.95% yield per process step to get even one working unit. Decades of experience in photonics manufacturing to date do not support that target being achieved.

In PsiQuantum’s photonic system, the number of process steps scales superlinearly with qubit count because every additional logical qubit requires proportionally more photon sources, interferometers, modulators, detectors, calibration channels, and synchronization events, each introducing multiple fabrication and operational layers. Doubling the qubit count roughly squares the number of required fusion gates and optical interconnects, since each new photon must interfere with several others to form a larger entangled cluster. This means fabrication complexity, calibration cycles, and control feedback loops all grow faster than linearly, while total yield drops multiplicatively as each layer, bond, and alignment adds new failure probabilities. At large scales, the process step count behaves closer to quadratic in qubits, with exponential penalties when including fault-tolerant redundancy.

What that means is that each order-of-magnitude increase in qubit count tightens the per-step yield tolerance by roughly two orders of magnitude. So there is your real challenge here even if you crack all the issues described above – it doesn’t scale.

PsiQuantum’s architecture is probabilistic from end to end. Every source, switch, and fusion gate only works part of the time, and photons that fail are thrown away. Even if fabrication were perfect, loss would still compound through every layer of routing and interference. Their claim of tolerating ten percent loss per resource generator is overly generous—at that point, the surface code’s error margin is already spent, leaving nothing for the rest of the optical chain.

Each photon must survive dozens of components: filters, couplers, modulators, and detectors. PsiQuantum’s own numbers show 30–50 percent loss just to produce one heralded photon, long before it’s entangled or fused. That means per-component losses must be far below their public 3 dB target, or the whole model collapses statistically. The limitation isn’t quantum mechanics; it’s probability stacking up faster than photons can survive it.