Skip to content

GPU Platforms

d-Matrix at OCP 2026: Read the Attention and Expert Split Behind the Rack

d-Matrix's October announcements separate a Corsair/H100 work split from a later Raptor roadmap. Read what each means for a complete inference system.

Leviathan SystemsPublished 2026-10-204 min read

In its October 12, 2026 Squadron announcement, issued at OCP Global Summit, d-Matrix said its multi-rack systems were shipping to priority customers and its Corsair servers and SquadRacks were generally available. Those are the supplier's reported availability categories. A particular buyer still needs confirmation for its configuration, access and delivery window.

The accompanying attention–FFN account gives enterprise buyers a more useful question than whether a proposal calls itself heterogeneous: where will the model's computation, learned weights and live sequence state run? That placement defines the serving arrangement being bought. Reading the October implementation alongside the company's September Raptor announcement also requires keeping two hardware generations and delivery horizons separate.

Follow the work inside the model

Attention lets a token's representation draw on information from other tokens in the available context. The feed-forward network, or FFN, transforms that representation. In a sparse mixture-of-experts model, a router selects expert networks to perform that work for a token. Hugging Face's explanation of MoE models distinguishes those selected experts from shared components such as attention. Experts are a feature of the model architecture; they are not present in every model.

In d-Matrix's same-day Corsair/H100 technical account, H100 GPUs handle attention, retain the key-value cache and route tokens to selected experts. Corsair holds the expert weights and performs their computation. Intermediate states travel to Corsair and results return to the GPUs repeatedly as processing advances through the model. The described communication uses Ethernet-based JetStream.

The separation is within the model's work. By comparison, prefill/decode disaggregation assigns prompt processing and subsequent token generation to separate serving pools. A proposal that says only disaggregated inference leaves the buyer unable to tell which separation it actually supplies.

The attraction is matching different work to different resources. Whether that helps the application depends on the actual computation and its coordination. An FFN label alone does not establish a memory bottleneck across every model, batch or operating condition.

Moving expert weights leaves sequence state to serve

Weights are the model's learned parameters. They and a conversation's retained state describe different memory populations. The key-value cache stores attention information from processed tokens for reuse during generation. A dynamic cache grows as tokens are added; other cache strategies can reserve or limit its extent. Moving a set of learned weights does not move or eliminate that sequence state.

For the described split, the GPU side therefore remains relevant as context grows. A buyer whose document assistant must retain longer conversations has a question about that side of the arrangement even when the expert accelerator's weight storage is unchanged. The distinction between model weights and inference memory helps explain why a weight-storage label cannot settle the allocation.

Corsair's SRAM-based design identifies a hardware characteristic. The enterprise outcome also depends on the remaining GPU work, communication and supported software. Faster access to one population of weights cannot, by itself, establish the response time of the whole service.

Corsair's implementation and Raptor's roadmap have separate clocks

The September 10, 2026 d-Matrix partnership release concerns next-generation Raptor XPUs integrated into NVIDIA MGX rack-scale infrastructure with NVLink Fusion. It states that initial availability of Raptor XPUs integrated into the MGX rack is expected in Q4 2027. That is a dated expectation for a later platform.

NVIDIA's September 10 counterpart confirms the planned Raptor integration. It describes an XPU scale-up domain and racks working alongside GPU systems for disaggregated inference. As a counterparty account of the collaboration, it supports the partnership's intent; it supplies no independent validation of d-Matrix's application benchmarks.

A common rack architecture may matter to a future infrastructure plan. It does not establish today's delivery, demonstrate a shared GPU/XPU memory domain, or make the October Corsair example an NVLink Fusion implementation. The product, interconnect and expected status must stay attached to each account.

Read a current offer and a later roadmap as different decisions

Consider a fictional enterprise preparing a document assistant based on a selected mixture-of-experts model. It receives a current Corsair-based proposal and a presentation about later Raptor integration.

For the current proposal, a price for the expert accelerator system is only part of the comparison. The enterprise also needs the GPUs still doing attention, the communication path joining the two sides and the software support for the intended application. Existing GPUs might be usable resources if the proposed arrangement supports them; describing them as already owned does not make their capacity unlimited or their operational support free.

Now change the proposed application from brief exchanges to longer document conversations. Keeping the same expert-weight storage does not establish that the retained GPU allocation still serves the requirement. Nor does adding expert hardware automatically resolve a limitation elsewhere. The buying unit is the complete supported serving arrangement, including the resources supplied by the enterprise itself.

The Raptor presentation answers a different planning question: which future integration is the supplier pursuing? Its announced rack architecture cannot fill an omitted dependency in the current Corsair proposal. If the enterprise elects to consider the later platform, its availability and supported application remain matters for that future offer.

Judge the service produced by the split

A vendor tokens-per-second headline cannot establish whether this enterprise's application meets its requirements. Use the GPU benchmark comparison guide for the complete workload, output-quality and latency comparison. The additional question illustrated here is whether the evidence and offer include both compute roles and the path between them.

Offload can be useful while retaining a communication cost and a software dependency. The value has to survive those dependencies in the complete application. That is why the October account matters: it makes the division of work concrete enough to examine, while the September announcement places a separate integration path on a future timetable.

Once equipment is selected, its placement and required connections can inform a physical deployment discussion. Leviathan's Complete Physical Infrastructure Deployment service is the relevant service page for scoping receiving, rack integration, onsite installation, IT cabling and physical handoff records.

Contact

Discuss Your Infrastructure Project

Tell us about your project, and our team will follow up to discuss the scope, requirements and availability.