Infrastructure Startups funded by Y Combinator (YC) 2026

Fivetran automates data movement out of, into and across cloud data platforms. We automate the most time-consuming parts of the ELT process from extracts to schema drift handling to transformations, so data engineers can focus on higher-impact projects with total pipeline peace of mind. With 99.9% uptime and self-healing pipelines, Fivetran enables hundreds of leading brands across the globe, including Autodesk, Conagra Brands, JetBlue, Lionsgate, Morgan Stanley, and Ziff Davis, to accelerate data-driven decisions and drive business growth. Fivetran is headquartered in Oakland, California, with offices around the world. 

Context.dev provides AI agents and software products with realtime web context at scale through a single API layer. Instead of stitching together multiple providers for crawling, markdown extraction, brand data, images, sitemaps, style guides, enrichment, and other web data needs, teams can use Context.dev to turn websites, companies, brands, and people into clean, structured, token-optimized context. We help developers build smarter agents, personalization workflows, sales and marketing tools, design automation, and B2B SaaS products that depend on accurate, fresh web data.

Miso Labs is building the world’s most emotive foundation models for voice. We believe that the next generation of AI interactions shouldn't just be functional—they should be human. By bringing warmth and lightning-fast speed to the voice layer, we empower developers to build voice agents that users truly love.

RightNow AI is a research lab building GPU infrastructure that lets teams own their AI stack instead of depending on closed-source APIs. Our platform RunInfra takes any HuggingFace model, auto-generates optimized GPU kernels, and deploys it serverlessly with pay-per-token pricing RightNow AI lab: https://rightnowai.co RunInfra platform: https://runinfra.ai.

Coding agents in the cloud don't work for majority of teams who require native capabilities; they lack XCode for iOS, Emulator for Android, Unity for game development teams. We are on a mission to cloudify every local development utility so that regardless of where it runs, your agent can compose an ideal set of capabilities on the fly so you don't need to run it on your laptop. Today, we provide remote services such as XCode, iOS & Android simulators to enable coding agents running in any sandbox to be able to build and iterate over mobile apps, too. Replit, Rork, Momentic, Minitap and more agent companies have built mobile development and testing experiences on Limrun platform. Coconote (Quizlet), SorceJobs (W25) and more companies with mobile apps use Limrun with their cloud agents, merging PRs without having to check out the code locally.

Wato gives AI agents shared memory, reusable workflows, connected tools, and living artifacts so work compounds across a team instead of disappearing into chat threads. Teams use Wato to give every agent access to trusted company context and turn useful runs into knowledge the next agent can build on.

The $2T global telecom industry was built for humans. Agents can't act in the real world today because without a phone number, they lack identity. AgentPhone gives every AI agent its own phone number to reach people and businesses through messages and calls.

Expanse unlocks wasted GPU capacity. We recover idle compute through three capabilities: resource prediction (right-sizing job submissions before they reach the scheduler), optimisation suggestions (code and config changes researchers can apply themselves), and failure prediction (catching jobs that will fail before they consume hours of GPU time). We’re four engineers. We ran HPC and GPU training workloads at the largest quant funds and national supercomputing centres. We faced this problem first hand and the only fix was to over-provision and burn millions. Ismaeel built the first multimodal HPC resource predictor as research at EPCC (Edinburgh’s Parallel Computing Centre), which beat every published baseline. This is the tool we wish we had.

We build distributed compute clusters with the cheapest CPUs and GPUs across Hyperscalers and Neoclouds for AI. Our mission is to bring frontier-grade infrastructure to everyone. We're starting by building large scale high performance computing (HPC) clusters for quantitative trading firms to run parallel simulation workloads such as backtesting. Our technology generalizes to critical AI workloads such as post-training with reinforcement learning, fine-tuning, long-horizon agents with high tool use, and batch inference.

Kugel is an enterprise-ready TTS model, available on-prem, with a focus on 25+ languages and low latency.

dripos sends first messages, handles replies, follows up, and books calls across iMessage, LinkedIn, Gmail, Slack, and group chats. You choose who it can message and when it can act.

Archal lets you test agents and code that touch third-party APIs without hitting the real services. We build working clones of popular SaaS platforms that hold state across requests and behave like the originals, so you catch bugs before your agent does something irreversible in production.

Inth makes privacy compliance live where the risk actually is: the codebase and runtime. Companies still rely on dashboards, questionnaires, and policy docs to prove compliance, but consent breaks in code, deletion breaks in code, and vendors, logs, AI systems, and agent-written changes all move user data in code. Inth sits in the repo, detects user-data privacy risk, maps it to files and owners, and generates the evidence regulators and enterprise buyers ask for. Its wedge is c15t, an open-source consent SDK with 2.6M npm downloads and adoption from teams like Zed, Infisical, and Expo. Inth is making every codebase compliant by default.

Clone any postgres DB regardless of size in 6s so agents can test their work

Smol machines solve the problem of "but app works on my machine" by shipping the machine. Smol machine enables devs to ship a lightweight virtual machine with any software, similar to how Electron ships an entire browser with the web app.

