Searching for X-Lab-Shaped Bottlenecks in Quantum Interconnect and Integrated Photonics
What’s already being built, what we’re unsure about, and where the FRO-shaped gaps might actually be
This is a preliminary post on the NSF X-Labs Quantum Interconnect and Integrated Photonics topic. We’re writing this as a provocation directed towards expert researchers, PMs and others in this field - we’d like your help to continue this very early investigation or spin up your own. In particular, we want to get this in the hands of people who can find the FRO-shaped bottlenecks here (and especially those who want to start new organizations to work on them), and start a conversation with them to get direct feedback. As always, if you are looking at proposing an X-Lab, please reach out via our short form and stay tuned for a webinar we’ll run on this.
—
Quantum Uncertainty
In our last post we walked through some select R&D gaps we see under the Sensing & Imaging X-Labs topic. The second topic — Quantum Systems: Interconnects and Integrated Photonics — is in some ways the harder landscape to read.
To date, there have been real companies in this space shipping products, two newly-announced billion-dollar US foundry plays, several hundred million dollars of VC behind photonic quantum computing efforts, an ecosystem of NSF and DOE centers that have been working in this space, and this recent related program from DARPA.
So the obvious question is: what’s left for an X-Lab? Where is a $10M-$50M/year, milestone-driven, independent team actually additive rather than duplicative?
We think there are real answers, but we want to remain cautious and curious about them to get to the genuinely interesting ones. What follows is our best current read based on a small number of early conversations.
(Again, we’re publishing this mostly to get pushback from researchers, engineers, entrepreneurs, VCs and PMs / other leaders in the field who know much more than we do - please comment below or reach out with your thoughts!)
The Shape of the Problem
Quantum computers don’t yet talk to each other very well. Inside a single dilution refrigerator, a superconducting processor might be a few hundred to a few thousand qubits today, connected to room-temperature control electronics by over a mile of cryogenic flex wiring running through the temperature stages. IBM, Fermilab, and others are betting that within a shared cold environment — multiple chips in one fridge, multiple fridges in one cluster — direct microwave links can scale to perhaps tens or even hundreds of thousands of qubits without conversion of microwave excitations to optical. But that path stops at the edge of the cold envelope. To scale beyond a single cold cluster — between buildings, across a campus, or across a country — microwave excitations have to be coherently converted into optical photons that can travel through fiber (or perhaps free space) at room temperature. Optical frequencies are about 40,000× higher than microwave, which means a telecom fiber at room temperature is, for quantum purposes, far quieter than a microwave cable at millikelvin. Once you’ve paid the transduction cost at each end, the photon can travel through ordinary fiber for kilometers without thermal contamination.
Other qubit modalities (neutral atoms, trapped ions, NV and SiV diamond color centers) sit closer to optical wavelengths natively but still need their own bridges to the telecommunications infrastructure we use to communicate. (PsiQuantum and some others cleverly use photons themselves as the qubits, which is a different story!)
The NSF solicitation lists, as examples of unmet R&D challenges, a number of compelling desiderata on this front: scalable modular architectures, heterogeneous quantum system interconnects, integrated photonic devices, transducers, optical interfaces, and components including entangled sources, interferometers, filters, switches, and detectors.
The Current State
We worked to put together this partial map of what’s already being built, because the additionality (how can an X-Lab be complementary, counterfactual, useful) question matters so much to us here:
QphoX (Delft) is shipping the first commercial transducer products, partnered with Rigetti and AFRL on superconducting optical readout, and recently selected for a €50M European pilot line (see here). The product volumes are small and the customer base is mostly research labs, but they’re developing real hardware deployed in real systems.
memQ, a UChicago spinout, just closed a $10M Series A and is building qubit-agnostic interconnects on a commercial silicon photonics foundry process. These are erbium-doped titanium dioxide thin films on silicon nitride, taped out via AIM Photonics.
