38/622 → 79/739; +4.58 percentage points of formation share.
What counts as "infrastructure" in this batch has quietly split into two distinct businesses: plumbing for AI agents themselves (Amika's sandboxed coding agents, Klaus AI's cloud OpenClaw instances, Moda's continual learning layer, Chamber's autopilot monitoring) and plumbing for the data those agents are starved for (Sciloop's Olympiad-level math problems, Luel's rights-cleared training data, Shofo's video library indexing billions of clips). That second group exists because frontier labs have exhausted the easy internet-scraped data and are now paying for provenance and difficulty, which is also why Sciloop specifically recruits IMO and IPhO medalists rather than generic annotators. The formation's own 1.48x growth looks almost sluggish next to the categories feeding it, with Cloud Computing at 3.10x, Big Data at 3.60x, and AI Compute Infrastructure at 3.07x, suggesting the compute and storage layers underneath are scaling faster than the tooling layer built on top of them. That gap matters because agent-generated code is starting to arrive at a volume and size that outpaces human review, which is exactly the lineage problem now being investigated: once a coding agent like the ones Amika hosts can produce a large, multi-file change in minutes, teams need machine-readable tracing from prompt to deployment to rollback, not just a diff. This wave became possible specifically once agentic coding models matured enough in 2024 and 2025 to be trusted with multi-step, sandboxed execution rather than single autocomplete suggestions, which is what turned "AI infrastructure" from a model-serving problem into a full software-factory problem. Companies like Cumulus Labs and The Token Company are really symptoms of that shift, patching the observability and cost overhead that shows up only once agents are actually shipping production changes at scale.
As coding agents produce larger changes, teams may need machine-readable lineage from request and context through generated code, review, deployment, and rollback.
Study agent-heavy engineering teams to measure failures in change attribution, reviewer comprehension, dependency tracking, deployment evidence, and safe rollback.
Companies like Amika and SF Tensor are building sandboxed cloud infrastructure for running many AI coding/research agents in parallel, but as agent fleets scale, teams will need visibility into which agent runs cost what, failed why, and consumed which compute — a distinct layer above raw sandboxing.
Check back around May 2027 for new YC batches, to see whether independent evidence for "Agent-orchestration observability and cost accounting layer" has appeared.
Hyperspell explicitly frames itself as a 'Company Brain' synthesizing documents and conversations for agents that are otherwise 'clueless' about company context, implying current connector-based approaches are fragmented; this points to demand for a shared protocol or standard for keeping agent context fresh across tools.
Check back around August 2027 for new YC batches, to see whether independent evidence for "Context/memory synchronization standard across enterprise data connectors" has appeared.
s2.dev frames durable streams as a 'missing building block' in data systems; as agent-based infra companies (Amika, Brickanta, Hyperspell) proliferate and need reliable event/data flow between agents and tools, demand for durable-stream primitives like s2.dev should grow beyond generic developer tooling into an agent-infra necessity.
Check back around June 2027 for new YC batches, to see whether independent evidence for "Durable streaming/event infra becoming a required substrate for agent pipelines" has appeared.
Sciloop sells expert-authored math/physics problems sourced from IPhO/IMO medalists specifically because frontier models fail on them; this narrow, credential-verified data-sourcing model suggests a next layer of marketplaces for other hard-to-scale expert domains (law, medicine, engineering) feeding AI labs.
Check back around August 2027 for new YC batches, to see whether independent evidence for "Verified/adjudicated data marketplaces for niche domain expertise" has appeared.
RightNow AI is a research lab for GPU kernel optimization. Our platform RunInfra serves top open-source models optimized for AI agents at a fraction of the cost of closed-source APIs RightNow AI lab: https://rightnowai.co RunInfra platform: https://runinfra.ai.
