Quantum Photonics Moves To A 300 mm Line

Nought point nought eight five decibels. That is the average loss Xanadu measured across a single chip facet in June, where a fibre meets a photonic die, and it is the number that explains everything the company announced on 6 October. A decibel is a ratio, and 0.085 dB is a transmission of about 98 per cent — two photons in a hundred lost at one interface. In almost any other branch of photonics that would be an unremarkable figure, the sort of thing a datacom engineer would not bother putting in a press release. In photonic quantum computing it was worth a press release, because the photon that does not arrive is not noise. It is the calculation.

On 6 October Xanadu Quantum Technologies (NASDAQ/TSX: XNDU) and GlobalFoundries (Nasdaq: GFS) announced a multi-year strategic partnership to move Xanadu's hardware onto GF's 300 mm production line in Malta, New York. The initial scope is narrow and deliberately so: industrialising Xanadu's superconducting nanowire single-photon detectors and its ultra-low-loss silicon nitride photonics platform. Not the full machine. Two components.

Silicon nitride is the waveguide. Xanadu's own description of its role is the useful one — the material serves as the optical circuit board, guiding photons across a broad spectrum of light with what the company calls negligible signal loss. Silicon nitride earns that job by what it is not: unlike silicon, its bandgap is wide enough that it does not absorb at the wavelengths of interest, and unlike silica it confines light tightly enough to make a chip-scale circuit. What is left is scattering from the sidewalls, which is a lithography and etch problem, which is to say a fab problem.

The detector is the harder half. Reading the output of a photonic quantum computer is not a matter of asking whether light arrived; it is a matter of asking how many photons arrived, one, two, three or four, in a single pulse. That is photon number resolution, and Xanadu has settled on superconducting nanowire single-photon detectors as its platform for it, the release calling SNSPDs widely recognised as the most reliable PNR technology. The company expects the choice to increase system speed significantly and to reduce substantially the chip overhead a utility-scale machine needs.

In a conventional optical link, loss is paid for with power. Launch more light and the receiver still sees enough. A photonic quantum computer cannot do this, because the quantum states it computes with are made of specific, countable photons in specific modes, and a lost photon is an erasure in the middle of the computation. Loss therefore does not degrade the signal-to-noise ratio; it degrades the answer, and the degradation compounds with every component the state passes through.

The consequence is a hardware budget that looks nothing like a telecom one. A state that traverses twenty components each transmitting 98 per cent arrives with about two-thirds of its probability intact. Error correction can repair the rest, but the overhead — the number of physical modes needed per logical one — climbs steeply as loss rises, and there is a threshold below which it does not converge at all. Minimising loss, Christian Weedbrook said in June, "is paramount to unlocking the full potential of photonic quantum computing". Which is why that month's benchmark was framed as a packaging achievement rather than a design one.

That work was done with Corning on customised fibre and fibre-array solutions and DISCO on wafer singulation — two names that tell you where the losses were hiding. Not in the waveguide. In the dicing of the facet and the alignment of the fibre to it.

Figure 1. A packaged Xanadu photonic chip with its fibre arrays attached. Fibre-to-chip coupling is the interface where the company measured 0.085 dB average loss per facet in June 2026. Credit: Xanadu Quantum Technologies.

Xanadu describes itself as building fault-tolerant quantum computers using light, designed to compute at room temperature, and that is accurate about the photonic circuit. It is not accurate about the detector, and the partnership makes the gap visible. A superconducting nanowire is superconducting only when it is cold — these devices are conventionally operated in the single-kelvin range or below. So a system built on SiN waveguides and SNSPD readout is a room-temperature optical machine with a cryostat bolted to its output.

This is not a contradiction so much as a division of labour, and it is arguably the photonic approach's advantage: only the detectors need cooling, not the whole processor, which is the opposite of the superconducting-qubit arrangement where everything lives in the dilution refrigerator. But it does mean the 300 mm line now has to make two very different things well, one of them a niobium-based superconducting film patterned into nanowires a few tens of nanometres wide, on the same wafer flow that produces the photonics.

Malta is not a research fab. GF's figures for the site are a 300 mm facility with capacity of 400,000 wafers a year on a 300 mm-equivalent basis, around 500,000 square feet, about 2,500 employees, and an existing silicon photonics offering alongside FDX FD-SOI, FinFET and RF-SOI. GF announced a dedicated quantum business, Quantum Technology Solutions, from Malta on 21 May 2026, alongside a letter of intent from the US Department of Commerce that was finalised in September 2026 as a definitive agreement worth up to $375 million.

What that buys a company like Xanadu is not lower unit cost — volumes are nowhere near the point where that matters. It is statistical process control. A research line produces a best device; a production line produces a distribution, with a known mean and a known spread, measured on every wafer and corrected when it drifts. For a machine that needs thousands of nominally identical low-loss components to behave the same way, the width of that distribution is the specification that decides whether the architecture works.

Xanadu has been assembling a supply chain in public for most of a year — Tower Semiconductor in February, EV Group, ASML, DISCO, Corning — and each agreement covered one process step. The GlobalFoundries deal is different in kind because it names a fab, a wafer size and a product scope, and because the stated output is the manufacturing foundation for Xanadu's forthcoming fault-tolerant demonstrators rather than a process study.

Neither company has published a target loss figure, a detector efficiency, a wafer count or a date, so the announcement cannot yet be scored. What it does settle is where the industry thinks the bottleneck sits. Not in algorithms, not in architecture, not in the physics of squeezed light — in whether a commercial semiconductor line can hold a waveguide sidewall and a superconducting nanowire to the same tolerance, wafer after wafer. That is a question foundries have answered before, for transistors. It has not been asked of photons that have to survive.