d-Matrix
d-Matrix is a privately held American semiconductor company headquartered in Santa Clara, California that builds accelerators, I/O cards and software for AI inference in data centers. It was founded in 2019 by Sid Sheth and Sudeep Bhoja, two engineers who left the data center interconnect company Inphi on the bet that serving trained models, not training them, would become the dominant cost in deployed AI.[3][5] The company's design thesis is memory-centric: instead of surrounding a logic die with high bandwidth memory, d-Matrix folds multiply-accumulate circuits into SRAM macros, an approach it calls Digital In-Memory Computing (DIMC).[3][8]
Its first commercial product, the Corsair PCIe accelerator, was announced at SC24 in November 2024 and entered full production in June 2026.[8][16] A successor, Raptor, moves the same idea into a 3D stacked DRAM package and is the first d-Matrix part slated for NVIDIA's NVLink Fusion and NVIDIA MGX rack ecosystem.[11][20] d-Matrix had raised $450 million in total as of November 2025, when a $275 million Series C valued it at $2 billion, and reported more than 250 employees.[4]
Founding and early years
Sheth and Bhoja incorporated d-Matrix in 2019. In a retrospective published on the company blog in June 2024, Sheth wrote that both founders were in their forties and had watched AI adoption accelerate from inside Inphi, whose data center interconnect products were being bought by every large cloud provider; from those customer conversations they concluded that the inference opportunity "would dwarf AI training."[5]
The first design was not the one that shipped. Sheth's account describes an initial charge-based analog in-memory compute architecture with a single-slope analog-to-digital converter, abandoned after the team concluded it could not fit converters on each bitline economically. The company pivoted to an all-digital scheme using SRAM bit cells paired with adders in a custom circuit fabric, which became DIMC. d-Matrix built and packaged its first chip, Nighthawk, within twelve months during 2020 and demonstrated it to investors at its Cupertino lab in 2021. The same post describes the company coming within two weeks of running out of cash before deciding to raise a much larger round than the $2 million it had been chasing.[5]
A second pivot followed in 2022. The first product had been aimed at non-generative transformer models such as BERT and T5; after GPT-3 and customer feedback, d-Matrix reworked the architecture for generative workloads, and that redesign became Corsair.[5]
Leadership
d-Matrix's own site lists Sid Sheth as Founder and CEO and Sudeep Bhoja as Founder and CTO.[1] Company press releases use both "founder" and "co-founder" for each man, sometimes on the same day: the Alchip and Andes releases of November 18, 2025 style Sheth "co-founder and CEO" and "Founder and CEO" respectively.[11][12] Sheth was styled President and CEO in company material published in 2024.[5]
| Name | Role | Note |
|---|---|---|
| Sid Sheth | Founder and CEO | Previously at Inphi[1][5] |
| Sudeep Bhoja | Founder and CTO | Presented 3DIMC at Hot Chips 2025 and Raptor at Hot Chips 2026[11][20] |
| PJ Jamkhandi | Chief Strategy Officer and interim CFO | [1] |
| May O'Neal | Chief Human Resources Officer | [1] |
| Sree Ganesan | VP of Product and Customer Engineering | [1] |
| Vid Jain | VP of Developer and Cloud Engineering | Founder and CEO of Wallaroo.ai before its 2026 acquisition[1][19] |
| Kristin Bryson | VP of Corporate Communications | [1] |
| Richard Ogawa | General Counsel and Chief Compliance Officer | [1] |
The board of directors listed on the company site draws one seat each from the main investors: Sasha Ostojic (Playground Global), Jeff Huber (Triatomic Capital), Michael Stewart (M12), Connie Sheng (Nautilus Venture Partners), Russell Tham (Temasek) and Per Roman (BullhoundCapital).[1]
Santa Clara has been the headquarters throughout. By September 2023 the company also had offices in Bengaluru and Sydney; the 2024 blog post adds design centers in Seattle and Toronto; and the November 2025 Series C fact sheet lists Toronto, Sydney, Bangalore and Belgrade alongside the Santa Clara headquarters.[3][5][4] The April 2026 GigaIO acquisition added a systems engineering team in Carlsbad, California, which d-Matrix said took it to six engineering sites across North America, Europe and Asia.[14]
Funding
Every round to date has been private. The figures below come from d-Matrix's own announcements.