There is a lot of other private sector activity. Trapped ion quantum computing company IonQ has acquired Lightsync to push on quantum photonics (and the foundry SkyWater!), Qunnect is building quantum networks and Welinq is building quantum memories, among others. Nu Quantum raised a $60M Series A in December 2025 (the largest round ever for a pure-play quantum networking company) to build datacenter-form-factor Quantum Networking Units, and co-founded the Quantum Datacenter Alliance with Cisco, NTT DATA, QphoX, Quantinuum, QuEra, and OQC to work on cross-modality interoperability. Misha Lukin and Hannes Bernien also created a second photonic interconnect company recently called Cavilinq and it appears to be commercializing the cavity-enhanced photonic links used in the Knaut demonstration we discuss below. Qunett (an MIT/UMaryland spinout from Englund and Guha — yes, distinct from Qunnect) is building entanglement sources, diamond quantum memory modules, and interconnects.
The Center for Quantum Networks (NSF ERC) runs BARQNET and QUCSON, with many of the senior US quantum networking PIs as faculty.
The “transducers” from quantum information in qubits to transmitted light are clearly key components in all this. The most recent transducer benchmark we took note of is Faraon’s group at Caltech: rare-earth ions achieving percent-level transduction efficiency at the single-photon level with 1.24 photons of added noise, and demonstrating interference between two transducers on a chip. That is a step forward, but we want to be careful: percent-level efficiency is not unity, 1.24 added photons is not shot-noise-limited, and the rare-earth route is one of at least three competing approaches — electro-optic, piezo-optomechanical, and rare-earth — each with different tradeoffs. And the microwave-to-optical conversion problem is not the only conversion in the stack: optical-to-optical quantum frequency conversion, which bridges the wavelengths many quantum memories actually emit at (often visible or near-IR) and the telecom band that fibers actually carry, is its own engineering layer with its own efficiency and noise budget.
Optical readout of superconducting qubits is also now a demonstrated capability. QphoX, Rigetti, and Qblox jointly demonstrated 81% single-shot readout fidelity of a superconducting qubit through QphoX’s piezo-optomechanical transducer — evidence that transducers are now interfacing with deployed superconducting qubits. Within the cold envelope, Cleland’s group at UChicago previously demonstrated deterministic multi-qubit entanglement through a one-meter superconducting cable connecting two nodes with 91.1% state-transfer process fidelity — an empirical foundation for the IBM/Fermilab cold-link scaling bet we mentioned earlier.
And, last week, two foundry plays made the news: namely, IBM’s Anderon ($1B proposed CHIPS award, 300 mm purpose-built quantum wafer foundry) and GlobalFoundries Quantum Technology Solutions ($375M proposed CHIPS award). This reflects a parallel push to onshore the quantum materials and packaging supply chains, much of which currently has single-supplier or offshore vulnerabilities.
So the question is: what’s the crucial platform technology that none of these will produce on their own?
A potential framing for X- Labs (that we’d like feedback on)
Knaut and colleagues at Harvard (Nature 2024) entangled two silicon-vacancy nodes in separate cryostats through 35 km of deployed Boston telecom fiber. The whole stack worked, but the entanglement generation rate was… about 1 Hz, orders of magnitude too slow for a fully fledged quantum network.
That number is perhaps an honest summary of where the field is. Every individual layer of the quantum interconnect stack has had record-setting demonstrations recently: 97.5% indistinguishable single photons from a single source, 80% efficient quantum memories at telecom wavelengths, 30%-class photon collection from color centers, 2 ps clock synchronization between national labs, percent-level transduction with near-shot-noise added noise. Each metric needs improvement, there need to be unified architectures to combine them (such as what DARPA’s HARQ program is working on), and then composing them into a single system becomes the bottleneck.
Excellent component-level science exists. The components were each optimized in isolation. Multiplying their point efficiencies across the stack gives you 1 Hz. Closing this rate gap by many orders of magnitude is not any one component’s job — it is a system problem that requires every layer to improve together, in concert, with shared characterization, shared protocols, and design budgets that account for the other layers’ presence. This type of systems problem may be good for an X-Lab.