* Ready to Use Instantly: Skip the complex server setup. It comes pre loaded with top AI Agents (like Claude Code and Hermes) so your agents can start working immediately. * Keep Your Work Safe: Unlike other systems that delete everything when they finish a task, these sandboxes remember your files, code, and settings forever until you delete them. * Pay Only for What You Use: Stop paying for idle servers. You pay a tiny fee to store your files, but you only pay for compute power during the exact minutes your AI is actively running. * Zero Infrastructure Headaches: You don’t need to be a senior DevOps engineer or manage complex cloud networks. One simple command handles the entire backend for you. * Connect to Any App: Allow your AI agents to securely log into and interact with everyday tools like Slack, Gmail, and GitHub in just a few clicks.
**Problem** Humans used to need laptops to do their work. Now they prompt agents instead, and those agents need computers too: always-on, in large quantities, and with unlimited capabilities. Building and maintaining that yourself is far from trivial. **Solution** Boat provides powerful, simple, permissive Ubuntu VMs in the cloud. In a sandbox, an agent can work for long periods on code or on knowledge work like legal, research, reporting, and auditing. **Why Boat and not any other cloud VM?** 1. It is ready out of the box: every sandbox ships with a desktop, chrome, languages, docker, snapshots, and templates, plus a CLI, SDK, API, and control pane for your fleet. 2. It is permissive and paradigm agnostic: any number of agents can use any number of sandboxes. Agents can run inside or use sandboxes as an external tool. 3. It is expertly crafted: we manage everything
Your agent worked fine yesterday. Today it loops, drops context, or calls a tool with the wrong arguments, and you find out from a user. Teams shipping AI agents change prompts, models, and tools constantly, and the same failures keep coming back in production: context loss, loops, broken tool calls, regressions after a refactor. Mirrors learns from your agent's production traces. It captures what worked and what failed in real sessions, then delivers the fix at runtime. Instead of waiting for someone to notice a regression and patch it by hand, the fix is applied automatically, so the same error never happens twice. Built for teams whose agents do things they can't undo, like issuing refunds, updating customer records, or sending outbound email.
Devin for AI Research Engineering
Hopper post-trains STT, TTS, and speech LLMs on your production calls, then serves and improves them against your live traffic
we build self-healing browser infrastructure that runs millions of agent operations a day without supervision. teams use it for large-scale data collection, agentic workflows, continuous monitoring, and anything else that requires acting on the web at volume.
We build infrastructure to power agentic systems such as software factories, multiplayer agent frameworks and agent coordination.
Login tells you who signed in. GrayPass verifies the expected person is still in control before consequential actions happen. We use cognitive biometrics like typing rhythm, cursor movement, scrolling, and response timing alongside session and device context to continuously establish assurance. Before actions like approving a payment, changing admin permissions, exporting sensitive data, or letting an AI agent act on someone’s behalf, GrayPass returns allow, challenge, hold, or deny and produces signed proof tied to that action.
Maritime is cloud infrastructure for AI agents. We help companies give every customer an agent running in its own isolated, persistent computer, starting at $1 per agent per month. Maritime handles hosting, scaling, and sleep/wake automatically, so developers can run thousands of agents without managing the infrastructure.
Measured-window members of Infrastructure that could not be assigned to a narrower shared YC tag or industry without forcing a classification.
A primary internal company segment within Infrastructure, derived deterministically from the shared yc tag “Developer Tools”. This is a classification of current window members, not a causal layer or a separate time trend.
A primary internal company segment within Infrastructure, derived deterministically from the shared yc tag “Data Engineering”. This is a classification of current window members, not a causal layer or a separate time trend.
A primary internal company segment within Infrastructure, derived deterministically from the shared yc tag “Deep Learning”. This is a classification of current window members, not a causal layer or a separate time trend.
A primary internal company segment within Infrastructure, derived deterministically from the shared yc tag “Robotics”. This is a classification of current window members, not a causal layer or a separate time trend.
A primary internal company segment within Infrastructure, derived deterministically from the shared yc tag “Open Source”. This is a classification of current window members, not a causal layer or a separate time trend.
These share a word with “Infrastructure” but are tracked as separate categories. The real overlap below is the actual share of Infrastructure’s companies also found in that category — not a guess.
Cambrian Signal, “Infrastructure,” publication yc-2026-10-02-semantic-v8, snapshot October 2, 2026.
cambrian.jafar.com/formation/infrastructure