The AI application layer is crowded. Value is shifting to infrastructure: post-training, inference, agents, security, sandboxes, observability, and compute.

Source note: this analysis is based on Redpoint Ventures’ May 2026 InfraRed report. The deck combines public company disclosures, broker research, PitchBook data, Redpoint survey work, and Redpoint estimates, so the numbers should be read as directional rather than neutral ground truth.

The Application Layer Correction

For the past three years, the software industry fixated on the application layer. Startups raced to build interfaces on top of frontier models, assuming that workflow integration would be enough to secure a durable advantage.

That assumption is now being tested. The market is going through an infrastructure correction. As the novelty of generative text wears off, application differentiation without infrastructure control looks thin. Companies that built mostly on commodity APIs are competing on workflow polish, brand, distribution, and taste, not on a deep technical edge.

Redpoint’s public-market cut shows the split clearly. Over the trailing twelve months, the deck shows application SaaS down 39 percent. These companies are also facing slower growth rates, partly because enterprise AI budgets are beginning to cannibalize traditional software spend. Redpoint’s CIO survey says AI investment often comes from replacing existing application contracts. At the same time, incumbents are moving quickly to embed AI features, squeezing thin application startups that offer little structural advantage.

Infrastructure software is on the other side of that trade. Redpoint shows infrastructure software up 38 percent. Consumption pricing lines up well with agentic activity and machine-generated code. Hyperscalers like Amazon Web Services, Microsoft Azure, and Google Cloud are seeing growth reaccelerate, helped by billions in new AI revenue.

The tailwind extends beyond raw compute. Companies providing observability, networking, databases, orchestration, and storage are selling into the operational mess created by AI adoption. Datadog benefits from the need to monitor reasoning loops, GPU utilization, and AI-generated code deployments. Cloudflare benefits from an internet where non-human traffic could overtake human traffic by 2027. When an AI agent performs a task, it can generate thousands of network requests where a human would generate a handful of clicks. Infrastructure absorbs that volume and captures the spend attached to it.

Scaling Is Moving Past Pre-Training

The underlying mechanics of AI development are changing. The first generative AI boom relied on a simple formula: scrape the public internet and scale pre-training compute. That formula is running into a data limit. The deck cites projections that usable human text could be exhausted by 2028. Synthetic data helps, but only up to a point.

Redpoint frames the next phase around three scaling vectors: post-training, test-time scaling, and agentic scaling.

Post-training shifts the bottleneck from raw data acquisition to judgment and verifiable environments. Models are refined through reinforcement learning in domain-specific settings, such as coding sandboxes or life science simulators. If the system can verify whether an answer is right, it can generate more useful training signal. That changes the data-labeling market. The valuable work moves from cheap internet-scale annotation toward expert judgment, task design, evaluation, and environment generation.

Test-time scaling changes the economics of inference. Instead of producing one immediate response, models can think longer, sample multiple reasoning paths, check their work, and revise before answering. That makes them more useful, but it is not free. Redpoint cites reasoning queries as consuming 30 to 100 times the inference compute of a standard chat query. The per-token price may fall, while the per-task cost still rises. That makes model routing, semantic caching, and cost-aware orchestration core infrastructure rather than optimization trivia.

Open source is gaining ground for cost and control. As closed APIs become expensive at scale, enterprises have stronger reasons to self-host open alternatives. Redpoint estimates a very large open source inference market, while assuming open models capture only a minority share of total inference workloads. The case is not purely financial. Smaller open models can run faster and more reliably in production environments.

The goal is domain-specific intelligence. By post-training open source base models on customer interactions, internal documents, and workflow data, companies can develop specialized models that understand their domain better than a general API can. The model matters, but the loop matters more: usage creates feedback, feedback improves the model, and the improved model drives more usage. Owning that loop creates a compounding advantage.

Agents Need Their Own Infrastructure

AI interaction is moving from simple tool use to process execution. Instead of humans prompting chat interfaces, self-directed agents operate across browsers, terminals, IDEs, and internal tools. Redpoint cites research suggesting that the length of a task an AI can complete autonomously is doubling roughly every seven months.

Agents do not fit neatly into old application patterns. They can be reactive, triggered by a human prompt, or proactive, triggered by an event. They can operate sequentially, in orchestrator-worker patterns, or as networks of agents handling open-ended work. The rise of agent frameworks and MCP servers shows this architecture is moving from demo to operating assumption.