| Round | Announced | Amount | Lead | Other participants |
|---|---|---|---|---|
| Series A | April 20, 2022 | $44 million | Playground Global | M12 (Microsoft's venture fund) and SK Hynix joining existing investors Nautilus Venture Partners, Marvell Technology and Entrada Ventures[2] |
| Series B | September 6, 2023 | $110 million | Temasek | Playground Global and M12 quoted in the release[3] |
| Series C | November 12, 2025 | $275 million | Co-led by BullhoundCapital, Triatomic Capital and Temasek | New: Qatar Investment Authority, EDBI. Follow-on: M12, Nautilus Venture Partners, Industry Ventures, Mirae Asset[4] |
d-Matrix stated the Series C valued the company at $2 billion and brought total capital raised to $450 million. Morgan Stanley acted as exclusive placement agent and Wilson Sonsini Goodrich and Rosati as legal counsel.[4]
Two figures in circulation do not match that release. Sheth's 2024 retrospective describes the Series A as "a $40M Series A from Playground Global and Microsoft M12" won in 2021, while the press release announcing it is dated April 2022 and gives $44 million.[5][2] CNBC reported in June 2026 that d-Matrix "has raised around $500 million so far, putting it at around a $2 billion valuation," against the company's own $450 million.[17][4] The $450 million figure is the one d-Matrix publishes.
Silicon and product timeline
d-Matrix ran three generations of test silicon before its first commercial part, and a fourth test chip after it. The pre-Corsair chips were named after hawks and were never sold as products; they were demonstrated to investors and evaluation customers.
| Product | Announced | What it was |
|---|---|---|
| Nighthawk | Announced April 2022 with the Series A, though the later Jayhawk release dates the platform launch to 2021; built during 2020 and demonstrated in 2021 | First proof-of-concept chiplet, initially analog in-memory compute[2][5][6] |
| Jayhawk | January 24, 2023 | Second-generation chiplet platform using the OCP Bunch of Wires die-to-die interconnect on organic substrates, 6 nm TSMC, 16 Gbps per wire, under 0.5 pJ/bit[6] |
| Jayhawk II | August 22, 2023 | Enhanced DIMC engine plus BoW interconnect, quoted at 30 to 150 TOPS/W on a 6 nm process[7] |
| Corsair | November 19, 2024 at SC24 | First commercial product, a PCIe Gen5 DIMC accelerator card; full production June 9, 2026[8][16] |
| JetStream | September 8, 2025 | I/O accelerator card for device-initiated accelerator-to-accelerator traffic over standard Ethernet[9] |
| Pavehawk | Hot Chips, August 2025 | 3DIMC test silicon proving 3D DRAM stacking, validated in d-Matrix labs; the vehicle for the technology that commercially debuts in Raptor[11] |
| SquadRack | October 14, 2025 at the OCP Global Summit | Rack-scale reference architecture with Arista, Broadcom and Supermicro[10] |
| Raptor | Detailed through 2025 and 2026 | 3D stacked DRAM accelerator, d-Matrix's 3DIMC technology; expected to tape out before the end of 2026[11][20] |
Manufacturing runs through TSMC. Corsair is built on TSMC's N6 node with Alchip Technologies as ASIC design and packaging partner, on organic substrates with LPDDR5 rather than HBM and CoWoS packaging, a choice d-Matrix describes as deliberate supply-chain risk reduction rather than a performance decision.[16]
Software and services
Aviator is d-Matrix's software stack. The company's product pages break it into seven parts: a tools suite for deployment and monitoring, a Model Factory of PyTorch model templates for distributed inference, a Compressor for block floating point numerics, an MLIR-based compiler, a distributed inference engine, a host runtime and on-chip firmware. It integrates with PyTorch and the Triton DSL and is built on open-source components including MLIR, PyTorch and OpenBMC.[21]
An early compiler effort was done with Microsoft. In November 2022 d-Matrix announced a collaboration using Microsoft's Project Bonsai reinforcement learning platform to train a compiler for its DIMC parts.[24]
In September 2026 the company announced d-Matrix Demo Cloud, a hosted evaluation environment that exposes Corsair hardware through an OpenAI-compatible API gateway so prospective customers can benchmark it without provisioning hardware. The launch post describes a sample deployment running heterogeneous speculative decoding, with a Qwen3 1.7B draft model on two Corsair cards and a Qwen3 235B target model on four NVIDIA H200 GPUs. At announcement the service was waitlisted.[22]
Acquisitions
d-Matrix made two acquisitions four months apart in 2026, both aimed at the layers above the chip.
