Generative AI’s Act o1: Why Reasoning Layers Are the Next Battleground

Sequoia’s latest observation, "Generative AI’s Act o1," marks the third installment in a series tracking the evolution of generative AI. The report notes that the pre‑train foundation‑model layer has largely stabilized, and the next competitive frontier will be the reasoning layer.

According to the analysis, the generative‑AI market is moving from fast, System 1‑style thinking to slower, more deliberate System 2 reasoning. OpenAI’s o1 model—referred to as the legendary Q* or Strawberry—demonstrates this shift by allocating extra compute to "stop and think" during inference, achieving a breakthrough in reasoning capability.

The authors claim a new scaling law is emerging: inference‑time compute becomes the primary lever for improving model performance, and o1 may represent the "AlphaGo moment" for generative AI.

Future focus is expected to move from pre‑training models to inference clouds. Dynamic scaling and elastic resources in the cloud allow models to devote more time to reasoning, leading to the slogan "No Cloud, No GenAI."

Application‑specific reasoning is highlighted as essential for real‑world AI deployments. While general‑purpose reasoning has advanced, successful large‑model applications still require domain‑specific reasoning capabilities. The report introduces the concept of a cognitive architecture—custom code and model interactions that mimic human thought to solve targeted tasks, also described as an "agentic application."

A shift from Software‑as‑a‑Service to Service‑as‑a‑Software is proposed, encapsulated in the phrase "Selling work, not software." This new delivery model emphasizes pricing based on delivered value or completed work rather than per‑seat software licenses, aiming to overcome the long‑standing bottleneck of AI product adoption.

The article stresses that foundation models are not products themselves; delivering a usable AI service requires a complex cognitive architecture, robust workflows, and deep domain expertise. It cites examples such as Sierra’s AI‑driven customer‑support agent (pay‑per‑issue) and XBOW’s AI‑powered penetration‑testing service (pay‑per‑service) to illustrate the "selling work" paradigm.

Finally, the report outlines three evolutionary stages of generative‑AI applications:

Act 1 (Technology‑driven): Early emergence of generative AI powered by transformers and diffusion models, focused on flashy demos.

Act 2 (Customer‑oriented): Shift toward solving specific client problems with end‑to‑end solutions, custom LLMs, integrated workflows, and assistant‑type tools.

Act o1 (Reasoning era): Marked by OpenAI’s o1 model, emphasizing inference‑time compute, System 2 thinking, and inference clouds, while software delivery may transition to "service‑as‑software" with a focus on selling work.

The authors conclude that the value chain—from GPUs to cloud infrastructure, foundation models, optional domain‑specific models, and finally LLM applications—offers ample white‑space opportunities, especially for startups that can position themselves in the reasoning and workflow layers.