Startups Directory

Cerebras Systems designs and manufactures wafer-scale AI processors (WSE-3) that are 57x larger than leading GPUs, enabling faster AI training and inference with dramatically reduced power consumption. The company offers both on-premise hardware systems and cloud-based AI services through its own data centers. Built by a team of veterans from SeaMicro and AMD, Cerebras solves the fundamental problem of data movement in AI workloads by keeping computation on-silicon, delivering measurable speed and efficiency advantages for large-scale machine learning.

Groq builds Language Processing Units (LPUs)—purpose-built AI inference chips that deliver up to 10x the throughput of comparable GPUs with ultra-low latency. The company offers GroqCloud, a fully managed inference platform with OpenAI-compatible APIs, and GroqRack for on-premises deployments. Founded by former Google TPU engineers, Groq has raised $1.75B and reached a $6.9B valuation, with a landmark $20B licensing deal with Nvidia announced in December 2025.

Ayar Labs develops TeraPHY, an industry-first optical I/O chiplet that replaces traditional electrical interconnects in semiconductor systems using silicon photonics. The solution delivers up to 1000x bandwidth density improvements and 1/10th the power consumption compared to conventional electrical I/O, enabling 8 Tbps of bandwidth at 10-nanosecond latencies. Built for AI accelerators, data centers, and distributed computing systems, TeraPHY eliminates GPU communication bottlenecks and solves the memory wall problem in high-performance computing.

d-Matrix develops specialized AI inference accelerator hardware for data centers, featuring the Corsair inference accelerator card, JetStream I/O networking accelerator, and SquadRack rack-scale system. The company uses digital in-memory compute architecture to deliver 3x better cost-performance and 10x faster token generation compared to traditional GPU-based inference solutions. d-Matrix targets the massive AI inference market—where most compute costs occur after model training—by optimizing for real-time latency, energy efficiency, and cost at scale.

Axelera AI develops purpose-built hardware and software platforms that accelerate AI inference at the edge using proprietary digital in-memory computing (D-IMC) technology. Their flagship Metis™ AI Platform combines CPU and memory into a single chip, eliminating energy-intensive data movement between cloud and device, delivering 10x better energy efficiency than standard computing. The company serves industrial, defense, robotics, and retail customers who need low-latency, privacy-preserving AI processing on devices rather than cloud.

Positron AI designs purpose-built FPGA-based hardware (Atlas) optimized for generative AI inference workloads, achieving 93% memory bandwidth utilization versus 10-30% in GPU systems while consuming less than a third of the power of Nvidia H100s. The company serves enterprises and cloud providers running transformer-based models at scale, from content moderation to token-as-a-service platforms. Their next-generation chip (Asimov/Titan) targets memory-intensive applications like video, trading, and multi-trillion parameter models with 6x more RAM than competing solutions.

Lightmatter develops photonic processors and interconnects that dramatically reduce energy consumption and latency in AI computing by replacing electrical connections with optical ones. The company's Envise chip combines traditional electronics with integrated photonic elements for AI acceleration, while its Passage interconnect technology enables data to move 100x faster between chips. They serve hyperscale cloud providers and AI data center operators, addressing the critical bottleneck of data movement energy consumption in modern AI workloads.

Astera Labs designs high-speed connectivity solutions for data center and AI infrastructure, enabling seamless integration of CXL, Ethernet, NVLink Fusion, PCIe, and UALink technologies through their COSMOS software suite. Their Intelligent Connectivity Platform addresses the critical bottleneck in AI clusters—connectivity—as GPU density scales from single nodes to thousand-GPU fabrics. The company offers a portfolio including Aries Smart Retimers, Taurus active electrical cable modules, Leo CXL Memory Controllers, and Scorpio X-series fabric switches, achieving 69% gross margins through a hybrid of premium hardware and high-margin software subscriptions.

FuriosaAI designs and manufactures high-performance, energy-efficient AI accelerators (NPUs) for data center deployments of large language models, multimodal applications, and computer vision workloads. Their flagship RNGD chip delivers 2.25x better inference performance per watt compared to traditional GPUs while consuming only 180W of power. The company serves enterprise customers and hyperscalers seeking cost-effective alternatives to Nvidia GPUs for AI inference at scale.

Hailo develops patented silicon AI accelerators that enable efficient deep learning inference on edge devices with minimal power consumption (<3W), eliminating cloud dependency while protecting privacy. The Hailo-8 and Hailo-10H chips integrate all required memory on the processor die, delivering 26+ TOPS of performance while maintaining exceptional power efficiency. Serving automotive, security, retail, industrial, and medical sectors, Hailo has secured over 300 customers including HP, Dell, BMW, Sony, and Bosch.

Enfabrica develops specialized networking chips and systems that solve critical data movement bottlenecks for AI and machine learning workloads at hyperscale. Their Accelerated Compute Fabric (ACF) SuperNIC chip delivers multi-terabit Ethernet connectivity and memory disaggregation, enabling GPUs, CPUs, and memory to communicate 4x faster than competing solutions. Built by former Broadcom and Google infrastructure leaders, the company bridges PCIe/CXL memory semantics with RDMA networking to reduce AI processing costs and improve efficiency.

Recogni develops AI inference processors that deliver 10x higher compute density, 10x lower power consumption, and 13x lower cost per query compared to GPU-based solutions. The company applies proprietary logarithmic number system (LNS) technology to enable efficient inference at scale for generative AI and autonomous vehicle applications. Recogni targets cloud data centers and automotive OEMs seeking to reduce the power, cooling, and infrastructure costs of AI deployment.

EnCharge AI develops analog in-memory computing hardware that performs AI computations directly within memory cells, eliminating energy-intensive data movement between processing and storage. The company's EN100 accelerator delivers 200+ TOPS of AI compute in an 8.25W power envelope for laptops and workstations, achieving up to 20x better performance per watt than competing solutions. Available in M.2 and PCIe form factors, EN100 brings advanced AI capabilities to edge devices and client platforms with unprecedented efficiency.