Key Points:
- Silicon Valley semiconductor startup d-Matrix has reached a $1.55 billion valuation in a new funding round.
- The company’s flagship Corsair inference accelerator has entered full volume production to meet surging data center demand.
- The Corsair chip utilizes a digital in-memory computing architecture that bypasses the high-cost memory bottlenecks of traditional GPUs.
- By-passing high-density DRAM for on-chip SRAM, the processor claims 10 times faster performance while using 5 times less energy.
The high-stakes battle to supply the physical hardware powering the global artificial intelligence boom has reached a major milestone. In a significant development for the semiconductor industry, California-based AI chip startup d-Matrix has reached a $1.55 billion valuation in a new funding round. The Wall Street Journal reported that the massive capital injection will fuel the commercial rollout of the company’s proprietary “Corsair” chip, which has officially entered full-scale volume production. By offering an innovative, non-GPU architecture designed specifically to run large language models and autonomous AI agents at a fraction of current operating costs, d-Matrix is emerging as one of the most formidable independent challengers to Nvidia’s multi-trillion-dollar monopoly.
Founded in 2019 by Indian-American semiconductor veterans Sid Sheth and Sudeep Bhoja, d-Matrix has spent the past seven years quietly building the foundational software and silicon required to solve the industry’s most pressing physical bottleneck. While Nvidia’s graphics processing units (GPUs) remain the undisputed champions for training massive foundation models, running these models in daily production—a process known as inference—demands a completely different architectural focus. Sheth, who previously scaled several multi-billion-dollar chip businesses, realized early on that general-purpose graphics chips are highly inefficient for inference workloads, prompting him to design a clean-sheet accelerator from the ground up.
The primary engineering challenge limiting the efficiency of modern AI inference is a physical barrier known as the “memory wall.” When running large language models, processors spend far more time and energy moving data back and forth between the computing core and external memory pools than they do actually performing mathematical calculations. To eliminate this data-movement bottleneck, d-Matrix has pioneered a digital “in-memory computing” (DIMC) architecture. Instead of shuttling data across the chip, the computational math happens directly where the data already lives, completely bypassing the memory wall.
This in-memory computing approach has enabled d-Matrix to make a bold, highly strategic design decision: to completely skip high-cost High-Bandwidth Memory (HBM3e/HBM4) and traditional DRAM in favor of on-chip Static Random-Access Memory (SRAM). This choice is highly significant because the global semiconductor supply chain is currently facing a severe memory crisis. Fueled by intense, competitive buying from cloud giants, Nvidia’s memory procurement costs have soared by 485%, making advanced HBM memory account for 25% of the total $7.8 million build cost of its latest supercomputers. By skipping DRAM in favor of SRAM, d-Matrix completely bypasses this massive cost and supply trap.
The resulting technical specifications of the Corsair PCIe card are truly remarkable. Each single-slot card packs 6,400 square millimeters of silicon, containing 4 gigabytes (GB) of what d-Matrix calls “Performance Memory” (SRAM), delivering an astronomical bandwidth of 300 terabytes per second (TB/s). For larger workloads, a dual-card configuration links 16 chiplets through high-speed, all-to-all interconnects, delivering up to 19.2 PFLOPs of mathematical performance and supporting up to 512 GB of off-chip memory capacity. This highly integrated, compute-dense design allows enterprise customers to run massive 100-billion-parameter models incredibly fast inside a single server rack.
These advanced specifications translate into massive, peer-beating efficiency gains in real-world testing. d-Matrix claims that when paired with existing data center infrastructure, its Corsair accelerator runs generative AI inference tasks up to 10 times faster than standalone Nvidia GPUs, while consuming up to 5 times less energy and reducing operating costs by a factor of 3. This step-change in performance and efficiency directly addresses the massive electricity and carbon-emission crises currently plaguing the data center industry, allowing a single server farm to handle the processing workload of ten.
This impressive technological performance has attracted substantial, highly strategic capital from global tech giants and investment groups. To date, d-Matrix has raised a total of $500 million in private funding, including a massive $275 million Series C round led by Singapore’s Temasek and including participation from Microsoft’s venture fund, M12. While the company declined to name its specific priority clients, Sheth revealed that several prominent cloud providers, sovereign AI infrastructure operators, and frontier AI laboratories have already committed to purchasing the Corsair cards, with 90% of current shipments heading to U.S.-based customers.
To support its transition into full-scale commercial volume production, d-Matrix has secured long-term supply and fabrication agreements with major semiconductor partners. The startup is collaborating with manufacturing giant TSMC to package its chiplets and partnering with Broadcom to integrate its high-speed connectivity solutions, ensuring its hardware can scale seamlessly in data centers. Furthermore, the company is already designing its next-generation “Raptor” processor, scheduled to launch next year, which will use advanced 3D DRAM packaging to support even larger reasoning and video-generation models.
The ongoing re-pricing of AI chip stocks highlights the immense, high-stakes nature of the global hardware race. As cloud giants collectively spend over $100 billion annually on data centers, they are actively looking to diversify their hardware suppliers to reduce their expensive dependency on a single GPU manufacturer. Even a minor 1.5% delay in chip procurement or regulatory approvals can cost companies millions in wasted factory overhead, prompting developers to explore alternative, more efficient systems. By securing its own manufacturing contracts, d-Matrix is proving that independent startups can successfully build their own domestic supply chains, easily exceeding the $1 billion funding rounds of traditional technology startups.
Ultimately, d-Matrix’s successful full-production launch and $1.55 billion valuation mark a vital turning point for the semiconductor industry. By proving that digital in-memory computing and custom chiplet packaging can successfully bypass the memory wall, the young startup is showing the rest of the tech world how to build sustainable, high-performance computing systems. As the first Corsair cards begin shipping to data centers over the coming weeks, this landmark hardware release proves that the next phase of the artificial intelligence revolution will not be dictated by raw GPU power alone, but will successfully rely on the perfect, highly efficient integration of processing and memory.
