> ## Content Index > Fetch the complete content index at: https://nexi.fund/llms.txt > Use this file to discover other available public pages before exploring further. # Nvidia Is Funding the Chip Startup Built to Beat It - URL: https://nexi.fund/d-matrix-nvidia-nvlink-inference-2026/ - Published: 2026-10-14T18:30:44.000Z - Updated: 2026-10-14T18:30:44.000Z - Description: Nvidia plans to take a stake in d-Matrix at a reported ~$5 billion, months after d-Matrix agreed to put its Raptor inference XPUs inside Nvidia's MGX racks through NVLink Fusion. The deal tests whether the rack, not the chip, decides who wins inference. - Author: Nexi.fund Labs - Tags: AI & Infrastructure, #mode-1, #hook-paradox, #track-A, #layout-thesis-spine, #brand-heavy Nvidia is preparing to fund a rival that designs chips to do the work its own GPUs are sold for. ๐ŸŽฏ The reported stake is small next to Nvidia's balance sheet. The signal is larger: Nvidia is standardising the rack, not only the chip. d-Matrix trades independence for distribution โ€” NVLink Fusion puts its Raptor XPUs inside Nvidia's own MGX racks. Compatibility now has to become a design win. Raptor does not ship until Q4 2027\. On 8 October, The Information reported that Nvidia plans to take a stake in d-Matrix, a seven-year-old inference-chip startup, as part of an effort to make its platform work with rival silicon. The company was last valued at $2 billion in November 2025\. The new round is being discussed near $5 billion. Three weeks earlier the two firms had already announced the partnership that makes the stake plausible. d-Matrix will build its next-generation Raptor XPUs into Nvidia's MGX rack architecture through NVLink Fusion, the interconnect Nvidia opened to outside chipmakers in May 2025. --- ## The $2 billion startup Nvidia just repriced d-Matrix started in 2019 and sells one thing: faster, cheaper inference. Its current product line โ€” Corsair accelerators, JetStream network cards, and the Aviator software stack โ€” is built around a single idea. Move the compute into memory instead of shuttling data between a processor and a separate chip. That choice defines both the product and its limits. The company claims Corsair delivers up to 10x faster performance, 3x lower cost, and 3โ€“5x better energy efficiency than GPU-based systems. It also concedes the architecture fits narrow workloads better than very large models. Investors have funded the bet in stages. A $44 million Series A in 2022 drew Playground Global, Microsoft's M12, and SK Hynix. A $110 million Series B followed in September 2023 with Microsoft and Temasek. The $275 million Series C in November 2025 โ€” led again by M12 and Temasek โ€” put the company at $2 billion. $5B reported target valuation โ†‘ $3B vs Nov 2025 #### d-Matrix valuation, reported From $2B in November 2025 to a reported \~$5B target, unconfirmed on size or timing. ยท *The Information, Oct 2026* The markup is the easy part to read. The harder question is why Nvidia would help a direct competitor get bigger. ## Why Nvidia would fund a rival Nvidia dominates the training market. Inference is the larger and more contested pool: it runs every time someone submits a prompt, and it repeats forever. That is where specialised silicon has a genuine opening, because a chip tuned for one job can beat a general-purpose GPU on cost per token. NVLink Fusion is Nvidia's answer. Instead of fighting every rival accelerator, it lets them plug into Nvidia's rack. A partner's XPU connects over NVLink scale-up, then Nvidia's Spectrum-X handles scale-out, MGX supplies the liquid-cooled chassis, and Vera CPUs, ConnectX-9 SuperNICs, and BlueField-4 DPUs handle the surrounding work. > Demand for inference is soaring, but capital, time and energy remain finite.โ€” Sid Sheth, co-founder and CEO, d-Matrix For d-Matrix, the deal collapses years of integration work. Sourcing high-speed interfaces, validating scale-up networking, certifying a rack design โ€” each step costs money and time, and Nvidia has already paid for all of it. In its own announcement, d-Matrix said Raptor is being designed from the ground up for NVLink Fusion and MGX, with a multi-year roadmap and a tie-up with connectivity vendor Astera Labs. The partner list tells the same story. Arm, Fujitsu, Intel, MediaTek, Qualcomm, SiFive, and Synopsys have joined NVLink Fusion. Amazon is designing Trainium4 to integrate with it. Nvidia is collecting the chipmakers that would otherwise compete on rack design, then selling them the fabric underneath. ## The economics being competed for d-Matrix's pitch rests on a split in how inference works. Generating a token has two phases: a compute-heavy prefill step, then a memory-bandwidth-heavy decode step. GPUs handle both; they are not equally good at both. The company's answer is speculative decoding. A small model running on a d-Matrix accelerator guesses the next tokens; a larger model on an Nvidia GPU checks them. If the guesses hold, the expensive model does less work per token. This is already in production, not on a slide. In July, the