We are making the browser layer for your AI agents. Agents don’t need to see UIs like us humans do. We give agents a semantic understanding of the web instead of brittle visual interfaces. Your agents can perform deep research, large-scale scraping, and complex automation—more efficiently with our knowledge graphs that turns the web into a native protocol for LLMs.

Interfaze is an AI model built on a new architecture that merges specialized DNN/CNN models with LLMs for developer tasks that require deterministic output and high consistency like OCR, Object detection, classification, Speech-to-text (STT) and more. Interfaze has outperforms large labs across 9 frontier benchmarks like OCRBench V2, RedCOCO and more. Full benchmarks: https://interfaze.ai/leaderboards Read paper: https://arxiv.org/abs/2602.04101

The AI-native observability platform for fast-moving engineering teams. Backed by tastemakers from the world's top AI companies: Vercel, Graphite, Daytona, Browserbase, LangChain, Mastra, Replit, and more.

Every computer ever made forces you to do one thing at a time. One cursor. One keyboard. One Hardware. One app in focus. Fifty years. Nobody questioned it. We did. We started over. We built InfinityOS, a fundamentally new computer OS on the cloud from ground up. Infinity runs every app in its own independent computer -- own GPU, own input, own environment. Windows and Linux in the same session. Humans and agents working side by side. No concurrency ceiling. Cold start in seconds. runs on the browser of any device..

Building the communication infrastructure needed for fully autonomous agents

Minicor connects AI startups to legacy desktop apps that have no API. Reliably, quickly and at scale. Describe a workflow and get an API endpoint. Each call triggers a self-healing desktop automation on a Windows VM. On-prem, cloud, or Citrix. Build through your coding agent via our MCP. Complete with the observability tools to manage your customer relationship: full video replay of every automation, granular logging, and slack enabled alerts.

Compression middleware that removes context bloat in milliseconds, lowering costs and improving end-to-end latency. Compression is especially effective across natural language workloads. In a blind LLM arena case study with one of our customers, compressed requests increased user preference, lowered costs, and lifted purchase volume by 5%.

AI SRE agents that automatically learn each customer’s system so they work just like an in-house engineer.

Terminal Use is an orchestration platform for background agents. Purpose-built for agents that use filesystems. Our platform is CLI-first, making it easy for your coding agents to experiment and continuously improve your deployed agents.

Piris Labs offers a full-stack inference service that eliminates the AI data movement bottleneck. By pairing proprietary photonic hardware with a vertically optimized software stack, we minimize the "memory wall" associated with expensive GPUs. This allows us to deliver the same performance as traditional clusters at a fraction of the cost. Our technology improves effective FLOP utilization and reduces latency, finally making the unit economics of trillion-parameter models sustainable. Founded by a team of MIT physicists and Meta AI experts.

Cascade helps companies take their proprietary data and workflows to align models for their use cases, resulting in higher throughput intelligence that works better, faster, and cheaper.

Wayco is building a unified data intelligence layer for medical treatment of legal cases. From the first intake to instant settlement.

We enable AI voice agents to collect payments over the phone through a single API call, handling card processing and PCI compliance across all payment gateways

Your agent processed a refund without looking up the order ID, costing you thousands. You only found out three hours later from a support ticket. Evals, output scoring, and observability can reduce the likelihood of mistakes like these occurring - but there's no solution that inspects and prevents an action as it’s about to execute. Salus does that. We’ve built an API that wraps around your agent and checks its actions at run time, blocking incorrect ones and providing immediate feedback to guide retries. Kevin and Vedant were roommates at Stanford, where they both studied computer science.

Autumn is building the first real-time signal intelligence platform for GTM teams. We monitor posts, commits, blogs, and announcements, surfacing buying signals the moment they appear. Define your ICP and the signals that matter, and we deliver a condensed, real-time feed filtered by intents.

Get your personal OpenClaw instance on the cloud and connected to your Slack or other messaging app in 3 minutes. Pre-installed security features and common skills.

Oximy helps enterprises understand and manage AI usage across their workforce. Used by security, finance, and transformation teams to track adoption, control spend, and manage governance as AI scales.

APIs weren't built for agents. Orthogonal fixes that. We enable agentic payments so any agent can instantly discover, access, and pay for hundreds of APIs. No subscriptions, no sales calls, just pay per use. API providers list once and get paid every time an agent calls their API.

We build intelligence for cultural institutions. Our first commercial product is for the literature industry with a focus on school and public libraries. It is a AI-native OS for libraries, that helps librarians save time on tedious tasks and lets them focus on their patrons.

Moda is the reliability & monitoring layer for agents. It surfaces patterns across agent hallucinations, laziness, forgetfulness, and tool call failures. While also providing analytics on user frustration, net promoter score, retention, and churn rates. In the next 10 years, with more AI agents spawning and trying to get PMF, developing an agent that actually works for users is critical. Everyone has access to the same models, thus the ones that win on a product level will be the ones who learn fastest from real usage. Moda is that learning layer.

Compresr provides an API that compresses LLM context without losing what matters. It’s a drop-in for agents and RAG that cuts token costs and improves accuracy.