| Target | Announced | What was acquired | Terms |
|---|---|---|---|
| GigaIO data center business | April 2, 2026 | SuperNODE platform, FabreX PCIe-based memory fabric, and a systems engineering team in Carlsbad, California. GigaIO, Inc. continued as an independent company focused on edge computing | Not disclosed; Sheth described it to Data Center Knowledge as "a business unit acquisition in which we are acquiring the unit's related assets"[14][15] |
| Wallaroo.ai | August 3, 2026 | Deployment and orchestration software, intellectual property, and the engineering, product and go-to-market teams. Founder and CEO Vid Jain joined d-Matrix | Not stated in the announcement[19] |
The stated rationale in both cases was the same. On GigaIO, Sheth argued that "inference is bigger than any one chip. It's now a systems problem," pointing to customers splitting workloads across CPUs, GPUs and inference accelerators.[14] On Wallaroo, he said the barrier customers reported was "the operational complexity" of deployment rather than raw performance, and that adding orchestration software gave d-Matrix an end-to-end stack from silicon to production.[19] The two deals had followed an existing GigaIO partnership that began in 2025, and d-Matrix credited the GigaIO team with accelerating SquadRack's path to production.[14][16]
Partnerships and named customers
| Partner | Announced | Nature |
|---|---|---|
| Microsoft (M12) | April 2022 onward | Investor from the Series A; Michael Stewart sits on the board[2][1] |
| Arista, Broadcom, Supermicro | October 14, 2025 | Co-developers of the SquadRack rack-scale reference architecture[10] |
| Alchip Technologies | November 18, 2025 | Joint development of a 3D DRAM inference accelerator, with Raptor named as the commercial debut vehicle; Alchip also partnered on Corsair's design[11][16] |
| Andes Technology | November 18, 2025 | d-Matrix selected the AndesCore AX46MPV RISC-V vector core for the Raptor architecture[12] |
| Gimlet Labs | March 12, 2026 | Gimlet Cloud to deploy Corsair alongside GPUs for agentic inference, targeted at select customers in the second half of 2026[13] |
| Parasail | July 7, 2026 | Inference cloud deploying Corsair alongside its NVIDIA Hopper and Blackwell fleet, splitting prefill onto GPUs and decode onto Corsair[18] |
| NVIDIA and Astera Labs | September 10, 2026 | d-Matrix joins the NVLink Fusion partner ecosystem; Raptor XPUs to plug into NVIDIA MGX racks, with Astera Labs supplying connectivity. Initial availability is a company projection for Q4 2027[20] |
Those are the deployments d-Matrix has put a customer name to. Speaking to CNBC in June 2026, Sheth declined to name Corsair customers, saying only that he had commitments from hyperscalers, neoclouds and frontier AI labs, that volume shipments began that month, and that roughly 90 percent of those customers were in the United States with the remainder in the Middle East and Southeast Asia.[17] In the same interview he said he had no intention of selling the company and called AI inference "a $1 trillion market in the making."[17]
Corsair received the AI Processor Innovation Award in the 2026 AI Breakthrough Awards, announced June 25, 2026.[23]
Competitive position
d-Matrix is one of several companies building silicon aimed only at inference, a group that includes Groq, Cerebras, SambaNova and Etched. What distinguishes d-Matrix commercially is its packaging choice rather than its performance claims: Corsair ships as an air-cooled PCIe card that goes into standard servers, so a buyer does not have to take a full rack-scale appliance or install liquid cooling to evaluate it.[16] Groq and Cerebras also lean on SRAM instead of HBM, a similarity CNBC noted in its June 2026 profile.[17]
