Y Combinator’s 2025 and 2026 infrastructure cohorts highlight how deeply infrastructure is being rebuilt around Artificial Intelligence workloads, data movement, and agentic applications. The list spans 228 companies, from established data engineering players like Fivetran to early stage teams building specialized stacks for inference, orchestration, and governance. Many of these startups position themselves as plumbing for Artificial Intelligence agents, offering APIs, observability, and runtime control so teams can move prototypes into production without rebuilding core systems from scratch. Others tackle long standing bottlenecks in cloud compute, storage, and networking, using new approaches like photonics or reinforcement learning to cut latency and costs while scaling across heterogeneous hardware.
A large subset focuses on making Artificial Intelligence agents safer, more compliant, and more reliable in live environments. Salus wraps agents with an API that inspects and can block incorrect actions at run time, while Truth Systems provides a compliance agent that translates policies into real-time guardrails for generative Artificial Intelligence usage. Chamber promises to let organizations run approximately 50% more workloads on existing GPUs by autonomously managing clusters and reallocating resources, and DeepAware targets GPU intensive data centers with an automation system that can slice energy waste by up to 30%. Janus automates Artificial Intelligence evaluations using high fidelity simulation environments, producing datasets that benchmark products and drive continuous post training improvement. Companies like Oximy and Nivara help enterprises understand how Artificial Intelligence is being used across teams, linking adoption and governance to measurable outcomes.
Another cluster targets core compute, storage, and data infrastructure tuned for Artificial Intelligence. Cumulus Labs operates a serverless GPU cloud with a proprietary inference engine that supports major large language and vision models, while SF Tensor’s Elastic Cloud platform automatically finds the cheapest GPUs across providers and can cut compute costs by up to 80%. S2.dev offers a serverless datastore for streaming data that treats streams like a web native resource, and PgDog provides sharding, load balancing, and pooling for PostgreSQL without application changes. Startups such as Kernel, Airweave, AgentMail, Modelence, Castari, and Metis are building the infrastructure needed for agents to access the internet, retrieve context, manage auth, and run in secure, autoscaling sandboxes. Others like Luel, Shofo, Mundo Artificial Intelligence, Shofo, and Mantis Biotechnology specialize in sourcing or structuring large, domain specific datasets, from rights cleared multimodal training corpora to human in computer models for biotech and sports. Together, these companies reflect a broad shift toward infrastructure that is explicitly designed around Artificial Intelligence centric workloads, from edge deployment and robotics to enterprise construction workflows and prediction markets.
