d-Matrix

d-Matrix builds in-memory compute chips optimized for efficient transformer inference at scale in data centers. Raised $450M including $275M Series C in Nov 2025 at $2B valuation; Corsair chip in deployment. Founded by ex-Synopsys and SiMa.ai leadership.

d-Matrix's digital in-memory compute architecture targets 10x better inference efficiency than GPUs for large language models.

d-matrix.ai AI AcceleratorsUSAprivate

At a glance

Total raised

$450M

last round series c

Headcount

—

no public estimate on record

Headquarters

Santa Clara

USA

Founded

2019

7 years active

Status

Private

AI Accelerators

Latest valuation

$2B

Series C · 2025

For subscribers

Subscribe to see the full profile

Including Funding rounds, Bull / Bear thesis, Stock + earnings, Roster changes, Patents, News, and Open roles.

Founding

50% off

/month

91 of 100 spots left

Full

$10

/month

Available once all 100 Founding Member spots are claimed.

Get full access

Already subscribed? Sign in →

Funding history

Total raised

$450M

Rounds tracked

Latest stage

Series C

Latest valuation

$2B

Date	Stage	Amount	Valuation	Lead investors
Nov 12, 2025	Series C	$275M	$2B	BullhoundCapital, Triatomic Capital, Temasek
Sep 6, 2023	Series B	$110M	—	Temasek
Apr 20, 2022	Series A	$44M	—	Playground Global

How they make money

Hardware (one-time)Pattern

d-Matrix sells application-specific integrated circuits (ASICs) optimized for large-language-model inference in data centers. Their target buyers are hyperscalers, cloud-service providers, and enterprise data-center operators who need alternatives to Nvidia GPUs for running transformer-based models at scale. Revenue is recognized upon shipment of the Corsair chip to customers, a classic one-time hardware sale model with no recurring software or license component disclosed. The economics of a fabless chip company like d-Matrix are structured around non-recurring engineering (NRE) costs for design and tape-out, plus per-unit cost of goods sold covering wafer fabrication, packaging, and testing. Gross margins for merchant silicon in data-center markets typically sit well above 50% once volume ramps, but the model requires large upfront capital — the $160M raised to date funds exactly that. What scales is the unit margin on each additional chip sold after the fixed NRE is recouped. There is no indication of a subsidized element in this model; pricing is tied to the performance-per-watt advantage d-Matrix claims over incumbent GPUs.

Generated by deepseek-v4-flash · June 2, 2026Analytical summary · not investment advice

Founder backgrounds

Sid Sheth

Co-founder & CEO

Holds an M.S.E.E. from Purdue University. Spent over a decade at Intel in engineering and management roles across the Pentium microprocessor and network processor groups. Was a founding team member at Aeluros, a mixed-signal networking IC company, from 2001 until its 2007 acquisition by NetLogic (now Broadcom). Subsequently joined Inphi Corporation as SVP and GM of the Networking Business Unit, where he built a business focused on AI data center interconnects that became over $1 billion in revenue before Marvell's $10 billion acquisition of Inphi. Co-founded d-Matrix in 2019 to develop Digital In-Memory Compute (DIMC) architecture for generative AI inference.

Sudeep Bhoja

Co-founder & CTO

Holds an M.S. in computer engineering from Purdue University. Served as Chief Architect at Big Bear Networks from 2001 to 2005, then as Technical Director at Broadcom in the Infrastructure and Networking Group from 2006 to 2012, specializing in high-speed interconnects. Joined Inphi Corporation as CTO in 2012, where he built DSP technology for high-speed data center optical interconnects until 2019. Co-founded d-Matrix in 2019 and serves as CTO, leading the development of the company's in-memory compute architecture for energy-efficient large language model inference in cloud data centers.

AI bull / bear

Bull case

d-Matrix is well positioned for the coming inflection point in AI infrastructure. The market is saturated with chips optimized for training, but inference — where models actually get deployed and monetized — is where the volume will be. d-Matrix's digital in-memory compute architecture directly targets that bottleneck: delivering up to 10x better inference efficiency than GPUs for large language models without the precision trade-offs that plague analog in-memory approaches. That matters because data center operators care about total cost of ownership per token, not peak flops. The Corsair chip signals a real product, not just a paper architecture. The founding team — pulled from Synopsys and SiMa.ai — brings deep silicon and AI accelerator experience. The $160M Series B in 2023 shows serious VC conviction at a time when capital was tightening. If inference workloads scale the way everyone expects, d-Matrix's efficiency edge could make it a standard fixture in the next wave of AI infrastructure buildout.

Bear case

d-Matrix is chasing a market where NVIDIA already owns the installed base and software stack, and where well-funded competitors like AMD, Intel, Groq, Cerebras, and SambaNova are all fighting for the same inference dollars. Digital in-memory compute is still an unproven architectural bet at data center scale; the 10x efficiency claim is theoretical until it survives real-world deployment across the chaotic, rapidly evolving landscape of transformer model variants. The $160M raised through Series B is meaningful but modest compared to the billion-dollar war chests of rivals. Software moat is everything in this market, and building a compiler stack, runtime, and framework integrations that actually works for production inference is a multi-year grind. d-Matrix also faces the classic startup dilemma: customers want to see large-scale deployment before committing, but getting to deployment requires customers to commit. Without a hyperscaler anchor deal or a clear path to volume production, the company risks getting squeezed between NVIDIA's ecosystem lock-in and the commoditization pressures of the open-source inference stack.

Generated by deepseek-v4-flash · June 2, 2026Refreshes quarterly · see disclosures

In the news

d-Matrix · last 30 days

No articles ingested yet for d-Matrix. Once the hourly news pipeline is live, every article the classifier tags as mentioning this company appears here with its one-line AI summary and sentiment.

Watch

Leadership

Sid Sheth

Co-founder & CEO

Sudeep Bhoja

Co-founder & CTO

—

Products

Corsair
d-Matrix's flagship AI inference accelerator built on a digital in-memory compute (DIMC) architecture that performs matrix math inside memory to slash the data-movement energy that bottlenecks GPUs. A PCIe card packs multiple chiplets with large on-card SRAM, targeting fast, cost- and power-efficient inference for generative-AI and large-language-model serving at data-center scale. Corsair is the core embodiment of d-Matrix's in-memory-compute bet.
JetStream
A custom networking accelerator (NIC/interconnect) that scales d-Matrix inference across many Corsair cards and servers with low latency, letting customers grow from a single card to rack-scale deployments. JetStream complements the compute silicon by addressing the cluster-interconnect problem central to serving large models, rounding out d-Matrix's full-stack inference platform.
Aviator
d-Matrix's software stack—compiler, runtime, and model-serving tools—that maps standard AI models onto Corsair hardware and manages deployment, so developers can run popular transformer models without hand-tuning to the in-memory architecture. Aviator is essential to making d-Matrix's novel silicon usable, providing the framework integrations and optimization needed for production LLM inference.

Most recent patents

17 patents on file, but none with both an extractable figure and an abstract on Google Patents yet.