Originally written: April 2024 | Published on Gorilla PE Insights: March 2026

This article is a public release of an internal partner memo shared in April 2024.

In April 2024, we introduced a framework to understand exactly how AI will replace human labor. At the time, the mainstream market narrative was singular and blunt: "AI is taking our jobs." Our question was different. We asked: "From where to where, and in what sequence, will this replacement actually occur?" From that question emerged the framework we call the Asymmetry of the Two Waves.

Part 1. The Asymmetry of the Two Waves — Why Not Just One?

Let’s broadly divide labor into two categories: Cognitive Labor, which consists of language, reasoning, and judgment; and Physical Labor, which involves spatial movement, manipulation, and physical force. The pathways through which AI penetrates these two domains are fundamentally different.

The bottleneck of Cognitive Labor is language. When a lawyer reviews a contract, an accountant analyzes a financial statement, or a consultant writes a report, they are ultimately reading and generating text. This is precisely the bottleneck that the Transformer architecture and RLHF (Reinforcement Learning from Human Feedback) have solved. LLMs have already crossed the threshold required to replace a significant portion of white-collar work.

It was in 2023 that GPT-4 passed in the top 10% of the Uniform Bar Exam and the top 1% of the Biology Olympiad. A 2023 MIT study by Noy & Zhang showed that workers using ChatGPT completed tasks 40% faster while increasing quality by 18%. GitHub Copilot has boosted developer productivity by 55%. The first wave is already here.

The bottleneck of Physical Labor is entirely different. Moving boxes in a warehouse or hauling materials on a construction site is not a matter of language; it is a problem of spatial awareness, balance, and force control. You cannot solve this with an LLM. It requires sensors, actuators (devices that convert electrical energy into physical motion), batteries, and highly advanced vision models.

The core thesis is this: The two waves arrive at different times, through different vectors. The LLM wave has already crashed ashore. The Physical AI wave is still building.

Part 2. The Bottleneck of Physical AI — Why the Delay?

The Scarcity of Physical World Data. Language models can use the entire text of the internet as their training data. Physical AI, however, requires data generated by actual robots moving in the actual physical world. You cannot scrape this from the web. You have to collect it yourself. You have to drive billions of miles.

The Problem of Edge Cases. On a factory floor, entirely unpredictable situations occur constantly. A dropped screw, a wet floor, a newly shaped component. As these edge cases accumulate, the system's real-world robustness increases. The entity that gathers this data the fastest secures a structural moat in Physical AI.

Why Tesla FSD Matters. As of 2024, Tesla has a fleet of over 6 million vehicles driving on real roads, vacuuming up data. For context, Waymo’s testing fleet sits at roughly 700 vehicles. This is a structurally different game of data accumulation. In FSD v11.3, the distance between driver interventions was 35 miles; by v13.2, it leaped to 724 miles. That is a 20x improvement. And crucially, this data is directly transferable to Optimus (their humanoid robot).

Part 3. The Intersection of the Two Waves

The Generative AI market is projected to grow from $67B in 2023 to $1.3T by 2032, a staggering CAGR of 39%. This market is already highly visible and is rapidly converging into an oligopoly.

The Physical AI market is a different beast. While it’s projected to grow from $30B in 2023 to $293B by 2032 (with Humanoids at a 48% CAGR and Drone AI at 32%), the market itself is still in its formative stages. The barrier to entry here is dictated by the sheer time required to accumulate physical data.

The ecosystem with the most formidable positioning at the intersection of these two waves is the one that possesses both LLM capabilities and physical world data simultaneously. As of April 2024, the structure closest to meeting that condition is the combination of xAI, Tesla, and X. While xAI trains cognitive intelligence using the real-time language data of X, Tesla FSD trains spatial intelligence using the physical data of the real world. Both currents flow within the same ecosystem.

Part 4. Counterarguments — Why This Thesis Might Be Wrong

The strongest counterargument is Tesla FSD itself. After more than a decade of development and billions of dollars invested, it still hovers at an L2+ autonomy level. The distance to true, full autonomy remains frustratingly unclear. This is the most glaring empirical evidence of just how difficult Physical AI truly is.

The counterargument is perfectly valid. However, we read the conclusion differently. When the decade of data and scars accumulated by Tesla FSD is transferred to Optimus, that first-mover advantage becomes practically impossible for a competitor starting from scratch to replicate. The harder the problem, the more insurmountable the first-mover advantage becomes.

The second counterargument is the fundamental limit of LLM hallucinations. To frame it using Judea Pearl’s Ladder of Causation, LLMs are exceptional at pattern recognition (capturing correlations), but they possess structural limitations when it comes to true causal reasoning in novel environments. This flaw could prove fatal in Physical AI. If a system encounters a new environment (an Out-of-Distribution scenario) with no pre-existing pattern, it fails.

Conclusion: The Question This Framework Poses

The first wave is already here. The second wave is inbound.

If this framework holds true, investors must ask themselves one critical question: At the intersection of these two waves, who is the first to accumulate the most irreplicable asset of all—physical world data?

That was the question we asked in April 2024.

[Subsequent Developments] 

In January 2025, Jensen Huang officially declared the era of "Physical AI" during his CES keynote, stating that "the ChatGPT moment for robotics is coming." The directional logic of the 'Asymmetry of the Two Waves' remained completely valid. However, observations now indicate that the timeline for Physical AI to reach true commercial scale will likely be 2 to 3 years longer than our early 2024 projections.