State of Startups and AI 2025 - Sarah Guo, Conviction

The current AI landscape represents the largest technology revolution of our time, but the path to value creation is more complex than it appears. While AI companies are achieving unprecedented growth, going from zero to over $100 million in annual run rate faster than ever before, the engineering and product challenges are significant. The focus should be less on the market hype and more on user uptake and the fundamental capabilities that are unlocking new applications.

Advancing Capabilities: Reasoning, Agents, and Multimodality

The past year has been defined by significant advancements in core AI capabilities.

• Reasoning: This has emerged as a new vector for scaling intelligence with compute. It unlocks a new set of use cases, particularly for sequential problems, systematic search, and high-stakes decisions where the ability to “show the work” is critical for user trust and adoption.

• AI Agents: Moving beyond simple chatbots, agents are systems that can plan, take ownership of a task, hold a goal in memory, and backtrack to try different hypotheses. This can range from simple, single-purpose tools to sophisticated systems that resemble a human colleague. The interest in this area is exploding, with a 50% increase in agent-based startups over the last year.

• Multimodality: Voice, video, and image generation models are rapidly maturing into real businesses, with companies like HeyGen and Midjourney surpassing $50 million in ARR. Voice is poised to be the first major enterprise application, enhancing workflows like medical consultations and lead generation. As controllability improves and costs fall, video and image applications will become more widespread, structuring previously untapped, high-bandwidth data sources.

A Competitive and Commoditized Model Layer

The market for foundational models is becoming increasingly competitive, which is a significant advantage for application builders. As Sam Altman noted, “last year’s model is a commodity.” The price of GPT-4, for example, dropped from $30 per million tokens to $2 in about 18 months.

This trend is driven by intense competition. While OpenAI initially dominated, players like Anthropic (Claude) and Google (Gemini) have captured significant market share. Furthermore, new, credible challengers like SSI and Thinking Machines are entering the market with orthogonal technical approaches, alongside potent open-source models. The key takeaway for builders is to plan for a multi-model world, leveraging tools that allow for flexible and cost-effective model orchestration.

The Application Playbook: Building “Cursor for X”

The success of coding assistants provides a powerful playbook for other domains. Companies like Cursor, Cognition, and Windsurf have demonstrated a killer application, with Cursor growing to half a million developers and a $100 million run rate in just 12 months with zero salespeople.

Code was the first frontier for several reasons:

1. Structured Language: Code is text with logical, structured rules.
2. Deterministic Validation: It’s possible to automatically test and verify if the output works.
3. Research Priority: Labs poured resources into code generation as a key benchmark for AGI.
4. Founder-Market Fit: Crucially, engineers built these tools for themselves, possessing an intimate understanding of the user’s workflow.

This last point is the blueprint for every other industry. The winners will not just be AI experts, but customer-centric builders who redesign workflows from first principles.

The value is not in the model itself but in the “wrapper” around it. A successful application like Cursor is a “thick wrapper” that:

• Uses multiple models for specific sub-tasks (e.g., diffs, merges, embeddings).
• Skillfully packages context and prompts the models.
• Automates repetitive tasks and standardizes best practices.
• Improves retrieval accuracy with usage over time.
• Integrates seamlessly into a familiar, safe, and fast user experience.

From a user’s perspective, the prompt is a bug, not a feature. The best AI products feel like they are reading your mind because they are built on a deep understanding of the user’s domain and needs, automatically collecting context and presenting solutions thoughtfully.

The AI Leapfrog and the Power of Co-Pilots

Counterintuitively, some of the fastest AI adoption is happening in conservative, low-tech industries. This “AI leapfrog effect” is visible in law (Harvey), medicine (Open Evidence), and customer service (Sierra), where companies are solving tangible problems and achieving massive scale by deeply integrating into professional workflows.

While the end-game may seem to be full automation via agents, the data shows that co-pilots are what’s driving revenue and adoption today. The “Iron Man suit” analogy holds: it’s best to start by augmenting the human user and then gradually increase autonomy as the underlying models become more reliable. Human tolerance for failure drops dramatically as latency and task complexity increase, making augmentation the most practical path to user trust and value in most domains today.

Execution is the Only Moat

In the AI era, traditional moats are less relevant. Execution is the new moat. Cursor didn’t invent code completion or the models it uses; it out-executed competitors by shipping a superior experience faster. In contrast, companies with first-mover advantage, like Jasper, faltered because their product was a “thin wrapper” that was easily replicated.

The advice for builders is to “build something thick and stay ahead.” Don’t be afraid to compete with the labs. Magical AI experiences that are deeply integrated into workflows build customer trust, drive adoption, and generate proprietary data and context that the labs cannot easily access. The opportunity for engineers and builders is to act as the translators, taking the magic of AI and applying it to solve real-world problems.

• Tags:
• AI Agents
• AI Applications
• Venture Capital
• Feature
• AI Strategy
• Co-Pilot
• Model Layer
• Execution