The most honest sentence in the announcement is the one that sounds least like an announcement. “We need to understand what these hybrid systems can fundamentally achieve before we can design and optimize them for real-world applications,” said Francesco Monticone, associate professor of electrical and computer engineering at Cornell, on a $7.5 million U.S. Department of Defense Multidisciplinary University Research Initiative (MURI) award his team will lead. The project, “Deep Nanophotonic Inference Systems,” was announced by the Cornell Chronicle on October 15, 2026. It funds a question, not a product: what can nanophotonic systems compute, in principle, before electronics has to take over?
That framing matters, because this desk has spent the month covering photonic systems that already claim results — a backpropagation-free training algorithm from Shanghai, a holographic multi-task classifier from CICT — each resting on assumptions about what the optics can do. This MURI goes underneath them, asking whether the field actually knows the answer to the prior question, with a multi-institution team, three-dimensional nanophotonic structures, and a program organized around limits first, devices second, and end-to-end systems third.
The problem the money is aimed at
Turning light captured by cameras and other sensors into useful information — recognizing a pedestrian, tracking a vehicle, flagging a feature in a medical image — is a task increasingly performed by AI, and it relies almost entirely on digital electronics. A sensor converts photons to electrons, and from there the computation is electronic: matrix multiplications on processors that draw power, generate heat, and demand memory traffic. The Chronicle states the premise plainly: hybrid optical-electronic systems could reduce the energy and computational resources required for AI-driven image analysis and other visual inference tasks, by letting some of the computation happen in the light itself, before conversion.
The idea is not new, and neither is the gap between the idea and a deployable system. Diffractive and nanophotonic systems perform linear operations — the matrix multiplications that dominate neural network inference — essentially for free as light propagates through engineered structures. But the field has been short on rigorous answers to the questions that determine whether that free computation is useful: how much can a structure of a given size and complexity actually compute, how accurately, and at what cost in fabrication tolerance, noise sensitivity, and electronic overhead? Those are the questions this project is funded to answer, and the order in which it answers them is the most interesting thing about it.
The applications named in the announcement — autonomous driving, vision-aided navigation, adaptive sensing, biomedical imaging — are worth reading as a list of places where the constraints are tightest, not as markets being addressed. Each is a setting where inference must happen close to the sensor, under power and latency budgets, often while moving. A camera on a vehicle or a drone cannot borrow a data center’s power budget. If optics can preprocess visual data at the sensor, the electronics downstream have less work to do. That is the deployment logic, and it is a conditional: if the optics can do enough, reliably enough, cheaply enough. The MURI is funded to establish the conditions, not to assume them.
Three aims, in the right order
The research is organized around three interconnected aims, and their sequence is a statement of method.
The first aim will identify the physical and information-theoretical limits of nanophotonic systems. This is the foundations work: given a volume of engineered material, a wavelength range, and a fabrication resolution, what computations are possible, and what do they cost in size, accuracy, and robustness? Limits are the least glamorous deliverable in a research program and the most consequential, because they tell every subsequent design effort where the walls are. A field that skips this step discovers the walls by collision.
The second aim will develop novel approaches for rapid simulation, inverse design, and fabrication of three-dimensional nanophotonic structures. The emphasis on three dimensions matters. Much of the diffractive computing literature works with flat, layered structures — metasurfaces stacked in sequence. Three-dimensional volumetric structures, with features at subwavelength scale throughout the volume, offer far more degrees of freedom for shaping light, and correspondingly harder design and fabrication problems. Inverse design — specifying the desired optical function and computing the structure that produces it — makes that design space searchable, but it is computationally expensive and sensitive to the gap between simulated and fabricated structures. “Rapid” is doing real work in this aim: a design method too slow to iterate cannot support the third aim.
The third aim will optimize how optical hardware and digital algorithms work together as complete end-to-end systems. This is where the project stops treating the optics and the electronics as separate problems. A nanophotonic front end that computes brilliantly but produces outputs the downstream electronics cannot use efficiently, or that demands calibration the system cannot maintain, is not a system. End-to-end optimization means the optical structure and the digital algorithm are designed jointly, with the interface between them — the sensor plane, the readout, the conversion cost — treated as part of the design rather than as someone else’s problem. As Monticone put it: “We want to understand what is fundamentally possible, use that understanding to design optimized photonic inference systems, and work with our collaborators to actually fabricate and test them.” Limits, then design, then fabrication and test. The program’s structure is its argument that the field has been building before it fully understood.