Scaling those agents requires new infrastructure. Traditional databases and orchestration systems were designed around human-speed software usage. Agents query systems far more frequently. They need durable execution engines like Temporal to manage long-running tasks that might fail and require state replay. They need analytical systems like ClickHouse to process high-velocity telemetry. They need state, memory, permissions, observability, and auditability.

Software development is the clearest early market because code has a clean reward signal: tests either pass or fail. That makes coding a natural reinforcement-learning environment. Better coding agents can improve the infrastructure used to build better models, creating a feedback loop that explains why model labs care so much about developer tools.

The old software development lifecycle is bending under that pressure. The bottleneck used to be writing code. Then it became reviewing code. Now it is often the execution environment. The IDE, CI system, sandbox, and test runner are collapsing into an agent control plane.

That is why sandbox infrastructure matters. Agents cannot safely test and iterate on a user’s local machine without guardrails. They need secure, isolated execution environments where they can run code, inspect results, and recover from failure. Companies like E2B, Tensorlake, Coder, and Daytona are part of this new execution layer. The sandbox becomes the place where agent work actually happens.

Security Becomes a First-Order AI Market

AI execution speed creates a security problem that cannot be treated as an afterthought. The volume of deployed code is rising as developers use agents to ship faster than review systems can keep up. Less experienced developers can accept AI suggestions uncritically, putting vulnerable code into production. Agents are also being deployed with broad permissions, API access, and weak identity governance.

Redpoint breaks the security opportunity into several fronts.

First, enterprises need to secure the AI pipeline itself. Model weights can represent hundreds of millions of dollars in intellectual property. Training data can be reconstructed through adversarial queries. Every employee prompt is a possible data leak. Autonomous agents need identity, permissions, and audit trails. The model, the employee, and the agent each need a control plane.

Second, AI gives defenders an advantage. Security operations centers are overwhelmed by alerts. Autonomous agents can triage alerts, investigate incidents, and perform root cause analysis, taking on work that analysts previously spent on false positives. Similar agents can run continuous offensive security work, executing reconnaissance and exploitation at machine speed instead of waiting for periodic manual penetration tests.

Third, AI labs are becoming security vendors. Companies like Anthropic, OpenAI, and Google are increasingly competing with traditional security incumbents. Their models find vulnerabilities in large codebases, changing the market for vulnerability management.

Finally, AI pulls compute back to the endpoint. For fifteen years, compute moved steadily toward the cloud and security tooling followed. Coding agents and local model inference reverse part of that trend. They need local filesystem and shell access. Traditional endpoint tools watch for malicious binaries at the kernel level. They are less effective against agents executing scripts in application space. The endpoint is unstable again.

Physical AI Is the Next Extension

The scaling logic of digital intelligence is moving into the physical world. Multimodality brings text, code, and structured data together with video, 3D geometry, and sensor data. Digital data is further along the curve, but physical data is starting to matter more.

The key concept is the world model. A video model produces a fixed clip. A world model produces an interactive environment that responds to action. It needs memory, physics, and low-latency response. That makes it useful for robotics because real-world training is slow, expensive, and sparse.

World models give robots something like synthetic practice. A robot can rehearse rare or dangerous events in simulation, compare possible actions, and commit to the one most likely to work. Combined with cheaper hardware, simulation scaling could push robotics toward the kind of acceleration that language models went through earlier.

What To Watch

The report’s core point is not that every infrastructure company wins. It is that the AI value chain is getting longer, messier, and more operational. That creates room for new infrastructure categories, but it also raises the bar. A generic AI narrative is not enough.

The first thing to watch is hardware diversification. Synchronous inference needs high-end GPUs and fast memory. Background agents can often use cheaper, older chips. Edge inference needs low-power silicon. As workloads fragment, the hardware stack should fragment too.

The second is agent control planes. Coding agents will force sandbox infrastructure, identity, permissions, observability, and durable execution to mature together. The companies that provide secure, low-latency execution environments for agents could become part of the new foundation of software development.

The third is the collision between AI labs and cybersecurity incumbents. Vulnerability discovery is no longer only the domain of human researchers or traditional security software. If frontier labs keep releasing models tuned for security analysis, incumbents will need to build, buy, or partner their way into equivalent capability.

The takeaway is simple: AI does more than create apps; it changes what infrastructure must do. Durability belongs to companies closest to inference cost, agent execution, security, and learning loops.

Source

This post synthesizes the findings, estimates, and market categorizations presented in Redpoint Ventures’ May 2026 InfraRed report.