Originally written: May 2023 | Published on Gorilla PE Insights: March 2026

This article is a public release of an internal partner memo shared in May 2023.

In May 2023, we set out to establish the first true benchmark of the AI race. At the time, market sentiment was violently split into two camps: one declaring that "the world will never be the same," and the other dismissing it as "just another tech bubble." Our judgment was that both camps were wrong. Yes, the world is fundamentally changing—but the beneficiaries of that change will not be evenly distributed. And no, this is not a bubble.

This piece is not merely a verdict on OpenAI as a single company. It is a documented attempt to recognize AI as a structural paradigm shift and to ask where the first definitive tollgate of that shift is being built.

Part 1. Why Now? — Breaking the Threshold of 60 Years of Research

From Neural Networks to ChatGPT: This Didn't Happen Overnight

To truly understand ChatGPT, you cannot start in 2022; you have to rewind to the 1960s. The concept of Artificial Neural Networks (ANNs) has existed for decades. However, for a long time, AI cycled through brutal "winters" and brief "springs." The theoretical ideas were there, but three critical components were missing: massive data, raw compute power, and the right architecture.

2012: The Resurrection of Deep Learning. At the ImageNet competition, AlexNet—based on Convolutional Neural Networks (CNNs)—crushed traditional methods. It achieved an error rate of 15.3%, less than half of the previous best record (26.2%). It was the exact moment deep learning proved it could approach human-level accuracy in image recognition.

2017: The Dawn of the Transformer. Google published the seminal paper, "Attention Is All You Need." This is the 'T' in GPT. The breakthrough was the Attention Mechanism, which processes the relationships between all words in a sentence in parallel. This fundamentally solved the structural limitations of legacy Recurrent Neural Networks (RNNs), which tended to "forget" earlier context in long passages. At the WMT (Workshop on Machine Translation) 2017–2018 competitions, Transformer-based models improved BLEU scores by over 30% compared to legacy RNNs. For the first time, a quantitative leap in linguistic fluency was measurable and proven.

2020: GPT-3. A model boasting 175 billion parameters debuted. It was qualitatively different from anything that came before it. Yet, its practical utility for the average user remained limited.

Late 2022: ChatGPT. OpenAI added RLHF (Reinforcement Learning from Human Feedback) into the mix. If GPT-3.5 was a "researcher's experimental sandbox," ChatGPT became a "tool anyone could use perfectly on their first try." It hit 10 million users in 40 days, and 100 million in two months. For context, it took Instagram 11 months to hit that same milestone.

Bubbles form in a vacuum. ChatGPT didn't. It emerged at the exact intersection of 60 years of foundational research, a decade of deep learning acceleration, the Transformer architectural breakthrough, and RLHF. A technology that shatters a threshold of this magnitude does not simply regress back below it.

Scaling Laws: The Brutal Physics of Getting Smarter — And the Inevitability of Oligopoly

The second key to unlocking the ChatGPT phenomenon is the Scaling Law. Empirically validated in a 2020 paper by OpenAI's Jared Kaplan, Dario Amodei, and others ("Scaling Laws for Neural Language Models"), the premise is ruthlessly simple: A model’s performance scales non-linearly as you simultaneously increase three variables: Parameters, Data, and Compute.

To put numbers to it: If you increase a model's parameters by 10x, the required data and compute to train it increase by roughly 1,000x. This is the core asymmetry of the Scaling Laws. DeepMind’s 2022 Chinchilla research further refined this, revealing that for compute-optimal training, the ratio of model parameters to training tokens should be roughly 1:20.

Beyond predictable performance gains, the Scaling Laws trigger another phenomenon: Emergence. Formally articulated in a June 2022 paper by Google’s Jason Wei et al. ("Emergent Abilities of Large Language Models"), the concept states: "Abilities that do not exist in smaller models suddenly appear in larger models—and this phenomenon cannot be predicted by simply extrapolating the performance of the smaller models." GPT-3.5 hovered in the bottom 10% of the Uniform Bar Exam. GPT-4 passed in the top 10%. By merely scaling parameters, its mathematical reasoning leaped to the top 1% of the Biology Olympiad. Even its creators couldn't predict that.

