AINEWS
The San Diego company is shipping suitcase sized Gryf and fixed site Manticore boxes for AI workloads that cannot reach a cloud and cannot fit on a phone, with a connectivity partnership with Mushroom Networks.
Hyperscale data centers run training and the heaviest inference. Phones run lightweight queries. A drill bit in a mine, a sensor on a defense site, and a camera on an automated farm generate more data than they can ship to a cloud and more work than a phone can do in a single battery cycle. San Diego-based GigaIO is betting its entire company on that gap, the edge, where compute has to live close to where the data is generated because neither the data center nor the device can physically serve the workload.
The bet became visible on April 2, 2026, when Bay Area inference-chip company d-Matrix announced it was acquiring GigaIO's data center business, including the SuperNode platform and the FabreX PCIe-based memory fabric that lets accelerators share memory across a rack. d-Matrix folded the GigaIO stack into its inference lineup alongside its own Corsair accelerators, JetStream networking, and the SquadRack reference architecture it co-developed with Broadcom and Arista. The deal let GigaIO keep its name, its engineering team, and two product lines that have nothing to do with the data center at all.
Two product lines survived the spin-off. GigaIO launched Gryf in 2025 as a carry-on suitcase-sized appliance that combines CPU, GPU, storage, and networking into a single box aimed at portable, data-center-class performance. Manticore is a larger on-premises device that sits in a fixed site, ingests data from sensors, cameras, and instruments, and processes it locally rather than shipping it to a cloud. Gryf is the mobile unit; Manticore is the fixed unit. Both run inference on the customer's premises, by definition, because the customer's site is where the data already is.
The architectural argument rests on three properties of that gap. The data center is unmatched for training and for the largest batch-inference jobs, but it cannot see what a drill bit in a mine is doing in real time over a satellite link that drops every other hour. The phone is unmatched for low-latency personal use, but a phone's power budget caps the model it can run. GigaIO CEO Alan Benjamin describes that fundamental power limitation as the structural reason the middle layer exists, and the workloads neither end can physically serve as the envelope edge compute is sold into.
In July 2026, GigaIO announced a partnership with San Diego-based Mushroom Networks to address connectivity bottlenecks between the far edge, meaning robotics and smart municipal devices, and the near edge, meaning a single building or manufacturing facility. The deal's framing is plumbing: Gryf- and Manticore-class hardware are only useful if the data they generate can move to a place that can also compute on it. GigaIO's source material does not name the bandwidth standard, the protocol, or any deployment counts, so the partnership is currently best read as a statement of direction rather than a shipped product line.
The verticals GigaIO cites for Gryf and Manticore, including defense, manufacturing, automated agriculture, oil and gas, and mining, share three properties: they generate more data on site than they can ship to a cloud over a stable network, they cannot wait for a round trip, and they cannot shrink the model down to phone scale. The reason an edge play can be made at all in 2026 is that inference accelerators and high-bandwidth memory have become cheap enough and small enough to fit into a suitcase or a server rack that lives outside a hyperscale lease. GigaIO is not the only company pursuing that envelope, and the hydrated source material does not include independent demand-side validation, customer counts, deployment figures, or revenue, so the architectural argument is firmer than the commercial one.
The d-Matrix deal, the Gryf and Manticore product split, and the Mushroom Networks partnership together describe a company that has chosen a single layer of the AI compute stack and removed every product line that did not fit. Whether that bet pays off depends on how many sites the connectivity plumbing can reach and how quickly the inference silicon inside the suitcase catches up with the inference silicon inside the rack.