a16z and YC's Views on AI-Era Infrastructure
I recently went through a16z's Big Ideas 2026, the Jennifer Li interview, and YC's Requests for Startups 2025–2026, focusing on their discussion of the infrastructure layer (databases, storage, messaging, middleware, compute). Worth sharing.
Core Thesis: Design Assumptions Changed, Not Feature Gaps
In Big Ideas 2026, a16z argues that existing enterprise backends were designed for a 1:1 human-to-machine model — one person triggers one request, the system responds once. That assumption breaks in the Agent era. A single Agent's goal might trigger 5,000 subtasks, database queries, and recursive internal API call fan-outs in milliseconds. To traditional databases and rate limiters, an Agent's traffic pattern is indistinguishable from a DDoS.
Their conclusion: the control plane needs to be redesigned. "Agent-native infrastructure" will emerge. The next generation of systems must treat thundering herd as the default state, not an exception.
They argue this isn't a matter of existing systems lacking features — the design assumptions are simply wrong for Agent scenarios. Concurrency models, access subjects, transaction boundaries, and traffic patterns have all changed.
Jennifer Li, who leads a16z's infra team, puts it this way:
"AI infrastructure is being rebuilt from scratch, not just upgraded. Everything running AI workloads today — storage, compute, orchestration, memory — was built for a different era. The entire stack is being rewired in real time."
This covers four components: storage, compute, orchestration, and memory. Memory is listed separately — it's a new addition in the AI era that doesn't exist as a concept in traditional infra.
a16z's Investment Framework
One interesting data point: a16z allocated $1.7B of its new $15B fund to the infra team — the largest single allocation of any area. Jennifer Li's investment framework covers the entire chain from chip design to the developer software stack.
Compute layer: GPU clusters, inference optimization, hardware architecture. a16z sees opportunity here, but what they're looking for isn't companies selling compute capacity — they're looking for companies that "control the decision layer": software that decides how compute is allocated, how models are deployed, and how costs are optimized. These control points become increasingly valuable as AI systems scale, analogous to how the orchestration layer worked during the cloud era.
Storage layer: AI workloads require storage with different design assumptions from traditional object storage — closer to compute, supporting unstructured and multimodal data, supporting vector retrieval. Storage is shifting from an afterthought to a core component of AI pipelines.
Orchestration layer: Agent workflow orchestration — how to decompose high-level goals into executable steps, how to handle long-running tasks, retries, and state management. This layer is currently chaotic with no standard answer.
Memory layer: A completely new layer that traditional infra lacks entirely. Agents need to remember context, task state, and past decisions across sessions. This isn't a database, isn't a cache — it's a new infrastructure primitive.
Data pipeline and data systems layer: Jennifer Li explicitly says traditional data pipelines need to be rebuilt. Enormous amounts of enterprise data are trapped in PDFs, Zoom recordings, and Slack messages — AI Agents can't consume them directly. Whoever can structure this data can unlock massive enterprise automation scenarios.
a16z's View on Databases: From System of Record to Execution Layer
a16z's take on the shifting strategic position of databases:
"In 2026, the strategic layer shifts upward. The system that understands intent and executes workflows becomes the product users care about. The database underneath becomes a commodity. Founders should build for execution, not storage."
Databases don't disappear, but their strategic importance is declining. The moat shifts from "who owns the data" to "who owns the execution layer."
Oracle's moves provide an interesting contrast — it's doing the reverse, trying to stuff message queues, vector search, and AI inference into the database, making the database itself the execution layer. a16z's view is that the execution layer is the new battleground, but it won't necessarily be won by database companies.
YC's Signals: Evolution of the RFS
YC's Requests for Startups is a relatively reliable signal source because it's a synthesis of seeing a large number of real pitches — not predictions made from thin air. Below are the infrastructure-related directions from 2025–2026.
Spring 2025 (the start of the Agent wave): This RFS explicitly called out several directions for the first time: B2A (software whose customers are Agents, not humans — API design, documentation format, and permission models need to be redesigned for machines); Devtools for Agents (tools for managing, deploying, and monitoring Agents); Secure AI App Store (security isolation for Agent runtime environments). This batch set the tone for subsequent ones: Agents are a new user type, and existing software and infrastructure weren't designed with this assumption.
Summer 2025: This batch focused on consumer directions with few infrastructure signals. But "Inference AI Infrastructure in the World of Test-Time Compute" appeared, indicating YC recognized that inference and training require two different sets of infrastructure — an important engineering judgment.
Fall 2025 (the most focused batch, only 6 directions): One of the 6 was directly "Multi-Agent Infrastructure," indicating YC had already identified multi-Agent system infrastructure as the highest priority at that point in time. Another direction — "The First 10-person $100B Company" — implied that AI infrastructure enables extremely small teams to operate extremely large systems; the leverage effect of infrastructure is being amplified to a new magnitude in this era.
