Unsupervised Learning• Apr 23, 2026• 54:52Interview

Has AI Infra Stabilized, FM Vibe Shift, & What's Next for Coding Agents

From Unsupervised Learning

Swix•Host of Latent Space & AI Community Organizer

Executive Summary

The AI coding market has become the primary battleground for major labs like Anthropic and OpenAI, who are generating billions in ARR and have designated it their highest priority (P0).
A successful strategy for startups, the "agent lab playbook," involves using foundation models to gather data, then training proprietary, domain-specific models to create a competitive moat.
AI agents are rapidly becoming the primary consumers of cloud infrastructure, with companies like Vercel reporting that over half their administrative traffic now comes from agents, not human developers.
While the AI infrastructure layer is showing signs of stabilization, vertical AI application companies are seen as more defensible than horizontal platforms due to the threat of being displaced by foundation model providers.

12 quotes

Concerns Raised

Midsize startups are at high risk of being displaced by foundation model providers.
The consumer AI application category has hit a plateau in user growth and engagement.
The rapid pace of change in the AI landscape creates existential risk and strategic instability, even for late-stage companies.

Opportunities Identified

The AI coding market is a massive, multi-billion dollar opportunity driving the revenue of major labs.
Vertical AI applications in sectors like finance and healthcare represent the next major growth areas.
Training domain-specific models is becoming more accessible, enabling startups to build competitive moats.
AI agents represent a new, rapidly growing customer segment for infrastructure and developer tool companies.

Key Themes

The AI Coding Wars

The AI coding market is the central battleground for major AI labs, with Anthropic's Claude Code and OpenAI's models competing fiercely for dominance. This vertical has proven to be a multi-billion dollar market, driving both labs to prioritize it above other applications and fueling the growth of specialized startups like Cursor and Cognition.

This intense competition accelerates innovation in developer tools and AI-assisted software engineering, fundamentally altering the developer workflow and creating a massive new market for AI products.

The Agent Lab Playbook

A key strategy for AI startups to compete with large labs is to first leverage state-of-the-art foundation models to build a product and gather high-quality, domain-specific data. They then use this data to train their own specialized models, which can outperform general models on specific tasks while reducing cost and latency.

This playbook provides a viable path for startups to build defensible businesses. It highlights that competitive advantage can be built not just at the application layer, but by creating a data flywheel to train superior, specialized models.

Vertical vs. Horizontal AI Strategy

The discussion contrasts the strategies of vertical AI companies (e.g., Abridge in healthcare) with horizontal infrastructure providers (e.g., LangChain). Vertical players, acting as the "outsourced AI team" for enterprises, are considered more robust and defensible against the expanding capabilities of foundation models.

This dynamic shapes investment theses and go-to-market strategies in the AI ecosystem. It suggests that deep domain expertise and last-mile integration are critical moats that are difficult for large, generalist model providers to replicate across many industries.

The Evolution of AI Infrastructure

While there are signs of stabilization in the AI development stack, the landscape remains volatile. A significant shift is occurring where AI agents are becoming the primary customer for infrastructure, surpassing human developers in traffic. Concurrently, the increasing accessibility of custom hardware and training tools is democratizing the ability to build specialized models.

This trend forces infrastructure companies to re-evaluate their target user, optimizing for programmatic access and agent-driven workloads. It also signals a maturation of the MLOps landscape, enabling more companies to move beyond simply calling APIs to training their own models.

The Next Frontiers: World Models and Embodied AI

Looking beyond current capabilities, the conversation identifies world models and spatial intelligence as the next major frontier for AI research. The goal is to move beyond next-token prediction to imbue models with a common-sense, physics-based understanding of the world, which is seen as a prerequisite for more advanced reasoning and robotics.

Progress in this area represents a step-change in AI capabilities. It could unlock new applications in robotics, simulation, and autonomous systems that are impossible with current LLM architectures, which lack true world experience.

Get started free

Topics

AI Coding Foundation Models Market Structure Anthropic OpenAI LangChain AI Infrastructure Agent Lab Playbook Vertical AI Enterprise AI AI Agents Model Training Vercel Cursor Cognition ARR (Annual Recurring Revenue)World Models

Processed Apr 23, 2026 yt-dlp + mlx-whisper + Gemini