We are building the world’s largest video library. We've aggregated billions of videos into a searchable index and use agents to find and label the exact datasets a lab needs on demand. If a lab needs 100K hours of cooking videos where someone is holding a pan, with reasoning annotations on top, our agents search the index, extract the matching subset, route it through our labeling pipeline, and deliver a custom dataset in days, not months.

Cumulus Labs lets engineering teams ship AI in production without needing a dedicated ML platform team. Right now, companies building AI products are forced to stitch together separate vendors for routing, observability, evaluation, fine-tuning, and inference. This fragmented approach is brittle, expensive, and is a common reason enterprises fail with AI. We replace that entire stack with a single unified platform. Developers can keep their existing code while instantly upgrading to a unified platform that handles routing, semantic caching, continuous shadow evaluation, simulated data, and one-click fine-tuning. Behind the platform is Ion, our proprietary inference engine running on a custom NVIDIA Grace GPU fleet. Ion uses in-house custom GPU kernels to deliver 30 to 50 percent more throughput than standard vLLM or SGLang, giving our customers SOTA inference economics.

Enabling physical AI at scale.

Chamber puts your AI infrastructure on autopilot, and saves your machine learning engineers hours of manual effort. Our agents continually monitor, detects failures, provide root-cause analysis, resolve issues, and optimize AI workloads and scale across clouds. It operates like an autonomous infrastructure team, helping save your research and engineers hours each day, and debugging workload performance issues. Your ML teams move faster, infra waste drops, and GPU bottlenecks disappear.

Edge AI is inevitable, but shipping it is painful: every device class behaves differently, runtimes vary, models are huge, and performance collapses under memory/power constraints. RunAnywhere turns that into an enterprise-ready workflow: one SDK to run models on-device, plus a control plane to manage models, enforce policies, and measure outcomes across thousands of devices.

Mantis is a digital twin company that combines LLMs with high-fidelity physics simulations in order to convert rare and difficult-to-access human behavior data into predictive models.

One API for KYC, KYB, AML, biometrics, and fraud — powered by 1,000+ data sources, 200+ signals, and in-house AI models. Global coverage. Integrate in minutes, pay as you go, 500 free checks/month. Start free at business.didit.me.

Luel is a sourcing and licensing platform for rights-cleared multimodal training data at scale. We work with frontier AI teams to provide high-quality bespoke data collections and off-the-shelf datasets.

Captain delivers the most accurate file search engine built for AI agents. We’ll index data from the sources folks already use like S3, SharePoint, and Google Drive, and easily scale multimodal, petabyte-level content search. We’re the Snowflake for Unstructured Data. Captain tops the Open-RAG-Benchmark with over 20% higher accuracy than standard RAG pipelines. We achieve this through robust data processing techniques like embedding normalization across modalities, ensuring that representations cluster by semantic content rather than data type.

Sciloop creates expert-level math and physics problems that frontier AI models can't solve, then sells the data to AI labs for training and evaluation. Our problems are created by IPhO and IMO medalists — the top 0.01% of STEM talent globally. On our benchmark, models like GPT 5.4 Pro and Gemini 3.1 Pro score 0-5% on our hardest problems. We work with AI labs to supply continuous, fresh training data that pushes the frontier of mathematical and scientific reasoning. Founded by Bilal and Osman, International Physics Olympiad medalists from MIT with hands-on ML research experience at MIT CSAIL.

Hyperspell is the Memory & Context Layer for AI Agents. AI agents are clueless geniuses. They crush humans on any standardized test, but wouldn't last a day at a real job. What today's super-intelligent agents are missing is the real world context they are operating in. A context that humans stitch together from hundreds of data points across dozens of interactions and channels. A context that grows with their tasks. Hyperspell gives AI agents this context by connecting to their user’s workspace data and building a personalized memory and context layer.

AI researchers should be pushing the boundaries of what's possible with new architectures and training methods. Instead, they waste weeks configuring cloud infrastructure, debugging distributed systems, and optimizing their GPU code. We know because we lived it: While training our own models across thousands of GPUs earlier this year, we spent more time fighting our infrastructure than doing actual research. That's why we're building two things. First, Elastic Cloud: a managed platform that automatically finds the cheapest GPUs across all providers, handles spot instance preemption, and cuts compute costs by up to 80%. Second, automatic kernel optimization that makes training code run faster by modeling hardware topology, often beating hand-tuned implementations. The problem is that getting high performance across different hardware is genuinely hard. NVIDIA's CUDA moat exists because writing fast kernels requires deep expertise. Most teams either accept vendor lock-in or hire expensive kernel engineers. Our goal is to break the CUDA moat. The compute bottleneck is the biggest constraint on AI progress. NVIDIA can't manufacture enough GPUs, and their monopoly keeps prices astronomical. Meanwhile, AMD, Google, and Amazon are shipping capable alternative hardware that nobody uses because the software is too hard. We're breaking that moat. If we succeed, anyone will be able to train state-of-the-art models without thinking past their PyTorch code.