To train a model at or beyond the GPT-4 level, you need a dataset equivalent in scale to the entire text of the public internet. Globally, there are perhaps four or five organizations that can absorb this cost. In the foundational AI model race, an oligopoly is structurally inevitable.

Part 2. The First Benchmark of the AI Race — The May 2023 Landscape

The Three Conditions for Victory

Winning the AI race requires three distinct pillars: a proprietary data flywheel, technological velocity, and a monetization backbone. As of May 2023, OpenAI is the only company that possesses all three simultaneously.

The Data Flywheel: The prompts and feedback generated daily by ChatGPT's 100M+ MAUs are fed straight back into model training. More users create more data, more data builds a better model, and a better model attracts even more users.

Technological Velocity: It took two years to go from GPT-3 to GPT-4. Compared to GPT-3.5, GPT-4 reduces the likelihood of responding with disallowed content by 82% and increases factual responses by 40%.

The Monetization Backbone (Microsoft): Microsoft poured $10B+ into OpenAI, securing roughly a 49% stake. GitHub, Azure, Bing, Office 365—OpenAI’s tech is being aggressively integrated into Microsoft's entire product suite, which commands a 90% global market share in office software. This distribution channel cannot be replicated by money or time alone.

The Data Proves It’s Not a Bubble

The MIT Study (Noy & Zhang, 2023): In a randomized controlled trial of 453 college-educated professionals, the group using ChatGPT completed tasks 40% faster while producing work of 18% higher quality. Notably, the most significant improvements were seen among those with lower initial writing skills. AI actually narrowed productivity inequality.

GitHub Copilot (Microsoft, 2023): In a controlled study of developers, the cohort using Copilot completed a Java web server task 55% faster than the control group.

The Competitor Landscape (As of May 2023)

Part 3. Counterarguments and Responses

Hallucinations: The structural limitations of LLMs are very real. However, this does not invalidate the thesis that "AI will replace a significant portion of cognitive labor." The Noy & Zhang study empirically proved that humans partnered with flawed AI still vastly outperform humans working alone.

The "$29B Valuation is a Bubble" Argument: Looking at forward revenue projections for 2024, OpenAI's multiple converges to under 30x Forward PSR. Compared to Snowflake trading at 112x or Palantir at 40x shortly after their IPOs, this valuation is highly defensible for a category king.

The Limits of Scaling: Concerns that we will soon exhaust all human-generated text on the internet are entirely valid. But paradoxically, as this constraint tightens, the relative advantage of players who already own proprietary, closed-loop data sources will only amplify.

Conclusion: The Race Has Begun

This piece is not a declaration that OpenAI will reign supreme forever. It is a snapshot of the first critical question we asked in May 2023: Is AI a structural paradigm shift, and who currently meets the conditions to clear its first gate?

Our judgment is that, as of May 2023, OpenAI was the sole entity checking all three boxes—data flywheel, velocity, and distribution. Google has the firepower to flip the board. If regulatory scrutiny around AI safety intensifies, entirely new layers (like Anthropic) could rise. The very rules of the game may change.

But the most important takeaway is simply this: the game has definitively begun.

[Subsequent Developments] 

The AI race converged into a tight oligopoly of foundation model players faster than expected. OpenAI's ARR surpassed $3.4B by 2025; Google Gemini rebounded with surprising strength; and Anthropic rapidly captured enterprise market share. The core question posed in May 2023—whether the AI race would inevitably consolidate into an oligopoly—proved accurate. The internal ranking of that oligopoly remains an ongoing battle.

[Key Citations]

Kaplan et al., "Scaling Laws for Neural Language Models," arXiv, 2020.01

Hoffmann et al. (DeepMind), "Chinchilla," arXiv, 2022.03

Wei et al. (Google), "Emergent Abilities of Large Language Models," arXiv, 2022.06

Vaswani et al., "Attention Is All You Need," NeurIPS, 2017

Noy & Zhang, "Experimental Evidence on the Productivity Effects of Generative AI," SSRN, 2023.03

This article does not constitute investment advice.