‘Inference Is Bigger Than Any One Chip’ – d-Matrix CEO on GigaIO Deal

Note: This content originally appeared on datacenterknowledge.com, click here to read the original piece.

The deal brings PCIe fabric and rack-scale system design in-house, as d-Matrix moves beyond silicon into full-stack AI infrastructure.

A push to build more complete, rack-scale AI inference systems is driving d-Matrix’s acquisition of GigaIO’s data center business. The deal adds interconnect technology and systems expertise as competition shifts beyond individual chips.

The deal builds on a collaboration that began in 2025 and is aimed at strengthening the company’s ability to deliver system-level AI infrastructure rather than discrete silicon.

Expanding Beyond the Chip

With the acquisition, GigaIO’s data center technologies, which include its SuperNode platform and FabreX PCIe-based memory fabric, are being integrated into the broader d-Matrix inference stack.

The move extends a platform that already includes Corsair inference accelerators, JetStream networking, Aviator software, and the SquadRack rack-scale reference architecture developed with Broadcom and Arista.

“Inference is bigger than any one chip. It’s now a systems problem,” said Sid Sheth, founder and CEO of d-Matrix. “To keep up with surging AI demand, workloads are increasingly disaggregated across CPUs, GPUs, and inference accelerators. That means data must move efficiently across chips, nodes, racks, and entire data centers in real time.”

He said the acquisition aims to accelerate delivery of low-latency, efficient, and scalable infrastructure.

Deal Structure and Strategic Rationale

The transaction is structured as a business unit acquisition, with ownership of GigaIO’s data center-related assets transferring to d-Matrix.

“It’s a business unit acquisition in which we are acquiring the unit’s related assets,” Sheth told Data Center Knowledge.

GigaIO will continue to operate independently while focusing on edge computing. Financial terms were not disclosed.

More broadly, the move reflects a shift toward owning more of the system stack as inference workloads become increasingly distributed and infrastructure design moves beyond individual chips.

Market Focus and Business Impact

Target customers include hyperscalers, frontier AI labs, and enterprise deployments where inference workloads are spread across heterogeneous compute environments.

According to Sheth, the acquisition is expected to accelerate the company’s revenue trajectory and expand how deployments are monetized.

“We expect it to pull in the revenue trajectory,” he said.

He added that system-level capabilities could support new pricing models and increase value per rack.

“Yes to both,” Sheth said. “We are positioned to enable premium AI inference services across a range of rack configurations, including heterogeneous systems optimized for disaggregated workloads.”

Competitive Landscape: Interconnect Approaches Diverge

The deal places d-Matrix more directly into a competitive landscape defined by differing approaches to system interconnects and scaling.

GPU-centric systems from Nvidia rely on proprietary high-speed interconnects such as NVLink and NVSwitch to tightly couple GPUs. AMD similarly uses Infinity Fabric to enable high-bandwidth communication across its platforms.

By contrast, PCIe-based fabric technologies such as FabreX extend standard PCIe connectivity across nodes, enabling disaggregated pools of compute and memory that can be composed at the rack or cluster level.

Stephen Sopko, analyst-in-residence at HyperFrame Research, said the deal strengthens the company’s move toward system-level ownership.

“This acquisition bolsters d-Matrix’s rack-scale AI inference capabilities by internalizing SuperNode architecture and the FabreX PCIe fabric,” Sopko told Data Center Knowledge. “That moves them from collaborative reference designs to a more complete, owned systems stack for low-latency enterprise deployment.”

He said PCIe-based fabrics can provide advantages in disaggregated environments.

“PCIe-based fabrics, like FabreX, enable flexible pooling of accelerators and memory without the overhead of traditional Ethernet fabrics,” Sopko said. “That is key to cost- and power-efficient scaling.”

Sopko added that the approach highlights a growing architectural divide.

“Heterogeneous inference is gaining real momentum in 2026, though broad production adoption still faces integration hurdles,” he said. “This deal sharpens d-Matrix’s contrast with NVIDIA and AMD’s platforms – doubling down on fabric-enabled disaggregation that lets operators mix CPUs, GPUs, and specialized accelerators.”

These approaches reflect different design priorities – tightly integrated, high-bandwidth systems on one hand, and more modular, composable infrastructure on the other.

Shift Toward Rack-Scale, Disaggregated Infrastructure

Adding GigaIO’s technology and engineering team aligns with the shift toward rack-scale architectures for distributed inference workloads.

PCIe-based fabrics such as FabreX enable high-speed communication between compute and memory across nodes, supporting more flexible system design.

That flexibility can improve utilization in AI inference environments, particularly as workloads are spread across heterogeneous resources.

The acquisition enhances the company’s ability to support these system-level deployments as inference becomes more distributed.

As part of the transaction, a team of systems engineers based in Carlsbad, California is being added, establishing a new engineering presence in Southern California.

The company now operates six innovation hubs across North America, Europe, and Asia.

View source version on datacenterknowledge.com: https://www.datacenterknowledge.com/data-center-hardware/d-matrix-buys-gigaio-data-center-unit-to-push-into-rack-scale-ai-inference