Spring 2026: AI-Native Enterprise Software — not adding AI features to existing systems, but rebuilding from scratch. YC's formulation: "only things that can't be retrofitted have a chance." This sentence is the core of the entire infrastructure rebuild thesis. Multi-Agent Systems also appeared in this batch, focusing on coordination, communication, and permission management between Agents.
Summer 2026 (the most recent and most aggressive)
Two directions directly related to infrastructure:
Software for Agents: Existing software is designed for humans — it has GUIs, mouse-click workflows, and ambiguities that humans tolerate but machines can't. Agents are a different type of user requiring machine-readable APIs, CLIs, and MCPs; identity systems, permissions layers, and payment infrastructure all need to be redesigned for autonomous programs. YC's exact words: "APIs, machine-readable documentation, command-line interfaces, identity systems, permissions layers, and payment infrastructure designed for autonomous programmes."
Company Brain: Every company's critical knowledge is scattered across emails, Slack, databases, and people's heads. AI Agents can't work in that state. A new infrastructure primitive is needed to structure fragmented knowledge, keep it updated, and turn it into skill files that Agents can execute. This isn't search, isn't RAG — it's a new infrastructure layer.
Signals from Specific YC Companies
The distribution of infra-direction companies in YC W26 and Spring 2026 gives more specific signals. Several clear sub-areas:
Auth/Identity: Agent identity management — who authorizes an Agent to do what. Microservices had OAuth for this, but the Agent identity model is completely different. Agents act on behalf of humans, making permission boundaries more complex.
Observability: What did the Agent do; what went wrong. Moda (W26) positions itself as "Datadog for agents" — Datadog was infrastructure for the microservices era; observability will be infrastructure for the Agent era too.
Testing: Testing and validation of Agent behavior. Traditional unit test frameworks can't handle the non-determinism of Agents.
Payments for Agents: Payment infrastructure for Agents completing transactions autonomously. Maven (W26) is building the payment layer for voice agents.
Communication for Agents: YC Spring 2026 accepted primitive.dev, described as "communication infrastructure for fully autonomous agents," built by former Microsoft/AWS/FB/Google engineers.
These sub-areas together represent the new infrastructure layer added in the Agent era: identity, observability, testing, payments, communication. All have counterparts in the traditional microservices era, but the design assumptions for Agents are completely different — the counterparts can't be directly reused.
One Direction That's Under-Discussed but May Have the Widest Impact
The impact of traffic pattern changes on existing middleware.
a16z says "thundering herd as default state"; YC's analysis also mentions "procurement is going machine-to-machine — reliability and determinism matter." But this direction is far less prominent in public discussion than Agent frameworks and RAG.
For example: existing message queues are designed for human-triggered business traffic with predictable peaks and smooth flow. Agent traffic isn't like that: bursty, recursive, extremely concurrent — and Agent-to-agent communication also requires a semantic layer (capability discovery, identity verification, permission checks), not just message transport.
Existing rate limiters are designed for human request rates; Agent rates exceed them by orders of magnitude.
Existing databases' connection pools and transaction models are designed for human-to-machine interaction response times; Agent concurrent access patterns are completely different.
These are design-assumption-level problems that can't be solved by tuning parameters — they require redesign. This is a real engineering problem, not a narrative.
What a16z Explicitly Doesn't Like
Jennifer Li has one clear reservation: many AI infra startups are valued on future demand assumptions, not supported by current real demand. a16z's view is that many AI infrastructure companies are overvalued.
What they believe is genuinely valuable: software that controls the decision layer (deciding how compute is allocated, how models are deployed), not companies that only sell compute capacity; companies that reduce complexity in heterogeneous hardware environments; companies that can truly turn unstructured enterprise data into formats Agents can consume.
What they don't like: companies without a control-layer moat, valued purely on AI narrative, with no real revenue to support it.
Summary
a16z and YC share a common underlying logic: AI's impact on traditional infrastructure isn't at the feature layer — it's at the design assumption layer.
Four core assumptions of traditional infrastructure fail simultaneously in the Agent era: requests come from humans at predictable rates; access subjects have UIs; system boundaries are clear and each layer is independent; data is structured.
The failure of these four assumptions corresponds to four categories of rebuild opportunities: traffic infrastructure, interface layer, middleware consolidation, and data structuring. The most certain opportunities right now are the interface layer (Software for Agents) and data structuring (Company Brain). Traffic infrastructure and middleware consolidation are less thoroughly discussed, but from an engineering perspective they are real problems.