There is also a potential approach around “tooling”. In the first half of the 20th century, you could pick a house out of a Sears catalogue for as little as $13,000 of today’s dollars, and receive all the pieces you need, and could assemble it yourself. Less than 20 years ago, if you wanted a dilution refrigerator, you had to design and build it for your space. It’s not hard to imagine a world where Bluefors and other fridge companies didn’t exist and it took an additional decade or two for a company to sell push-button dilution refrigerators. More optimistically, one could also imagine a world where each of the most time-consuming engineering steps of building quantum interconnects and photonic chips could not only be treated as solved problems, but as common and widespread methods and practices, provided for all as detailed recipes, low-cost services, or ready-to-assemble kits, like the now-coveted Sears Homes.
Some specific areas where work may be needed
We are not certain about any of these, and we’re simply not experts in these fields. We’re putting these down to get input and in the hopes you’ll suggest better ones.
Deterministic, indistinguishable single-photon sources at scale. Individual quantum-dot sources now achieve 97.5% indistinguishability with 20.8% fibre-coupled brightness, stable over 10 hours, and there are commercial products from Quandela, Sparrow Quantum, and AegiQ. Interference between remote sources has been demonstrated across pairs of QD-cavity emitters. So the gap is not “can a single source emit a good photon” — that’s largely solved. The gap is N-source indistinguishability at the scales quantum networks and photonic computing actually need: hundreds to thousands of sources on a wafer with the spectral, temporal, and polarization homogeneity to interfere with each other. Quantum dots are intrinsically inhomogeneous and the strain and electrical tuning that work for two-source matching do not yet scale. The alternative — massive multiplexing of heralded spontaneous parametric down-conversion sources — replaces the emitter inhomogeneity problem with a lossy-switching-network problem (PsiQuantum’s Omega chipset is the first attempt to do this at production-fab scale). Either route, the bottleneck is no longer the single source, it is the manufacturing and the multiplexing.
Photon collection efficiency from solid-state emitters. Sources might be good, but getting them coupled to the mode you care about can still be hard. A typical NV center radiates omnidirectionally. Only a few percent goes into the zero-phonon line, of which only a fraction can be collected by high-NA optics, of which fewer still couples into single-mode fiber — end-to-end collection efficiencies for unstructured emitters are often below 0.1%. Photonic crystal cavities and other methods have pushed individual lab demonstrations into the tens-of-percent range. None of this is yet manufacturable at scale or integrated with downstream optics in a way that survives packaging and cryogenic cycling.
Detectors that compose with the rest of the chip. Single-photon detection at the wavelengths quantum networks use is mature. The remaining detector questions are mostly about architectural fit. Superconducting nanowire single photon detectors have generally required dedicated substrates and hybrid integration with photonic chips. Most latch above ~100 MHz click rate. Kilopixel arrays exist as demonstrations but megapixel arrays don’t. Photon-number resolution and ultra-low jitter and high count rates and large arrays are hard to get on the same device. Graphene/hBN-based platforms are potentially interesting because of the architectural fit for a quantum interconnect chip where the detector is part of the same fab process as the routing and the readout electronics. And for the longer term, higher temperature or (moonshot) room temperature single photon detectors would be a big unlock. (PsiQuantum’s Omega integrates SNSPDs directly onto silicon photonic chips at GlobalFoundries, which is one example of monolithic integration.)
Interconnect and System Architectures. This is basically an X-Lab version of HARQ, e.g., to create an interconnected theoretical architecture that offers 1000X improvement on some meaningful application. That could be a quantum internet or it could be just a modular quantum computer running an application. Where are the strongest benefits to be had through heterogenous versus homogenous integration – which modalities are the most complementary versus redundant, creating greatest value in integrating them? When to pick the computing modalities best suited to the available transducer technologies or vice versa? What are acceptable error rates and other specifications? QEC requires 1% error rates in theory and closer to 0.1% in practice. Are those the right benchmarks for networking? If not, what are they? And what is an acceptable transmission rate for useful systems?
