The network is becoming
part of the computer.

Efficient Computer takes dataflow into volume production

Efficient Computer announced agreements for more than $97 million in Series B financing and said its Electron E1 processor was in volume production. Its spatial dataflow architecture maps operations onto interconnected processing elements. Within the fabric, operations execute when their inputs arrive, without a central program counter sequencing each step.

This changes the execution architecture while retaining digital logic and familiar programming languages. A graph of operations remains configured on the chip, reducing repeated instruction fetching and passing intermediate results directly between operations.

For connected equipment, the potential consequence is more processing directly within a sensor or controller, rather than a separate compute or cloud resource. A vibration monitor, for example, could analyse measurements locally and send a fault alert plus a short diagnostic recording, reducing routine uploads and dependence on a remote server.

Q.ANT opens its photonic programming tools

Q.ANT released an open-source software development kit for its photonic processors, with Python and C interfaces, examples and a CPU simulation backend. Developers can start on ordinary computers, with access to physical processors planned through IONOS cloud services and on-premises installations.

The toolkit exposes mathematical operations suited to the optical hardware, including nonlinear functions. For telecom researchers and product development teams, it offers a way to test candidate workloads such as signal classification or analysis of network telemetry. The announced cloud and on-premises options also create a practical choice about where the accelerator sits: a shared facility can spread its cost, while a local installation avoids repeatedly transporting large datasets to it.

AnalogAI licenses memory that also computes

South Korea's AnalogAI selected SST's memBrain SAGE intellectual property for edge processors intended to combine on-device learning and inference. The design uses analogue computation within flash-memory arrays, targeting operation at or below one watt. SST has implemented the underlying IP in 40nm and 28nm processes.

Computing conducted directly within memory reduces transfers between separate storage and processing blocks. On-device learning adds another possibility. For instance, a robot or monitoring device could adapt to local or task-specific conditions, without sending every new example to a cloud training service. This could alter both traffic volumes and application design, especially on private wireless networks. Local reactions could also continue during a connection outage, while the network carries fleet-level coordination, selected observations and model updates.

The UK sets out a neuromorphic roadmap

NeuroWare launched the UK Roadmap for Neuromorphic Computing, developed with more than 100 stakeholders. Its vision extends to 2050, covering industrial adoption, skills and infrastructure. Near-term proposals include access to testbeds, stronger benchmarking and engagement with regulators responsible for sectors such as aviation and healthcare.

For telecoms, neuromorphic computing belongs in the discussion about where sensing data is processed, as well as some other network-specific applications. A low-power event-driven processor could sit beside a radio, camera or industrial sensor, extracting useful observations before anything crosses a network. Other uses might relate to physical security or condition-monitoring of network equipment and passive infrastructure. That could distribute processing across many small devices and gateways. It also widens the context for integrated sensing and communications (for wireless or fibre networks), as well as preventative maintenance. Various useful combinations of sensing and compute will develop outside cellular equipment and 6G programmes – for instance around subsea cables or on satellites.

BrainChip makes neuromorphic sensing easier to evaluate

BrainChip launched its AKD1500 PCIe development card for standard PCs, alongside demonstrations of radar classification, human presence detection and fall detection. The applications process sensor inputs locally using its event-based Akida architecture. The card complements existing modules, chips and licensable IP.

Local sensing can turn a continual stream of measurements into a small number of events. A fall detector might send an urgent alert and supporting evidence, while keeping routine observations within the premises. Network requirements then depend on how promptly and reliably those events are delivered, with additional capacity for diagnostics and updates. This gives Wi-Fi providers, private-network operators and building-system integrators a potential role in connecting and managing distributed sensing.

Extropic adapts transformer models to probabilistic hardware

Extropic published Z1T, a family of sparse transformer-like models designed around its Z1 probabilistic processor. The study combines Z1 with conventional programmable logic and releases model weights and training recipes. Its energy projections exclude the final dense output calculation and data movement between the processors.

The proposed system divides inference between probabilistic chips and conventional processors, with intermediate results passing between them. Extropic explicitly explores this arrangement at rack and datacentre scale. That makes the physical layout and communication fabric part of the computing architecture: closely connected stages need enough bandwidth and predictable transfer times to keep each other occupied. Different processor types may therefore require different patterns of connection from an otherwise uniform accelerator cluster.

Imec increases the density of superconducting circuits

At the Applied Superconductivity Conference, imec presented niobium-titanium-nitride circuits with three metal layers and a design density of 3.8 million Josephson junctions per square centimetre. It also demonstrated 30nm superconducting wires. The research platform uses processes compatible with 300mm semiconductor manufacturing.