inference provider Parasail began deploying Corsair accelerators alongside Nvidia Hopper and Blackwell GPUs, splitting prefill and decode across the two architectures. Parasail says the setup produces up to 10x faster token generation and up to 3x better energy efficiency. It is one of the first commercial examples of heterogeneous, disaggregated inference. The next chip extends the idea. Raptor stacks a DRAM die and an SRAM compute die into a single package โ€” d-Matrix calls it a two-story design โ€” to keep more weights close to the compute. The company previewed the 3D DRAM work at Hot Chips 2026; technical details have been published through IEEE. Raptor is expected to tape out before the end of the year. ## What has to be true for this to work The reported investment is Nvidia's first move into an inference-chip startup since its $20 billion licensing agreement with Groq in December. That deal was structured as licensing, not a stake, which is a clue to how Nvidia prefers to handle rivals: absorb the technology, leave the balance sheet alone. An investment in d-Matrix is a softer instrument. It buys alignment without owning the company, and it keeps a second supplier of inference silicon inside Nvidia's ecosystem rather than outside it. That does not make the round safe. The diligence question for any inference-chip financing right now is narrow: does NVLink Fusion compatibility show up in actual design wins within two quarters, or only in a press release? Nvidia's capital does not guarantee that hyperscalers choose Raptor over Nvidia's own Blackwell-generation parts. Raptor arrives in Q4 2027, which leaves a long window for Nvidia to improve its own inference economics and close the gap the startup is betting on. The valuation, if it lands near $5 billion, prices in the optimistic path: compatibility converts quickly, design wins follow, and d-Matrix becomes the default second source in Nvidia racks. The pessimistic path looks like Groq โ€” a strong technology folded into a licensing deal on someone else's terms. ### Will a rival chip ship inside Nvidia racks by 2027? ๐Ÿ”ฎ **By the end of 2027, at least one non-Nvidia inference XPU will be shipping inside a mainstream Nvidia rack in volume.** Probability: 65% โ€” the design work is public and Nvidia is deliberately opening the rack, but volume depends on hyperscaler qualification cycles that are still unstarted. #### ๐ŸŸข Optimistic scenario (30%) Raptor tapes out on schedule and ships in MGX racks at volume. Nvidia's stake gives d-Matrix a credibility floor with hyperscalers. **Consequence:** a durable second source for inference silicon, priced at a premium. #### ๐ŸŸก Base scenario (50%) The partnership holds, shipments arrive late, and design wins concentrate in a few large customers. **Consequence:** d-Matrix stays a specialised supplier inside Nvidia's ecosystem rather than a rival to it. #### ๐Ÿ”ด Pessimistic scenario (20%) Nvidia's own inference silicon improves faster than d-Matrix can qualify, and the stake becomes a precursor to licensing. **Consequence:** the $5 billion valuation proves to be a peak, not a step. The pattern is familiar from other challengers. As we wrote in a [recent piece on Positron AI](https://nexi.fund/positron-ai-inference-silicon-series-c-2026/), the memory-first inference bet against Nvidia is a real market and a hard one to defend. d-Matrix's difference is that it no longer has to defend the market alone. It has handed the rack to Nvidia, and taken Nvidia's money in return. That trade is a clean test of where value sits in AI hardware. If interconnect standards matter more than the chip at the end of them, Nvidia wins whether a customer picks its GPU or someone else's XPU. If the silicon still decides, d-Matrix โ€” and the $5 billion being discussed for it โ€” is the better bet. [ Nvidia reported to plan a stake in inference-chip startup d-Matrix The primary account of the reported investment and the \~$5B valuation, attributed to The Information. Used for the deal's terms and timing. DataCenterDynamics, Oct 2026 ](https://www.datacenterdynamics.com/en/news/report-nvidia-plans-to-invest-in-inference-chip-startup-d-matrix/?ref=nexi.fund) The clearest single record of what is confirmed versus rumoured in the round. [ Chip startup d-Matrix to use Nvidia chip-linking technology in AI servers Reuters' coverage of the NVLink Fusion partnership and the Raptor XPU plan. Used for the technical and commercial terms of the tie-up. Reuters, Sep 2026 ](https://www.reuters.com/business/media-telecom/chip-startup-d-matrix-use-nvidia-chip-linking-tech-ai-servers-2026-09-10?ref=nexi.fund) Independent confirmation that the partnership is real and dated, not a rumour. [ d-Matrix adopts Nvidia NVLink Fusion for rack-scale XPU deployment Nvidia's own account of what NVLink Fusion and MGX give a partner, with the CEO quote. Used for the architectural detail. NVIDIA Blog, Sep 2026 ](https://blogs.nvidia.com/blog/d-matrix-nvlink-fusion/?ref=nexi.fund) The vendor's framing of why the rack, not the chip, is the unit of competition.