Foundry-compatible quantum-grade photonics — and characterizing the ceiling. Quantum interconnects need loss budgets, phase stability, polarization control, and isolation from thermal and acoustic noise that classical photonics doesn’t. Academic labs have shown extraordinary numbers: silicon nitride waveguides at 1 dB/m loss integrated with single quantum emitters at 930 nm, anneal-free SiN at 1.77 dB/m on a CMOS-compatible process, and resonator Q factors in the tens of millions. So the physics is well demonstrated. What does not yet exist as a foundry process is the consistent achievement of those quantum-grade specs in a production environment, characterized statistically across wafers, with design rules anyone can use. The new foundries (Anderon, GF) will help. Running those measurement campaigns and publishing the design rules could be a public-good contribution.
Connecting the pieces, not just running them. Each platform has shown it can do the basics of a quantum network on its own. Delft has built a working stack on diamond NV nodes, superconducting groups have improved entanglement quality to >94% fidelity on short links, and there is even a draft architecture standard being written. What does not yet exist is hardware from different platforms talking to each other — a Delft NV setup does not speak to a Harvard SiV setup does not speak to a trapped-ion setup. The supporting infrastructure has the same problem: Caltech and partners demonstrated 5-picosecond clock synchronization between Fermilab and Argonne over deployed fiber in 2022, but a productized version that ships timing alongside quantum signals across many nodes of a real network does not exist. And the field does not yet have a shared way to measure how good a quantum link is — no community benchmark analogous to “Tbps” in classical networking.
First-generation quantum repeater primitives — entanglement swapping at relays, memory-enhanced scaling, multi-node coordination — have been demonstrated in academic settings, and groups at Delft, Stony Brook, Hefei, and elsewhere are building toward chains that compose them. But no one has yet built a working quantum repeater chain that distributes entanglement at useful rates beyond what’s possible with direct fiber transmission. Getting from there to second-generation quantum repeaters, with error correction built into the links themselves, depends on closing all of these gaps together.
Use Cases. There are use cases for quantum networks but they haven’t been developed to be made practical in most cases. There is lots of opportunity here, e.g., can we make a useful version of quantum non-local games or can we set up versions of quantum money or one-time programs that are useful new cryptographic primitives. And how much is quantum-quantum interconnect versus quantum-classical interconnect the real bottleneck for scaling algorithms? An X-Lab that made the use cases and technical performance targets clearer would help unlock the field.
End-to-end entanglement generation rate. The empirical answer to “how good is the whole stack” is the rate at which two remote nodes can be entangled. The current state of the art is striking. Knaut and colleagues demonstrated entanglement between two SiV-based nodes in separate cryostats, in separate Harvard labs, through 35 km of deployed Boston-area urban telecom fiber. The whole stack works. The success rate is approximately 1 Hz. A useful quantum network at scale needs entanglement generation rates many orders of magnitude higher than that. (Also, from Qunnect, “The collaboration achieved record swapping rates of 1.7M+ pairs/hour locally and 5,400 pairs per hour over deployed fiber.”) Closing this rate gap is partly about improving each layer independently — better photon collection, lower-loss optics, faster spin readout — and partly about multiplexing many channels in parallel, which requires multi-mode quantum memories, frequency- or time-multiplexed sources, and switching infrastructure. This is the integration problem mentioned above.
