May 11, 2025

CB Insights Research AI 100: The most promising artificial intelligence startups of 2025 April 24, 2025 FREE DOWNLOAD: THE COMPLETE AI 100 LIST Get data on this year’s winners, including product focus, investors, key people, funding, and Mosaic scores. First name Phone number The AI space is evolving at an unprecedented rate. Since the start of 2024, thousands of new AI companies have formed, and funding to AI companies has surpassed $170B, primarily driven by titans like OpenAI and Anthropic. Given this momentum, the ecosystem is larger and more challenging to navigate than ever. Our annual AI 100 list is designed to cut through this noise and highlight the next wave of AI winners, with a focus on early-stage players that are showing strength in terms of market traction, investor quality, and talent. FREE DOWNLOAD: THE COMPLETE AI 100 LIST Get data on this year’s winners, including product focus, investors, key people, funding, and Mosaic scores. First name Phone number Leveraging CB Insights datasets such as deal activity, industry partnerships, team strength, investor strength, patent activity, and our proprietary Mosaic Scores , we selected 100 winners out of a cohort of 17K+ companies. We also analyzed CB Insights’ exclusive interviews with software buyers and dug into Analyst Briefings submitted directly to us by startups. Below, we map out the winners, categorizing them based on their core offering. Key trends and category definitions follow. Please click to enlarge. Data as of 4/23/25. Key takeaways on the AI 100 AI agents dominate the conversation. These applications, which automate tasks and processes for human users, are the next wave of genAI. Having made their way into virtually every horizontal and enterprise function, AI agents are also coming for infrastructure and verticalized applications . AI agents and supporting infrastructure make up 21% of this year’s companies, and the investors we spoke with consistently cited this space as a priority. ML security has become table stakes. The need to secure AI applications has grown in lockstep with the proliferation of genAI and agentic AI. 46% of strategy team leaders point to security as the primary barrier to genAI adoption, according to a recent CB Insights survey. Machine learning security companies are hardening AI algorithms and foundational models like LLMs, while also defending against increasingly sophisticated AI-powered attacks. AI observability and governance are critical gaps. Widespread use of AI is exposing the technology’s cracks — hallucinations, lack of orchestration, and output inaccuracies. It’s clear that AI ubiquity can’t exist without robust monitoring. Companies are rising to meet this need. Startups in this year’s list cover areas like observability and governance, while a small cohort also monitors AI agents to ensure reliability and compliance. The future is physical. Looking ahead, AI will evolve beyond software AI agents to a physical state. Advances in disparate areas of AI development — including robotics, multimodal image and voice models, edge computing, synthetic data, and spatial intelligence — provide the scaffolding for physical AI, which pairs AI software with hardware to take action in physical environments. Industrial humanoids represent an early manifestation of this, while future permutations could include fully autonomous defense drones, home companion robots, and more. Vertical applications are exploding. In 2024, the horizontal companies in this AI 100 cohort received more funding than their vertical and infrastructural counterparts — $1.6B compared to $1.2B each for infrastructure and vertical. But in 2025 so far, the funding picture looks very different: Vertical winners lead the way with $1.1B in funding raised. Category breakdown AI INFRASTRUCTURE On the foundation model front, infrastructure newcomers are rapidly releasing models that rival industry leaders, signaling a maturing market where technical excellence and novel approaches increasingly compete with raw computing power. We identified winners across large language, edge, reasoning, small language, and multimodal models. Meanwhile, as AI applications — particularly agents — become more autonomous and widespread, the need for robust monitoring, governance, and cybersecurity solutions has grown in lockstep. We’ve heard this in our conversations with AI investors, as well. Mozilla Ventures , a lead investor in Credo AI , views governance as a strategic imperative. Mohamed Nanabhay, Managing Partner, notes:  “…We think that the AI governance sector itself will take on a crucial role of creating value for enterprises, allowing companies that leverage governance to deploy AI faster through the reduction of risk with a greater competitive advantage as a result.” Category definitions: DATA Synthetic data: Artificially generated or altered information that mimics real-world data without privacy concerns. Aaru uses a multi-agent approach to create population simulations for predictive decision-making applications like consumer behavior and electoral modeling. Data preparation & curation: Tools and platforms that clean, transform, label, and organize data to make it suitable for AI training and deployment, encompassing data cleaning and specialized data processing. Unstructured , for instance, helps organizations capture unstructured data from various documents and convert it into AI-friendly formats such as JSON to train LLMs. Vector databases: Solutions that provide enterprises with an easy way to store, search, and index unstructured data at a speed, scale, and efficiency that current relational (and non-relational) databases cannot offer. For example, Qdrant provides an open-source vector database that allows developers to build production-ready applications that use nearest neighbor search functionality. DEVELOPMENT & TRAINING Foundation models: Pre-built AI algorithms and architectures that can be deployed, fine-tuned, or integrated into applications, spanning general-purpose foundation models and specialized domain-specific models. This category includes large language, edge, reasoning, small language, and multimodal AI models. For instance, Archetype AI ‘s Newton model processes multimodal sensor data and natural language to provide insights and predictions about physical environments. Agent building & orchestration: This category covers AI agent development platforms for building, orchestrating, and monitoring agents. Companies like LangChain provide a framework for building context-aware reasoning applications with tools for debugging, testing, and monitoring app performance across the entire application lifecycle. Computer vision & spatial intelligence: Technology that enables AI systems to understand, interpret, and interact with physical spaces and 3D environments, including mapping, navigation, and spatial data processing capabilities. Notably, World Labs develops Large World Models (LWMs) that enable AI systems to perceive, generate, and interact with both virtual and real 3D environments using spatial intelligence. OBSERVABILITY & EVALUATION AI observability platforms: These platforms monitor, measure, and assess AI model performance, reliability, and outputs, including tools for testing, benchmarking, and continuous improvement of AI systems. For instance, Arize ’s platform allows teams to monitor, diagnose, and improve the performance of AI models and applications in production through tools based on open-source standards that integrate with existing AI infrastructure. Governance: Solutions that establish policies, processes, and controls for responsible AI development and deployment, covering risk management, compliance, ethical oversight, and transparency requirements. For example, Credo AI offers a platform that automates AI oversight, risk management, and regulatory compliance while providing AI auditing to ensure system integrity and fairness. Machine learning security (MLSec): Technologies that protect AI systems from vulnerabilities, attacks, and data breaches, including techniques for securing model training, inference, and data pipelines. Solutions developed by companies like Zama enable computation on encrypted data, allowing for privacy-preserving machine learning across industries that require data privacy and security. ACCELERATED COMPUTING & HARDWARE Edge: Platforms that provide the infrastructure and models to operate AI on “edge” devices such as tablets, IoT, autonomous vehicles, or smartphones. For example, EdgeRunner AI constructs an ensemble of small, task-specific models that work together to solve complex problems locally on devices, ensuring data privacy and security for heavily regulated industries. Photonics: Solutions that use light (photons) instead of electrons for data processing, with the potential to significantly increase computing speeds. Companies in this category provide memory, interconnects, and system architecture. Xscape Photonics develops bandwidth-efficient photonics solutions to support AI/ML infrastructure. Quantum: Companies providing novel techniques like model compression and hardware to support quantum commercialization. Multiverse Computing provides AI model compression technology to enable quantum AI workloads and processing. Chips: Traditional chips, in addition to chips to support new AI technologies. Etched develops chips designed specifically for transformer inference, capable of processing extensive data for applications such as real-time voice agents and content generation. FREE DOWNLOAD: THE COMPLETE AI 100 LIST Get data on this year’s winners, including product focus, investors, key people, funding, and Mosaic scores. First name