The company's more recent positioning is explicitly not GPU replacement. Since 2026 its announcements have described Corsair and Raptor as parts in a heterogeneous rack where GPUs handle the compute-bound prefill phase of an inference request and d-Matrix silicon handles the latency-bound decode phase, which is how both the Parasail and Gimlet deployments are structured and what the NVLink Fusion collaboration formalizes.[18][13][20]
The architectural criticism most often raised against the approach is capacity. Rick Bahr, an adjunct professor of electrical engineering at Stanford, told CNBC that the downside of the SRAM-centric design is that it cannot handle very large reasoning models.[17] d-Matrix's own answer is Raptor, which replaces the off-package capacity tier with stacked DRAM bonded directly to the compute die.[11]
What d-Matrix does not disclose
d-Matrix is private and files no public financial statements. It has not published revenue, gross margin, unit shipments, or the identity of the hyperscale and frontier-lab customers it says have committed to Corsair. The $2 billion valuation is the company's own figure from the November 2025 Series C and has not been revalidated by a later priced round.[4][17] Performance claims such as "10x faster" and "3x better energy efficiency" are d-Matrix projections against specified GPU baselines on specific model configurations, and the company labels several of them preliminary on its own product pages.[21]
See also
- d-Matrix Corsair
- d-Matrix Raptor
- AI accelerator
- AI chip
- Inference
- Groq
- Cerebras Systems
- SambaNova Systems
- Etched
- NVLink Fusion
- NVIDIA MGX
- High Bandwidth Memory
References
- ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10"About d-Matrix." d-Matrix. Accessed September 15, 2026. d-matrix.ai/about
- ^1 ^2 ^3 ^4"d-Matrix Announces $44 Million in Funding to Build a One-of-a-Kind Compute Platform Targeted for At-Scale Transformer AI Datacenter Inference." d-Matrix. April 20, 2022. d-matrix.ai/...transformer-ai-datacenter-inference
- ^1 ^2 ^3 ^4"d-Matrix Announces $110 Million in Series B Funding to Make Generative AI Commercially Viable with First-of-Its-Kind Inference Compute Platform." d-Matrix. September 6, 2023. d-matrix.ai/...its-kind-inference-compute-platform
- ^1 ^2 ^3 ^4 ^5 ^6"d-Matrix Raises $275 Million to Power the Age of AI Inference." d-Matrix. November 12, 2025. d-matrix.ai/...on-to-power-the-age-of-ai-inference
- ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9Sid Sheth. "Transforming AI: d-Matrix's Pivotal Moments in Pursuit of Gen AI Inference At Scale." d-Matrix blog. June 26, 2024. d-matrix.ai/...ursuit-of-gen-ai-inference-at-scale
- ^1 ^2"d-Matrix Launches New Chiplet Connectivity Platform to Address Exploding Compute Demand for Generative AI." d-Matrix. January 24, 2023. d-matrix.ai/...ng-compute-demand-for-generative-ai
- ^"Generative AI Compute Front-Runner d-Matrix Reaches New Milestone in Efficient AI Inference." d-Matrix. August 22, 2023. d-matrix.ai/...milestone-in-efficient-ai-inference
- ^1 ^2 ^3"d-Matrix Unveils Corsair, the World's Most Efficient AI Computing Platform for Inference in Datacenters." d-Matrix. November 19, 2024. d-matrix.ai/...atform-for-inference-in-datacenters
- ^"d-Matrix Announces JetStream I/O Accelerators Enabling Ultra-Low Latency for AI Inference at Scale." d-Matrix. September 8, 2025. d-matrix.ai/...jetstream
