Deep|GOOG:LLM Narrative Shift, Anthropic vs Gemini, and AI Overview

GOOG May Be Approaching Its Fundamental Bottom

Two months after our extensive coverage of how Anthropic would impact AWS and drive ASIC business, we're seeing signs of a potential narrative shift.

For our latest quarterly insights on LLMs, please refer to our recent article, much of which was echoed in Semi Analysis's latest LLM piece. However, we believe our understanding of frontline LLM progress is more detailed than Semi Analysis's assessment.

Deep|LLM 1Q25:AI Agents Are Going To Transform The Entire Investment Landscape

We haven't yet established a complete framework for analyzing GOOG, so we're not ready to publish a GOOG Preview. Like MSFT, META, and AMZN (which we've already covered), GOOG faces complex key debates.

Given the urgency, I want to share our latest findings on GOOG, including updates on Anthropic vs Gemini, Agent developments, AI Overview/Mode progress and search experts interviews.

We expect to release 2-3 quarterly research updates on the autonomous driving industry this week. These were originally scheduled for last week but were delayed for additional review following the Twitter dispute between Elon Musk and Trump.

LLM Narrative Shifts

While our detailed perspective was presented in last week's LLM Q1'25 report, here's a distilled version.

We've observed three significant inflection points in frontier LLM research:

Bullish: Post-Training Surpassing Pre-Training

The first shift occurred last July when post-training consumption exceeded pre-training at leading LLM companies. This marked the first time post-training replaced pre-training as the most crucial step. At that time, frontier researchers were bullish because post-training and pre-training created a perfect training loop.

Post-training requires high-quality data (including Chain-of-Thought) but in smaller volumes, making it ideal for marginal improvements. Post-training models generate intermediate reasoning data during inference that, while not directly usable for post-training, can serve as pre-training data. Researchers were confident about LLM training generalization.

This positioned post-training as the key to continued pre-training scale-up.

The capital markets lagged frontier research by about two months, with NVIDIA actually hitting its low in August.

Bearish: Post-Training Generalization Proving More Difficult Than Expected

The second shift occurred in late November when we discovered that pre-training progress at top LLM companies had hit significant walls in data and generalization.

Regarding the data bottleneck: While pre-training relies on public datasets, post-training depends heavily on talented PhDs. Simply put: the number of synthetic data models correlates directly with the number of brilliant PhDs available. This creates a largely manual data synthesis approach, with each researcher responsible for specific domains (e.g., group theory, dynamic programming). The bottleneck emerged because PhD hiring couldn't keep pace with GPU acquisition.

As for generalization: PhD-synthesized data concentrated in coding and math domains, creating highly specialized datasets. Models needed to generalize beyond these areas. Frontier researchers hoped self-play algorithms would succeed (evolving from AlphaGo to Alpha Zero), but progress was slow.

This significantly hindered post-training's ability to drive continued pre-training scale-up. Between November and March, top model companies hit pre-training walls. Although Anthropic claimed its pre-training hadn't plateaued, much of its post-training work was actually being completed during the pre-training phase, unlike OpenAI's separate Reasoning Model.

The capital markets again lagged by two months, with the AI narrative only shifting noticeably in February with DeepSeek's emergence.

Bullish: Cursor and Agents

The third shift occurred this March when Cursor proved Anthropic's approach correct.

Cursor grew from under $100M ARR six months ago to $300M ARR in April, then quickly reached $400M and $500M ARR in May and June.