Picks, Shovels, and Blast Radius: YC Is All-In on the Agent Infrastructure Layer

The wave right now

The 2023–2026 infrastructure cohort is not building tools for training models. It is building for one moment: the day an agent touches production. Three forcing functions define the wave — GPU economics that make every token a budget line, agents that break every assumption classical DevOps was built on, and enterprises that suddenly care where inference physically runs. YC has responded by funding the picks-and-shovels layer at a density that looks less like portfolio diversification and more like a thesis. If you're building infra this quarter, the question isn't whether the wave is real — it's which sub-layer is already a knife fight and which still has room.

The landscape today

The compute squeeze is spawning an efficiency layer

When the scarce input is GPUs, intelligence *about* GPUs becomes a product. Expanse reads cluster telemetry to predict job fit and unlock capacity you already paid for; Zibra Labs aggregates the cheapest CPUs and GPUs across hyperscalers and neoclouds into one scheduling engine; RightNow goes lower still, generating kernels and optimized runtimes for self-hosted open-source inference. The contrarian end of this bet is Piris Labs, wagering that photonic hardware — not better software — breaks the cost curve. My read: the software plays here only work if the gains are *measured and material*; "we make GPUs cheaper" is now a claim every buyer has heard four times.

F26

RightNow

Enabling Model-Hardware Co-Design at Scale

ActiveGPU Inference Optimization2 ppl

P26

Expanse

Unlock wasted GPU capacity.

ActiveGPU Efficiency & Cost4 ppl

P26

Zibra Labs

Distributed Compute for AI

ActiveDistributed HPC Compute2 ppl

W26

Piris Labs

Inference at Light Speed

ActivePhotonic AI Inference Hardware4 ppl

Agents need somewhere to act that isn't production

The single sharpest pattern in this cohort: isolation as a product. Agents are write-capable, tireless, and occasionally wrong — so the blast radius has to be engineered down to zero. Ardent clones any-size Postgres databases in seconds so agents can migrate and test on branches; Indexable forks entire environments — files, processes, memory, database — into snapshots; Limrun extends the idea to native mobile, putting Xcode and Android simulators inside the agent's sandbox. This is the most defensible corner of the wave because the technical bar (stateful, fast, full-fidelity cloning) is genuinely high and the switching costs once wired into CI are real.

P26

Indexable

sandbox infrastructure for AI agents

ActiveAgent Sandbox & Testing3 ppl

P26

Ardent

Database sandboxes for Agents

ActiveDatabase Cloning & Sandboxing2 ppl

P26

Limrun

Extend your cloud agent with XCode, Android and more

ActiveCloud Development Environment Extensions1 ppl

Trust is the new uptime

Nobody ships an agent they can't audit. Archal runs eval scenarios against stateful clones of real SaaS services and fails the build when behavior regresses — evals as a CI gate, not a dashboard. Sazabi rebuilds observability around agent executions, with conversational debugging instead of Grafana archaeology. And Codag attacks the unglamorous bottleneck: compressing torrents of infrastructure logs into schema-valid capsules an agent can actually reason over. The bet underneath all three is the same — agent output is now a production artifact, and production artifacts get regression-tested.

S26

Archal

The eval platform for autonomous software

ActiveAgent API Testing Platform2 ppl

P26

Sazabi

The AI-native observability platform for fast-moving engineering…

ActiveAI-Native Observability Platform10 ppl

S26

Codag

Systems log compression for agents.

ActiveLog Compression Infrastructure1 ppl

Agents are getting inboxes, phone numbers, and shared memory

The most speculative sub-pattern is the standards land-grab: giving agents real-world identity and coordination primitives. primitive is email infrastructure for agents; AgentPhone issues them dedicated phone numbers behind a unified webhook; Wato gives teams of agents shared memory and reusable workflows so work compounds instead of evaporating in chat threads. Highest ceiling in the theme — whoever owns the agent identity layer owns a tollbooth — but it depends entirely on a multi-agent future arriving on schedule.

P26

primitive

Communication for agents

ActiveAgent Communication Infrastructure4 ppl

P26

AgentPhone

Phone Numbers for AI Agents

ActiveAgent Communication Protocol2 ppl

P26

Wato

Shared memory, tools, and workflows for agents across teams

ActiveAgent Workflow Platform2 ppl

The cohort signal

This is a deliberate program bet, and the batch math proves it. Winter 2026 seeded the theme — Compresr, Piris Labs, Terminal Use and peers. Then Spring 2026 nearly tripled the density: roughly two-thirds of the current-cohort companies in this memo come from that single batch. Summer and Fall 2026 show fewer entries but *narrower* ones — Codag's log compression, Archal's CI-integrated evals — which is what a maturing theme looks like: the land-grab batch has happened, and later batches are funding wedges, not platforms.

One more tell: an unusually large share of this cohort arrived via rename or pivot — Archal, Indexable, Zibra Labs, Compresr among them. Founders who started elsewhere are steering *into* agent infrastructure mid-flight. That's conviction flowing toward the theme, but it also means the obvious ideas are being arrived at simultaneously from multiple directions.

Lessons from the last cycle

YC's prior infra generation tells you how this movie usually ends: acquisition by the platform above you. Heroku defined developer experience for a decade and was acquired; Firebase became Google's app platform; CoreOS was absorbed into Red Hat OpenShift; Cloudant went to IBM back in 2014. The exception is instructive: Fivetran is still independent and still the leader — because it picked a boring, durable job (moving data) that outlives any single platform shift. The lesson for this wave: DX premium is real, but margin accrues to whoever owns the compute underneath you. Pick a workload that survives the next model generation, or price in the acquisition.

If you're building here

Three openings worth your quarter:

• Verticalized agent evals. Archal proves the CI-gated eval pattern, but it's one company against an enormous surface — voice agents, browser agents, regulated industries. You'd have to believe enterprises won't ship agents without regression gates. (They won't.)
• Sovereign inference. KugelAudio's EU-hosted on-prem TTS and RightNow's self-hosted endpoints both point at the same buyer: the compliance-bound enterprise that cannot send tokens to a US API. You'd have to believe open-weight models stay good enough and that sovereignty budgets keep growing. Both look safe.
• The token P&L. Compresr and Codag treat context as a cost line. The opening is upstream: whoever gives a CFO a single pane for AI spend across providers owns the renewal conversation.

The tarpits, by name: generic agent sandboxes and browser layers — StableBrowse, Indexable, and smol machines are already converging on adjacent territory, and the frontier labs build inward from here. GPU cost optimization without proprietary workload data — Expanse, Zibra Labs, and RightNow make this a four-way race before you arrive. And "the OS for agents" platform plays like ProjectX: platforms get built *after* workloads exist, not before.

The meta-belief for the whole theme: agents reach production at enterprise scale within 18 months. If you believe that, the trust and isolation layers are the safest ground in startups right now. If you don't, almost nothing here survives contact.