Superconducting digital logic could move some processing or control into the same cold environment as quantum devices and specialised sensors. This would change the boundary between room-temperature computers and cryogenic equipment. Fewer signals might need to cross that boundary, but the local circuits would still require power, connections and interfaces to the outside world. Imec's work on integration with photonics is relevant to how those signals could eventually be carried.

DNA assembly performs programmable calculations

A Maynooth-led team published a scaffolded DNA computer in Nature. Its molecular components assemble so that the intended computational result is thermodynamically favoured. Demonstrations included adding 25-bit numbers and detecting parity. Small programs ran in under a minute, and the system could be reused. A preprint appeared in 2025.

Here, computation takes place through the physical evolution of molecules. Designing the possible molecular states makes the desired answer energetically favourable. It is a distinctive approach to computing, albeit with a more natural connection to chemical measurement and biological experiments than to today's network infrastructure.

Optical processing reduces work at the fibre receiver

A Nature Photonics paper describes programmable optical processing in a 1,300km experimental transmission system using three-mode fibre. Photonic processors at intermediate nodes adjust how the optical modes mix, reducing modal dispersion and the complexity of digital equalisation at the receiver. The work first appeared as a May 2025 preprint.

This puts computation directly into the transmission path. Some work needed to recover the signal moves from receiver electronics into programmable optical components along the route. It illustrates a physical change in distributed computing within a telecom network, below its packet and software layers. The approach is particularly relevant to future systems that use several spatial modes to increase fibre capacity, where separating the received signals can become computationally expensive.

Quantum software gains a more useful performance framework

NVIDIA added CUDA-Q Logical to its open-source platform for designing and testing fault-tolerant quantum workflows. It also incorporated Sandia's QUOPS benchmark, which assesses the size and speed of computations a machine can execute successfully. The September qStack workshop separately examined quantum systems engineering and error-correction overheads.

For a cloud or telecom provider offering access to quantum computing resources, useful performance includes classical processing and communication around the quantum job. Some applications may alternate between remote job submission and substantial local computation. Others need rapid feedback close to the quantum hardware. QUOPS helps characterise the machine, while a service provider must also account for queueing, data preparation and network delays when describing the experience delivered to a customer.

Alice and Bob explores simpler cat qubit control

Alice & Bob and ENS Lyon reported that a coupler, driven by a steady voltage of about one microvolt, could produce the photon-pair dissipation used in cat-qubit stabilization (Cat qubits are an advanced type of superconducting qubit that encode information in Schrodinger-inspired superpositions of coherent light states, which can resist bit-flip errors). Changing the voltage enabled different photon-conversion processes on the same device. The company highlighted the experiment following an August preprint.

Scaling quantum hardware also means scaling the wiring and control electronics around it. A steady-voltage drive could simplify part of that arrangement, although the complete device still needs appropriate filtering, connections and calibration. The relevance is chiefly inside the quantum machine, where the physical control system affects how many components can be integrated and operated together.

Infleqtion entangles 30 logical qubits using 80 atoms

Infleqtion reported an experiment connecting 30 logical qubits encoded in 80 physical atoms on its Sqale processor. Its technical account describes approximately 1,000 physical operations, efficient entangling gates and software recovery of some results affected by atom loss. The downstream circuit used distance-2 encoding, with limited error protection.

Logical qubits combine physical resources to detect or manage errors, but their capabilities vary with the encoding and operations used. This experiment brings encoding, atom movement and compilation into the same account of system performance. It is primarily a computing milestone within one processor. The clearest link to telecoms is understanding what increasingly complex quantum machines can reliably execute, and over time what that might mean in terms of inter-site connectivity.

Fujitsu integrates diamond spins with photonic circuits

Fujitsu announced a diamond-spin quantum computer prototype developed through work with TU Delft and QuTech. It integrates tin-vacancy centres in diamond with photonic circuits and translates quantum programs into physical control sequences. Fujitsu plans a multi-module prototype in 2027 and is aiming to link this approach with superconducting quantum hardware.

Optical interfaces are central to the proposed modular architecture. Separate units would need to exchange quantum information as part of one computation, making interconnection a key determinant of machine performance. Fujitsu's prototype extracts photons through locally-integrated waveguides, so extending such connections over telecom fibre would require attention to wavelengths and losses, as well as the successful entanglement rate and how long quantum states can be retained.

US CHIPS awards target quantum manufacturing bottlenecks

The US Department of Commerce finalised the CHIPS research awards, including up to $375 million for GlobalFoundries and $100 million for Quantinuum. GlobalFoundries' award supports a domestic quantum foundry and integration capabilities. Quantinuum's work targets manufacturing needs for trapped-ion systems, including integrated photonics and specialised optical components.

