AI applications have transcended the realm of mere "concepts" and are now delivering tangible results.

Recently, Madrona Venture Group, a prominent U.S. venture capital firm, unveiled the Intelligent Applications 40 (IA40) for 2026.

This list is far from ordinary; it represents a collective endorsement by the primary market of AI applications.

Madrona teamed up with 72 investors from 54 leading VC firms and corporate investment entities to nominate and assess over 450 companies. By integrating data scores from PitchBook, they identified the 48 most noteworthy AI application and infrastructure companies of the year.

This year, Madrona's theme is “Unlocking the Value of AI,” emphasizing the realization of AI's true potential.

This shift is evident in revenue growth. Among the companies selected this year, at least 26 have disclosed clear revenue or Annual Recurring Revenue (ARR) figures, with at least 9 boasting ARR exceeding $100 million.

What's even more compelling is that as revenues surge, companies' strategies are evolving in tandem.

On one hand, AI applications are expanding their reach along workflows: from being a tool for a single task, they are now assuming control of entire workflows, integrating clients' meetings, projects, customer data, and business rules into the product.

On the other hand, as AI agents (Agents) truly begin to serve enterprises, the focus of underlying infrastructure is shifting, with companies addressing Agent costs, stability, and other critical aspects garnering more attention.

Silicon-based Insider will leverage these 48 companies to delve into some of the transformations occurring in AI applications.

/ 01 / AI Applications: From Selling Tools to Dominating Workflows

In 2026, two significant shifts are emerging at the application layer.

Firstly, product boundaries are expanding, transitioning from optimizing individual steps to delivering holistic outcomes.

Among the 33 application companies, 18 are no longer content with their initial function but are now tackling more steps before and after the core task.

For instance, Granola initially focused on creating meeting minutes but now connects information from various team meetings, enabling users to directly query clients, projects, and historical decisions.

Gradial, initially centered on marketing content generation, now also handles translation, review, compliance, and publishing.

Serval started with IT Q&A and has now begun directly processing IT service processes like work orders and permission requests.

Despite operating in different domains, their strategies are remarkably similar: enter enterprise processes with a core function, then gradually assume control of other steps necessary to complete the task.

The rationale behind this is straightforward. Once AI automates the simplest step, what truly impedes progress becomes subsequent reviews, collaboration, and execution.

Take Gradial’s marketing example: AI can generate marketing content in mere seconds.

However, for a large enterprise to actually launch it, translation, brand review, compliance checks, CMS entry, and publishing are required—potentially involving over ten steps.

Even if "generation" is swift, the entire process may still take two weeks or longer.

Thus, Gradial gradually expanded from content generation to downstream marketing operations.

Take its partnership with T-Mobile as an example: Gradial seamlessly integrates with T-Mobile’s CMS, asset library, and work order system, incorporating page building, asset processing, content updates, as well as brand, compliance, and QA checks into the workflow.

The outcome?

A large marketing campaign at T-Mobile that previously took a month and 1,000–1,200 cumulative hours now takes about 80 hours with manual QA after integrating Gradial, reducing launch time by approximately 90%.

As products begin to compress overall work delivery time, the value clients are willing to pay for also increases.

Secondly, AI applications are starting to comprehend clients' business context.

Among the 33 application companies, 26 can continuously read and utilize clients' business information and historical records.

This context may encompass a company’s past meetings, customer history, project progress, or internal knowledge, business rules, and past decisions.

Sierra serves as a prime example. It provides AI customer service, enabling Agents to directly handle customer service issues.

In Rocket Mortgage’s mortgage scenario, Sierra’s Agent doesn't merely respond based on the user’s current statement but retrieves existing loan information and business system data to complete credit checks, personalized rate displays, loan plan adjustments, and repayment and loan status queries.

What does this signify?

When the same customer re-engages in the conversation, the Agent encounters not an isolated input but a comprehensive context built on their loan progress, historical interactions, and current business status.

The result is striking: customers who initiate the mortgage process with Sierra’s digital assistant have a final closing rate three times higher than non-users; those utilizing both AI chat and human bank advisors see their home purchase and refinancing conversion rates increase fourfold.

This reflects a shift in the source of AI application value.