- ^1 ^2"d-Matrix Announces SquadRack, Industry's First Rack-Scale Solution Purpose-Built for AI Inference at Datacenter Scale." d-Matrix. October 14, 2025. d-matrix.ai/...squadrack
- ^1 ^2 ^3 ^4 ^5 ^6 ^7"d-Matrix and Alchip Announce Collaboration on World's First 3D DRAM Solution to Supercharge AI Inference." d-Matrix. November 18, 2025. d-matrix.ai/...olution-to-supercharge-ai-inference
- ^1 ^2"d-Matrix and Andes Team on World's Highest Performing, Most Efficient Accelerator for AI Inference at Scale." d-Matrix. November 18, 2025. d-matrix.ai/...celerator-for-ai-inference-at-scale
- ^1 ^2"d-Matrix and Gimlet Labs to Deliver 10x Speed Ups, Massive Power Efficiency for Frontier AI Workloads." d-Matrix. March 12, 2026. d-matrix.ai/...gimlet
- ^1 ^2 ^3 ^4"d-Matrix Boosts Rack-scale AI Capabilities With Acquisition of GigaIO Data Center Business." d-Matrix. April 2, 2026. d-matrix.ai/...acquisition-of-gigaio
- ^Shane Snider. "'Inference Is Bigger Than Any One Chip': d-Matrix CEO on GigaIO Deal." Data Center Knowledge. April 3, 2026. datacenterknowledge.com/...rack-scale-ai-inference
- ^1 ^2 ^3 ^4 ^5 ^6"d-Matrix Corsair AI Inference Platform Enters Full Production to Meet Customer Demand." d-Matrix. June 9, 2026. d-matrix.ai/...-production-to-meet-customer-demand
- ^1 ^2 ^3 ^4 ^5 ^6Katie Tarasov. "Upstart chipmakers keep challenging Nvidia. This time it's Microsoft-backed D-Matrix." CNBC. June 9, 2026. cnbc.com/...dia-d-matrix-chip-production-microsoft
- ^1 ^2"Parasail to Combine NVIDIA AI Infrastructure with d-Matrix Accelerators to Achieve 10x Faster Token Generation." d-Matrix. July 7, 2026. d-matrix.ai/...parasail-d-matrix-accelerators
- ^1 ^2 ^3"d-Matrix Acquires Wallaroo.ai to Speed up Deployment of Heterogeneous AI Inference Workloads." d-Matrix. August 3, 2026. d-matrix.ai/...d-matrix-acquires-wallaroo
- ^1 ^2 ^3 ^4 ^5"d-Matrix Adopts NVIDIA NVLink Fusion Rackscale Infrastructure for Ultra-Low Latency AI Inference." d-Matrix. September 10, 2026. d-matrix.ai/...d-matrix-rackscale-nvidia
- ^1 ^2"Aviator Software." d-Matrix. Accessed September 15, 2026. d-matrix.ai/aviator
- ^"Introducing d-Matrix Demo Cloud: Ultra Low Latency Inference, Ready to Test in Minutes." d-Matrix blog. September 15, 2026. d-matrix.ai/...-inference-ready-to-test-in-minutes
- ^"d-Matrix Corsair Inference Accelerator Wins 2026 AI Breakthrough Award." d-Matrix. June 25, 2026. d-matrix.ai/...tor-wins-2026-ai-breakthrough-award
- ^"d-Matrix Unlocks New Potential with Reinforcement Learning Based Compiler for At Scale Digital In-Memory Compute Platforms." d-Matrix. November 15, 2022. d-matrix.ai/...digital-in-memory-compute-platforms
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation. Every suggestion is reviewed for sourcing before it goes live.
1 revision · v2 · 2,753 words · full history
Fact-checks are independent of edits: a reviewer re-verifies the article against its sources and stamps the date. How we verify
Research and drafting on this wiki are AI-assisted, under named human editorial standards. How AI is used here
Reviewer note: Independently fact-checked against 35 primary documents (152 claims) plus a resolution sweep of all 58 reference URLs. 16 defects found, 6 material, all corrected, including three uncited Vera Rubin architecture claims that NVIDIA documentation contradicts.
Cite this page: AI Wiki. "d-Matrix." aiwiki.ai, updated 15 Sept 2026, fact-checked 15 Sept 2026. CC BY 4.0. https://aiwiki.ai/wiki/d_matrix