Quantum processors rely on a surrounding supply chain of optical components, packaging and control electronics. Some manufacturing expertise overlaps with that used in telecom photonics, even where the wavelengths and device requirements differ. Better access to these processes could also help smaller companies develop interfaces between quantum modules or components for sensing. The relevance to networks is therefore partly industrial: who can manufacture the hardware used to connect the next generation of machines.

Xanadu and ASML tackle optical losses in quantum chips

Xanadu and ASML announced a collaboration on lithography for photonic quantum hardware. They will investigate how manufacturing processes can reduce optical losses, including those associated with roughness at the edges of fabricated structures. The aim is to improve the components needed for larger, error-corrected photonic systems.

Lower loss can change the feasible layout of a photonic quantum machine. Photons may need to pass through a succession of components and connections, so losses accumulate as the system grows. Improving fabrication could allow a more complex circuit or ease the demands on other parts of the design. Related process knowledge may also benefit optical interconnects and sensing, although different materials and wavelengths limit how directly an improvement transfers between products.

Backline connects quantum systems to fast classical control

Xanadu and AMD released Backline, an open platform built around PennyLane (an open-source quantum software platform) for communication between quantum hardware and classical processors. It connects CPUs, GPUs and programmable logic through a common development environment. The announcement reports local control-loop times below three microseconds in its tested configuration.

Quantum feedback and error correction create a distributed computing problem with extremely short communication latency budgets. Control measurements must reach classical processors and the resulting decisions must return quickly enough to be useful. Such loops favour very close hardware and predictable interconnects. A remotely accessed quantum service can therefore contain two quite different networks: a local fabric supporting operation of the machine, plus external connectivity for users, data transfer and less time-sensitive work.

Eurofiber prepares a metropolitan testbed for quantum links

Eurofiber and Delft Networks established an entanglement-ready dark-fibre loop between Delft and Rotterdam in the Netherlands. The testbed is intended to evaluate links between quantum memories, over existing metropolitan fibre infrastructure, including the effects of attenuation and synchronisation. It gives the partners an operational setting for quantum-network engineering.

An entanglement network could connect quantum processors or sensors so that they share a resource needed for a joint task. Its performance depends on the quality and rate of entanglement delivery, plus how long endpoints can preserve the resulting states. Conventional communication is also needed to coordinate operations. This could make route characteristics part of deciding where quantum equipment is installed and which nodes can usefully work together.

Mesa Quantum funds compact timing and navigation systems

Mesa Quantum raised $11.8 million to advance chip-scale quantum sensing, moving from components towards integrated timing and navigation units. The company is targeting applications that need to maintain operation when satellite positioning or timing signals are unavailable or unreliable, with funding supporting integration and production work.

Given ongoing threats to conventional timing platforms such as GPS, this is very timely. Telecom networks depend on timing for functions such as aligning transmit and receive periods for cellular radios. A better local reference can extend holdover, the period during which equipment stays within its required timing accuracy after losing an external source. Compact atomic clocks could give operators more flexibility over where timing is generated and how it is distributed, including at remote sites or within private networks. This is one of a number of novel approaches being attempted or deployed around PNT (positioning, navigation and timing).

A programmable quantum photonic processor operates in orbit

Researchers reported operating a programmable quantum photonic processor aboard a nanosatellite. The experiment manipulated two photons in an integrated circuit and observed quantum interference. It combined onboard generation, processing and detection of non-classical light, in equipment that had undergone launch vibration and stresses, and then operated in the space environment.

The authors identify future possibilities in onboard processing of Earth-observation data and distributed quantum networks. Both would affect satellite connectivity, in different ways. Local processing could reduce the data that needs to be downlinked. A quantum-network payload would instead need optical links and coordination with other nodes, likely via inter-satellite and ground-to-space lasers. The present result establishes experience with generating, manipulating and detecting non-classical light in orbit, including maintaining the conditions needed for photon interference.

SK hynix demonstrates processing in memory within a server

At AI Infra Summit, SK hynix showed its AiM processing-in-memory chip, an AiMX accelerator card and a server containing the card. Visitors could also try an associated language-model service. The architecture places computational functions within the memory itself, to reduce transfers to a separate processor.

Processing within memory changes where work happens inside a server. If it improves the performance and power budget of a complete inference service, it could also change which workloads fit in a small enterprise installation or an edge facility. Suitable operations could remain close to stored data, while other stages run on conventional processors / cloud and AI. The balance between these components will influence internal connections and how much work needs to cross a datacentre network.