Future of Compute

Many important optimization problems require finding the best solution among an enormous number of possibilities. For conventional computers, the computational effort can grow rapidly with problem size, motivating the search for alternative computing concepts in which physical systems themselves perform part of the computation. In this talk, I will discuss our work on coherent optical networks as a platform for such physical optimization. The central idea is to encode the variables of an optimization problem in the phases of coherent optical fields and to engineer their interactions such that the gain function of the optical network represents the cost function to be minimized: configurations with high optical gain correspond to low values of the cost function. The resulting system is closely related to models of interacting spins and, for sufficiently complex interactions, to spin glasses – systems characterized by frustration and rugged optimization landscapes with many competing solutions. We experimentally realize such networks using photon Bose–Einstein condensates in optical microcavities. By coupling several condensates and controlling their interactions, we create programmable optical spin systems whose collective dynamics favors configurations of high optical gain and thereby searches for low-cost solutions to the encoded optimization problem. I will introduce the basic operating principle, present our experimental results on optical spin-glass simulation, and discuss the opportunities and challenges involved in scaling this approach towards larger programmable networks and dedicated optical processors for hard optimization problems.

The energy cost of digital neural networks is growing unsustainably. These networks rely mainly on large matrix-vector multiplications, followed by simple, fixed nonlinear activation functions. An alternative is to let device physics perform complex nonlinear operations directly. Reconfigurable nonlinear-processing units (RNPUs) are nanoscale, multi-terminal silicon devices that apply a nonlinear operation to a set of input voltages, where the nonlinearity can be tuned through control voltages. Using differentiable surrogate models, we simulate large networks of these devices and train their responses in software. These tunable nonlinear elements are a natural fit for analogue Kolmogorov-Arnold Networks (KANs), a recent type of network that gains its expressive power from learnable nonlinear functions rather than matrix-vector multiplications. System-level estimates indicate that RNPU-based KANs need orders of magnitude less energy, and far less chip area, than digital networks of comparable accuracy. Multi-terminal devices also present new opportunities for training. Whereas conventional networks typically use univariate nonlinear activation functions, RNPUs apply a multivariate nonlinearity. We show that which signal is connected to which terminal strongly affects the functions a device can express. By letting the training algorithm learn this wiring, we obtain more expressive RNPU-based models with negligible additional hardware or operating cost. Altogether, physics-based nonlinear processing can lead to more compact, energy-efficient AI hardware, but requires rethinking both network models and the algorithms that train them.

“Second-order topological insulating (SOTI) states in three-dimensional materials host helical one-dimensional hinge states that can support unconventional superconductivity. Here, we provide evidence for topologically protected hinge states in BiSb(3%) nanoflakes through an unconventional magnetic-field interference pattern and demonstrate a 4π-periodic supercurrent, evidenced by the suppression of the first and third Shapiro steps. These results highlight the potential of SOTI hinge states as a platform for Majorana bound states, which are of interest for topological quantum computing due to their non-Abelian nature and potential robustness against local perturbations. Moreover, the spatially separated hinge channels could enable superconducting interferometer architectures for manipulating and processing quantum information, positioning BiSb as a promising platform for future quantum technologies.

Ref: Coexisting Topological Hinges and 1D Rashba States in Bi0.97Sb0.03 Revealed by the Josephson Effect, ACS Nano 20 (31), 21636–21645 (2026).”

Holes in silicon are an attractive candidate for spin qubits applications. The p-type symmetry results in a low hyperfine constant, reducing environmental nuclear spin noise [1]. One particularly promising candidate is an acceptor based qubit, where the hole bound to an acceptor atom in silicon is used for qubit encoding [2]. These acceptor atoms don’t break the crystal symmetry in the silicon lattice and therefore have a 4-fold degeneracy at the Γ-point, which needs to be lifted in order to create a well defined qubit basis. In order to do this, we use gate induced strain to induce heavy hole-light hole splitting and create either a heavy hole or light hole based acceptor qubit [3]. To apply gate induced strain, we use the difference in thermal expansion coefficient between the Si substrate and Al metal gates. The qubits need to be operated at cryogenic temperatures allowing the Al gate structure to induce strain in the silicon substrate and thus the Acceptor atoms, which are placed using ion implantation and incorporated in the Si substrate. The acceptor states are read out using a metal oxide semiconductor(MOS) based single hole transistor(SHT), used as a local charge sensor. Tuning the SHT allows it to remove the weakly bounded hole from the acceptor, bringing it to its neutral state, and read out the spin state using elzerman readout.

[1] D. V. Bulaev and D. Loss, Electric dipole spin resonance for heavy holes in quantum dots, Physical Review Letters 98, 097202 (2007)

[2] J. Salfi, J. A. Mol, D. Culcer, and S. Rogge, Quantum computing with acceptor spins in silicon, Nanotechnology 27, 244001 (2015)

[3] D. van der Bovenkamp, C. S. A. Müller, B. D. Pantiru, I. Bošnjak, M. Cignoni, Q. Torrent Nicolau, M. E. Bal, S. Wiedmann, J. Ridderbos, and F. A. Zwanenburg, Gate induced strain on a two-dimensional hole gas in silicon, arXiv:2607.07932 (2026).