d-Matrix drinks the Nvidia Kool-Aid with NVLink Fusion and MGX rack designs

AI inference-chip startup d-Matrix will license Nvidia's NVLink Fusion and MGX rack designs for future Raptor accelerators, targeting systems of up to 144 accelerators on a single all-to-all NVLink fabric by end-2027. A 144-XPU rack is projected to provide roughly 2.3TB of memory capacity and 7.2PB/s of aggregate memory bandwidth, sufficient for models exceeding 4 trillion parameters at 4-bit precision. The agreement further expands Nvidia's ecosystem beyond GPUs, as d-Matrix plans to use Vera CPUs, NVSwitch, BlueField/ConnectX networking and SpectrumX Ethernet alongside its accelerators.

Analysis

NVDA’s licensing strategy should be valued less as IP revenue and more as an attach-rate mechanism for networking, CPUs and rack components. Third-party inference silicon that scales inside NVLink/MGX architectures expands the addressable inference market without requiring NVDA to win every accelerator socket; it also raises switching costs for hyperscalers and enterprises once NVSwitch, BlueField/ConnectX and software workflows are embedded. Over 6-18 months, this supports a higher mix of recurring platform revenue and reduces the bear-case that custom accelerators necessarily displace Nvidia economics.

The non-obvious loser is open-scale-up networking: a viable NVLink-compatible inference ecosystem makes Ethernet-only and alternative interconnect architectures less compelling for latency-sensitive decode workloads. MRVL benefits only if its custom-silicon engagements remain complementary to Nvidia’s fabric; otherwise NVLink standardization shifts bargaining power toward NVDA and can compress ASIC gross-margin expectations. QCOM has limited near-term read-through: edge inference remains a separate power-constrained market, but a lower-cost datacenter decode tier could delay some enterprise demand for distributed edge deployments.

Near term, this is not a material earnings catalyst because the adopter is private and product qualification remains ahead. The 1-3 month catalyst is additional public design wins or disclosed NVLink-linked networking/CPU content; the key falsifier is hyperscaler adoption of competing open fabrics or evidence that licensed partners use minimal Nvidia content beyond the interconnect. Consensus may underappreciate that NVDA can monetize inference growth even where its GPU is absent, but may overestimate the pace: enterprise deployments will require software maturity, reliability validation and economics versus GPU-only clusters.

Key Decisions for Investors

• Maintain/add NVDA on 6-12 month weakness rather than chase the headline; express through a long NVDA / short SOXX pair to isolate platform-share gains. Reassess if networking and CPU attach fail to accelerate in the next two earnings reports or if management signals material licensing-funded incentives are diluting returns.
• Avoid treating MRVL as a direct beneficiary until management identifies incremental NVLink-related design-win revenue, content per rack, and margin structure. Set an alert around the next earnings call: disclosed dependence on Nvidia-controlled fabric without offsetting custom ASIC content is bearish for multiple expansion.
• For a tactical 1-3 month view, favor NVDA over QCOM in a pair trade, sized modestly: datacenter inference infrastructure has a clearer monetization path than an indirect edge-demand read-through. Exit if NVDA underperforms QCOM by 10% after the next Nvidia results absent a negative guidance revision.
• Watch public disclosures from AWS, MediaTek, Marvell and other licensees for rack-scale deployments. A second or third production design using NVSwitch plus Nvidia NICs would justify increasing NVDA exposure; absent production commitments, treat this as strategic optionality rather than a forecast revision.