Previously, much of the application layer’s capability stemmed from the model itself. However, as foundational models become more accessible, what truly becomes exclusive to the application layer is the context originally accumulated in clients' CRMs, business systems, and historical records.

The barrier for AI applications lies in how well you understand your clients.

/ 02 / Agent Infrastructure Shifts Focus

As application companies begin to assume control of more steps, the underlying infrastructure market also shifts its focus.

In 2025, when Madrona discussed Agent infrastructure, the emphasis was on connection protocols like MCP and A2A, Agent orchestration represented by CrewAI and LangChain, and evaluation tools represented by Haize Labs and Braintrust.

By 2026, the list’s focus has clearly evolved.

Last year’s highly regarded evaluation and orchestration companies like CrewAI and Haize Labs have largely exited the scene. This year, Baseten, OpenRouter, and Fireworks have newly entered the official list, and together with Together AI and Fal—five of this year’s 12 infrastructure companies now focus on inference and routing, a significant increase from last year’s two, accounting for over 40%.

In other words, as Agents truly commence operations, the costs, scheduling, and operational issues concealed behind a single model call are also being brought to the forefront.

Around these issues, three directions are clearly gaining momentum.

Firstly, inference optimization.

After Agents truly execute tasks, a single task often involves multiple rounds of model and tool calls. As the chain lengthens, Token costs and latency escalate together. What Baseten and Fireworks do is make model inference more cost-effective and faster.

For example, Baseten focuses more on resolving compute scheduling issues: during periods of low traffic, it reduces idle GPU resources occupied by models; when traffic surges, it activates more GPUs and can schedule GPUs across different cloud providers and regions.

Fireworks, on the other hand, optimizes GPU usage efficiency by separating different computational aspects of model inference to allocate resources, reducing redundant computations through caching and request scheduling, enabling the same GPU to handle more requests.

For Agents, these optimizations ultimately boil down to one question: can tasks be completed at a lower cost and faster speed?

And this is precisely what must be addressed for Agents to truly enter production environments.

Secondly, model routing.

An Agent’s task often encompasses multiple types of work: some require strong reasoning, while others are simple extractions, and model calls must also consider stability.

More troublingly, even if the model is determined, the inference supply behind the same model is not stable—prices, speeds, capacities, and tool call performances of different providers may continuously fluctuate.

What OpenRouter does now is two-layer routing:

Auto Router initially selects a model based on task type and cost requirements; Auto Exacto then chooses a more suitable inference provider among those offering the model and dynamically adjusts ranking based on speed, tool call performance, and Benchmarks—updating approximately every five minutes.

This means that with every step an Agent takes, model routing can dynamically decide which model to use. As models and inference supplies become increasingly fragmented, how to choose among them is also becoming a separate layer of infrastructure.

Thirdly, runtime environments.

After Agents truly execute tasks, besides calling models, they may also need to log into web pages, run code, process files, and complete a series of operations consecutively.

The model decides what to do next, but these actions themselves require a stable execution environment.

Browserbase initially provided dedicated cloud browsers for Agents because many websites lack directly callable APIs when Agents truly execute tasks.

Browserbase places browsers in the cloud and manages the runtime status, identity, and operation records of each browser task, enabling Agents to directly operate web pages.

Now, Browserbase has also taken over the infrastructure for running Agent code through Functions and Runtime—Agent code can run directly beside the browser, with browser startup, task execution, logs, and results managed by the same platform.

This runtime environment has already begun to handle large-scale calls from clients.

After Airtable launched its Agent in April this year, it executed over 300,000 browser tasks through Browserbase in less than three months, averaging over 3,000 daily.

In other words, as Agents shift from single calls to continuous execution, runtime environments have also become production infrastructure.

/ 03 / Conclusion

From applications to infrastructure, what IA40 reflects this year is essentially the same trend.

The deeper AI integrates into application environments, the more its value derives not just from the model itself.

For application companies, providing just one function is no longer sufficient. They need to delve deeper into complete workflows, grasp more of their clients' business context, and truly participate in completing work.

For infrastructure companies, opportunities no longer arise solely from "helping developers build Agents" but from the continuous inference, scheduling, and execution needs that emerge after Agents truly commence operations.

Whoever is closer to real work and controls more execution steps has a better chance of capturing the next stage of value.

Text/Yuanyuan