Why this team, and why now
Monticone is not arriving at this question cold. In November 2025, the Chronicle covered work by Monticone and Yandong Li, then a postdoctoral researcher, published in Nature Communications, that derived scaling laws for free-space optics and photonic circuits: how large an optical system must grow as the tasks it performs become more complex. Using optics-specific pruning methods grounded in wave physics, they found an optical system performing the same task could be 1% to 10% the size of its conventional counterpart — and, extrapolating to the linear operations of large language models at 100 billion to 2 trillion parameters, that a free-space optical setup could in principle perform those computations in a device roughly 1 centimeter thick. The same study found diminishing returns in inference accuracy as the device grows larger, suggesting the right answer is sometimes a balance of size and task performance rather than maximum size.
That earlier work is directly continuous with the MURI’s first aim. It also carries a useful piece of honesty: Monticone has said publicly that he is personally skeptical about whether optical computers will replace or drastically accelerate GPUs, while arguing that for imaging and computing in resource-limited edge scenarios, optics could work extremely well, and that space is not necessarily the bottleneck some feared. A program led by someone willing to state the skeptical case is a program whose limits work is more likely to be believed.
The team structure fits the MURI model. Co-principal investigators come from Stanford University, the University of Pennsylvania, Vanderbilt University, and the Massachusetts Institute of Technology, with international collaborators at Sorbonne University and Lund University. The announcement names institutions, not individuals, so this article assigns no names to those seats. A researcher’s CV corroborates the Army connection, identifying the award as an Army Research Office MURI from the fiscal year 2026 solicitation, co-written with MIT’s Marin Soljačić — consistent with the solicitation itself, whose Army topic 13, “Optimized Photonic Inference Systems,” describes volumetric nanophotonic structures with subwavelength features performing optical inference tasks such as image classification and vision-aided navigation, and photonic-structure/algorithm codesign. That topic description reads almost as a summary of this project’s three aims.
The funding scale fits the program’s structure. MURI awards under the fiscal year 2026 solicitation run up to $7.5 million over five years, at up to $1.5 million per year, and this award sits at that ceiling — a program-level fact, not a claim about this award’s disbursement schedule, which the announcement does not specify. It places the project in the standard MURI frame: multi-year basic research, closely managed by a Defense Department program officer, on a problem no single group could take on alone.
The honest ledger
The most important thing to say about this project is what it has not done, because the announcement is easy to misread as a result. There is no device, no benchmark accuracy, no energy-per-inference figure, no latency measurement, and no comparison against a digital baseline. The $7.5 million funds research, not a product, and the project’s own framing — foundations first — is an explicit refusal to promise the numbers that would make a deployment case. Any coverage that converts “could reduce energy and computational resources” into “will cut AI energy use” is inventing a claim the source does not make.
Second, the hybrid framing cuts both ways. A system that splits computation between optics and electronics inherits the costs of both domains: the fabrication precision and alignment sensitivity of photonics, plus the conversion, readout, and control electronics that this desk’s earlier coverage has repeatedly flagged as the unmeasured part of the energy story. The third aim — end-to-end optimization — is the project’s acknowledgment of that problem, but it is a research aim, not a solved interface. Until a complete pipeline is measured, laser to sensor to readout, against a digital baseline on the same task, the energy case for hybrid visual inference remains a direction of travel, not a result. That discipline applies here with more force than to the CICT and Shanghai papers, because here there is not even a laboratory system to measure yet.
Third, the application list should be read with care. Autonomous driving and biomedical imaging are domains where a wrong answer has consequences, and where tolerance for the noise sensitivity and fabrication variation of nanophotonic hardware is correspondingly low. The limits work must address robustness — how performance degrades with fabrication error, misalignment, and environmental drift — before any of those applications is plausible, and the announcement does not promise a robustness result. These are the motivating use cases for a foundational program, not its near-term deployments.
