d-Matrix Acquires GigaIO's Data Center Business, Betting That Inference Is a Systems Problem

d-Matrix, the in-memory compute startup backed by $450 million in funding at a $2 billion valuation, has acquired GigaIO's data center business. The deal, announced April 2, brings GigaIO's FabreX PCIe memory fabric and SuperNODE rack-scale system under the d-Matrix roof. Financial terms were not disclosed.

The acquisition converts what started as a 2025 partnership into full vertical integration. d-Matrix now controls every layer of its inference stack: the Corsair accelerator, the JetStream networking card, the Aviator software layer, and now, the rack-scale interconnect fabric that ties it all together.

"Inference is bigger than any one chip. It's now a systems problem," CEO Sid Sheth said in the announcement. That's the thesis - and the acquisition is the most concrete bet d-Matrix has made on it.

What d-Matrix gets

The headline asset is FabreX, a PCIe-based memory fabric that uses Non-Transparent Bridging (NTB) to virtualize CPU, GPU, and accelerator resources across servers into a shared 64-bit virtual address space. FabreX delivers sub-200-nanosecond cross-server memory access at up to 512 Gbps on PCIe Gen4, roughly 10x lower latency than d-Matrix's own JetStream NIC, which operates at 2 microseconds over standard Ethernet.

The second asset is SuperNODE, a single-node server supporting up to 32 accelerators. GigaIO used SuperNODE to post 46,755 tokens per second on Llama 2 70B using 16 AMD MI300X GPUs, claiming the highest single-node throughput in the MLPerf 4.1 Inference database. One caveat: those results were explicitly marked as unverified and were not officially submitted to or verified by MLCommons. They should be treated as self-reported performance data, not independent benchmarks.

d-Matrix also gains GigaIO's Carlsbad, California engineering team, making it the company's sixth global hub.

Why FabreX matters for d-Matrix

Without FabreX, d-Matrix's rack-scale story relied on JetStream and standard Ethernet. That works. JetStream's 400 Gbps bandwidth and 2-microsecond latency are respectable. But Ethernet-based scaling hits a wall for tightly coupled inference workloads where chips need to share model weights and KV-cache data across nodes with minimal overhead.

FabreX fills that gap. PCIe fabric at sub-200ns is a different class of interconnect, closer to what NVLink provides within a single server than what Ethernet provides between them. For d-Matrix's target workload - models up to 100 billion parameters distributed across a single rack - this level of latency matters.

The practical architecture likely combines both: FabreX for tight scale-up within a rack, Ethernet via JetStream for looser scale-out across racks. That two-tier approach mirrors what NVIDIA does with NVLink (intra-node) and InfiniBand (inter-node), but built on PCIe and Ethernet rather than proprietary interconnects.

The SquadRack connection

The 2025 collaboration between the two companies wasn't abstract. GigaIO CEO Alan Benjamin appeared at d-Matrix's SquadRack launch at OCP Global Summit in October 2025, publicly endorsing the platform. SquadRack is d-Matrix's reference architecture pairing Corsair accelerators with JetStream NICs, Supermicro X14 servers, Broadcom PCIe switches, and Arista Ethernet switches.

d-Matrix claimed SquadRack delivers 3x cost-performance, 3x energy efficiency, and 10x faster token generation compared to "traditional GPU accelerators." Those numbers lack named baselines. Which GPU? At what batch size? On which model? Until d-Matrix publishes apples-to-apples comparisons, those figures are marketing, not benchmarks.

SquadRack was targeted for availability through Supermicro in Q1 2026. With the acquisition landing on April 2, the question is whether SquadRack is actually shipping, and whether FabreX is already integrated or remains a roadmap item.

The Corsair foundation

At the center of all this is the Corsair accelerator, presented at Hot Chips 2025. The card uses d-Matrix's Digital In-Memory Compute (DIMC) architecture: 8 chiplets on TSMC 6nm, 2 GB of on-chip SRAM with 150 TB/s internal bandwidth, and 256 GB of LPDDR5X for capacity. It runs at 2,400 TOPS INT8 in a PCIe Gen5 x16 form factor drawing around 600 watts.

d-Matrix claims 60,000 tokens per second at 1ms latency for Llama 3 8B on a single server, and 30,000 tokens per second for Llama 3 70B at rack scale. The 38 TOPS/W efficiency figure was competitive when Corsair launched in Q2 2025, but the inference accelerator market moves fast, and those numbers need fresh context against current Groq, Cerebras, and NVIDIA offerings.

Competitive context

d-Matrix isn't the only company deciding that inference requires owning the full stack.

NVIDIA acquired Groq for approximately $20 billion in late 2025, buying the fastest LLM inference engine on the market and folding it into its own ecosystem. Cerebras has partnered with Amazon on disaggregated inference, combining Trainium3 chips with the Wafer-Scale Engine for a hybrid approach. SambaNova raised $350 million in a Vista-led Series E in February 2026 and claims its SN40L architecture can replace 320 GPUs with 16 chips for serving 671-billion-parameter models.

What separates d-Matrix is its starting point: a PCIe card, not a proprietary rack or wafer-scale chip. That's both the advantage and the constraint. PCIe cards slot into existing server infrastructure, which lowers the barrier to adoption. But PCIe bandwidth caps out well below what NVLink or custom wafer-scale interconnects can deliver. FabreX extends that ceiling with NTB-based pooling, but it's still PCIe underneath.

For models up to 100B parameters - the sweet spot d-Matrix is targeting - that may be enough. For the 400B+ frontier models that are driving headlines, it probably isn't. d-Matrix is making a bet that the volume market for inference will be mid-sized models served at massive scale, not the largest models served at any cost.

What to watch

GigaIO Inc. continues as an independent entity focused on edge computing, though the announcement didn't clarify leadership or what happens to existing GigaIO data center customers.

The real test comes when d-Matrix ships the integrated platform and publishes independent, verifiable benchmarks. The company has a clean architectural story (Corsair for compute, FabreX for scale-up, JetStream for scale-out, Aviator for orchestration), but the performance claims are all first-party. No MLCommons-verified results. No named GPU baselines. No third-party validation of the combined system.

d-Matrix has the funding, the engineering team, and now the interconnect IP. What it needs next are receipts.