The quantum memory itself. We have spent this whole post talking about how to connect quantum systems and have not said much about the thing being connected. Quantum memories store quantum information at network nodes; their coherence time, multi-mode storage capacity, and write/read efficiency together set what a network can actually do. The progress has been remarkable. Integrated rare-earth-doped crystal memories just demonstrated 80% storage efficiency for weak coherent pulses, 70% for telecom-heralded single photons, and 20-mode storage at >50% efficiency. Welinq is shipping commercial neutral-atom memories in standard 19-inch racks at room temperature, with >90% storage-and-retrieval efficiency and 200 µs coherence — their QDrive system was sold to the Slovak Academy of Sciences in January 2026. The 29Si nuclear spin in the Knaut demonstration achieves second-scale coherence. So the individual records are excellent. What does not exist is a platform that simultaneously combines long coherence (seconds or better), many storage modes, high efficiency, telecom-compatible interfacing, room-temperature or modest-cryogenic operation, and a path to manufacturable packaging — and is qualified statistically across many devices rather than one heroic demonstration.
Moonshot: A full quantum random access memory. QRAM is a fundamental primitive for many quantum algorithms and yet it does not exist as hardware. Bucket-brigade photonic spin-network architectures have been proposed, but it has been argued that bucket-brigade QRAM may not survive realistic noise without fault-tolerant gates, which would push it well beyond current proposals. We are genuinely unsure whether this is X-Lab-shaped at this point. But it would force integration.
What we want to hear
If you work in this space, we have specific questions:
Which bottlenecks above feel real to you, and which feel already closed, or otherwise missing the mark?
What did we miss? In particular: what *capability* does a working quantum interconnect ecosystem actually need that no organization has an ability and mandate to execute on today?
Where would a $10M-$50M/year, milestone-driven, full-time engineering team move the needle — and where would it just duplicate efforts?
Is QRAM a good systems moonshot? Is there a better one? Is it too early?
What engineering step or technique are you or your team surprisingly good at that could, if you had sufficient capital, unblock quantum science the way BlueFors did by scaling access to dilution refrigerators?
Integrated photonics more generally
We’ve talked a lot about quantum “interconnect” per se in this post. But there’s a lot of “integrated photonics” that’s useful for computing and other areas, beyond quantum interconnect per se, e.g. PsiQuantum/Xanadu/Photonic Inc‘s work and new companies like MothQuantum. Integrated photonics is also central inside quantum computers themselves, not just between them. Neutral-atom and trapped-ion systems do their qubit trapping, cooling, and control through laser beams that today come from bench-top optics — and the path to thousands or tens of thousands of qubits runs through replacing those bulk-optical setups with photonic integrated circuits that can deliver shaped beams to individual qubits from a chip. The same integration challenges (low-loss multi-wavelength operation, dynamic addressing, foundry-process scale) apply. Is there a platform opportunity here that individual companies aren’t set up to pursue?
Seizing the opportunity
If you are looking at proposing an X-Lab, please reach out via our short form.
We will be holding a webinar soon for people planning to propose FROs and wanting to interact with Convergent around that.
NSF is also doing an official webinar.
Initial X-Labs applications to NSF are due in mid-July.
Thanks to Jay Lewis, WIll Zeng and Evan Miyazono for comments on drafts and Isabelle Phinney, Nathalie De Leon, Christophe Jurczak and others for preliminary conversations that informed our early learnings.
Thinking about applying for an NSF X Lab this summer?
Join us for a practical 30-minute webinar covering
Background on the FRO model, its relation to X-Labs, and how Convergent Research has incubated FROs in the past
Our perspective on what you need to submit an application in July
Our perspective on some common applicant questions
Some advice on what to do if you don’t yet have an entity / host organization to apply under for Phase 0 in July.
In short, most teams will likely need to get an entity with a SAM UEI that they can submit under (which can take several weeks to set up) – there are limitations on how many initial applications Convergent or any other single org could submit as a lead entity for in July.
How to request feedback on your proposal draft
Date: June 5, 2026
Time: 2:00–2:30 PM ET
Please register your interest here (whether for the webinar or more broadly).
You skipped Nu Quantum
Super interesting but my subject matter is along the second x-labs RFP, will you do an analogous dive there or go through your contact form?