Rivvor demonstrates wireless interconnects for AI computing
Contactless blades exchange data via Bounded Electro-Magnetic Corridors (image: Rivvor)
Replacing a computing module can involve tracing cables, reaching past neighbouring equipment and reconnecting high-speed interfaces. Rivvor has built a system in which data travels between modules without those cables or mating connectors.
At the Open Compute Project Global Summit in San Jose, the California startup demonstrated its Valen Edge Series: three AMD Ryzen computing blades communicating through millimetre-wave wireless links. Each blade carried 128 gigabytes of memory.
Announced on 29 October 2026, the demonstration delivered 15 gigabits per second per link and application-memory-to-application-memory latency of 250 nanoseconds, according to Rivvor. The company says it developed the working system in 11 months.
Founded in 2025 and based in Roseville, Rivvor is led by co-founder and chief executive Nazym Paltachev. Its team combines processor design and wireless engineering: chief science and product officer Timour Paltachev led research and development teams at AMD and worked on the US Department of Energy’s CORAL-2 programme, while chief technology officer Konstantin Tiutin brings experience in millimetre-wave, virtual-reality and satellite systems.
Why AI computing needs fast interconnects
Interconnects are the data links between computing components. They move information between the processors and other devices doing the work. When an AI model runs across multiple graphics processing units, or GPUs, those processors exchange intermediate results as they generate an answer. Communication speed therefore affects how quickly the system can complete its task. Running a trained model to produce an output is known as inference; edge computing places that work near the devices or users it serves.
The figures reported for Valen describe two aspects of that communication: bandwidth, the amount of data a connection carries per second, and latency, the delay involved in transferring it. Large transfers depend on sufficient bandwidth, while frequent exchanges also make a workload sensitive to the delay each transfer introduces.
Providing those connections takes space alongside processors, memory, power delivery and cooling equipment. Adding modules also increases the number of interfaces technicians must install and manage, making the layout of the rack part of the communications problem.
How Rivvor’s wireless interconnects carry data
Rivvor addresses that layout through wireless links engineered into the enclosure. Valen uses millimetre waves—radio waves with wavelengths measured in millimetres—to carry data over the short distances between computing blades.
Rivvor’s broader architecture includes what it calls Bounded Electromagnetic Corridors (BEC). These guide signals through structural paths within computing equipment, with modules coupling to those paths across a controlled gap. The surrounding structure therefore becomes part of the communications system.
The corridor supplies the route, while the module’s interface transfers data into or out of it. Mechanical supports hold the module securely in place, separating its physical attachment from the high-speed data connection. [4]
Rivvor also describes directional links that establish paths through an enclosure with known geometry and endpoint positions. Both approaches form part of a connectivity layer, or fabric, linking computing and switching modules under the control of software and a fabric controller.
Software-controlled connections between computing modules
Within that fabric, software establishes, calibrates and monitors the available paths. Where the installed hardware supports multiple routes, it can change connections without manual recabling. This extends software control from directing network traffic to managing the physical wireless paths carrying it. [4]
In its June announcement, the company outlined steerable beams and software-controlled connections, together with wireless network interface cards and rack-level controllers intended for integration into equipment manufacturers’ designs. It positions the technology alongside copper and optical links. [6]
For a workload that exchanges substantial amounts of data between particular modules, this architecture is designed to make the available connections configurable around those exchanges. The hardware supplies the possible routes; software manages which of them are used.
Rivvor describes two ways of delivering its architecture: integrating the connectivity layer into a customer’s enclosure and modules, or supplying a complete modular computing system with the connections built in. Its design spans compact groups of blades, populated racks and selected engineered links between racks. [5]
For servicing, Rivvor describes a replacement process in which the fabric detects an equipped module and establishes and monitors its links. This makes connection management part of the system’s control software rather than a separate cable-handling task.
From millimetre-wave links to sub-terahertz interconnects
Valen provides a physical implementation of Rivvor’s approach. Its October announcement places that system within a longer development programme, including a move towards sub-terahertz links operating between 100 and 500 gigahertz.
Rivvor plans to use a purpose-built chip for this next stage, targeting more than 400 gigabits per second per link. The company is also holding a further live demonstration in Santa Clara on 12 November.
Beyond data centres, Rivvor identifies robotics, industrial equipment and remote computing as applications for its contactless architecture. These share a practical interest in replaceable electronics and accessible connections, particularly where equipment is serviced repeatedly or physical access is difficult.
With Valen, Rivvor has brought its wireless interconnect concept into a working modular computer. The opportunity extends beyond carrying bits across a gap: designing communications into the enclosure gives equipment builders another way to make computing infrastructure easier to assemble, adapt and service.
Liked this article? You can support our independent journalism via our page on Buy Me a Coffee. It helps keep MoveTheNeedle.news focused on depth, not clicks.