Fourth, a sourcing limitation to state plainly. This article is based on the Cornell Chronicle announcement, which names institutions but not individual co-investigators, gives no project duration or disbursement schedule, and reports no technical specifications for the structures to be studied. The Army topic description and the program’s funding structure are corroborating context from the public solicitation, not details of this award. Readers who want the proposal’s technical specifics will need the award documentation or the team’s forthcoming publications.
Where this fits, and what to watch
This MURI sits underneath the month’s other stories on this desk, and that is its significance. The Shanghai B-ONN paper asked how a photonic network can learn without a reciprocal backward path; the CICT holographic network asked how one fixed optical system can serve several tasks at once. Both assume the optics can carry a useful share of the computation, and both run into the same unanswered questions about capacity, robustness, and electronic-surround cost. This project is funded to answer those questions at the level of physics and information theory, before the next generation of devices is designed around answers that may be wrong.
That makes it a program to watch over years, not a story to check next week. Three developments would indicate it is producing something the field can build on.
First, publications from the limits aim: quantitative bounds on what a nanophotonic structure of a given volume and fabrication resolution can compute, with task-complexity, accuracy, and size trade-offs stated precisely enough for a device designer to use. The 2025 scaling-laws paper is the template; the MURI should extend it from free-space optics and circuits to volumetric 3D structures and to the information-theoretic side — not just how big a system must be, but how much information it can carry and transform.
Second, fabricated 3D structures from the second aim, with measured performance against the predicted limits. The gap between inverse-designed structure and fabricated device is where many nanophotonic programs stall; a fabrication result that closes it, even on a narrow task, would be more informative than another simulation.
Third, an end-to-end demonstration from the third aim: a nanophotonic front end and digital back end designed jointly, measured as a complete pipeline, conversion and readout costs included. That is the measurement that would let the energy case be stated as a number rather than a direction.
The Pentagon does not usually fund physics for its own sake, and it is tempting to read a Defense Department award as a signal that a technology is near deployment. The signal here runs the other way. The award funds the questions that must be answered before deployment is even well-defined: what can light compute, how much room does it need, and what does the partnership with electronics cost? A field that can answer those questions is a field ready to build; a field that cannot is building on hope. This project is an attempt to replace hope with bounds — and bounds, stated honestly, are the most useful thing a research program can produce.
Sources and further reading
This article is based on the Cornell Chronicle announcement: Syl Kacapyr, “$7.5M grant to establish foundations for visual computing with light,” Cornell Chronicle, published October 15, 2026 (updated October 16, 2026), https://news.cornell.edu/stories/2026/10/75m-grant-establish-foundations-visual-computing-light — the $7.5 million award, the project title “Deep Nanophotonic Inference Systems,” Francesco Monticone as lead, co-principal investigators at Stanford, the University of Pennsylvania, Vanderbilt, and MIT, international collaborators at Sorbonne University and Lund University, the three research aims, the metamaterials platforms, the second MURI on space-time metamaterials led by the University of Pennsylvania, and the application areas (autonomous driving, vision-aided navigation, adaptive sensing, biomedical imaging) are all as reported there, including both quoted passages from Monticone. No device, benchmark, energy, or latency figure appears in the announcement, and none is claimed here. Program context — the fiscal year 2026 MURI solicitation’s Army topic 13, “Optimized Photonic Inference Systems,” and the program’s funding structure of up to $7.5 million over five years at up to $1.5 million per year — is from the public solicitation (ONR announcement N0001425SF002 / ARO announcement W911NF25S0004 / AFOSR announcement NOFOAFRLAFOSR20250002); the identification of this award as an Army Research Office MURI is corroborated by a proposal collaborator’s CV. The 2025 scaling-laws work is from the Cornell Chronicle, “Researchers reveal how small optical computers could get,” November 20, 2025, reporting Monticone and Yandong Li’s Nature Communications paper: the 1% to 10% size finding, the roughly 1-centimeter-thick estimate for LLM-scale linear operations at 100 billion to 2 trillion parameters, the diminishing-returns finding, and Monticone’s stated skepticism about replacing GPUs. For the complementary device-level stories this month, see this desk’s coverage of the Shanghai B-ONN backpropagation-free training paper and the CICT holographic multi-task classifier.