Leaked Google memo warns: 'We have no moat'

Google and OpenAI are locked in a public slugfest, but an internal memo from Mountain View just threw cold water on both of them: there is no moat, and you're fighting the wrong war.

A document recently leaked from a Google researcher and published by SemiAnalysis lays out a harsh reality. While the two tech titans have been obsessing over each other's every move, open-source developers have been quietly eating their lunch. Independent developer Simon Willison noted that whether genuine or not, it stands as one of the most compelling analyses of the LLM landscape yet.

$100 models vs. $10M monsters

The central argument is brutally simple: proprietary models are losing their edge at blinding speed.

According to the researcher, open source is doing with $100 and 13 billion parameters what Google struggles to achieve with $10 million and 540 billion parameters. And they are doing it in weeks rather than months. People are already running foundation models on a Pixel 6 at 5 tokens per second, while personalized AI can be fine-tuned on a beefy laptop in an evening.

The turning point arrived in early March when Meta's LLaMA leaked. It lacked instruction tuning and RLHF, but the open-source community pounced anyway. Within weeks, projects like Alpaca and Vicuna 13-B surfaced, stacking human evaluation, multimodality, and quantization on top of each other. The barrier to frontier experimentation dropped from a corporate supercomputing cluster to a single motivated developer.

The LoRA revolution

The memo pinpoints why big tech's playbook is failing: training giant models from scratch is the hard path, and LoRA changed the rules.

Low-Rank Adaptation allows model updates to be shrunk by thousands of times. Fine-tuning runs cost roughly $100 and take under a day on consumer hardware. Crucially, these improvements are stackable. One person trains instruction following, another adds reasoning, and a third layers on tool use. These cheap updates accumulate without throwing away earlier work.

By contrast, spending millions to retrain a massive model from zero wipes out prior iterative gains. Coupled with the realization that curated synthetic data scales better than sheer data volume, the open-source world is outpacing corporate research labs through decentralized, breadth-first iteration. As the author notes, this mirrors how Stable Diffusion culturally eclipsed DALL-E.

Why it matters

This memo tears through the industry hype and exposes the structural vulnerability of closed-source AI.

For developers and individual tinkerers, it confirms that you don't need a multi-million-dollar compute budget to build meaningful AI tools. Running small, stackable models locally is not only viable; it is often faster, cheaper, and completely unencumbered by corporate guardrails or API billing.

For the AI industry, it poses an existential commercial question: who will pay for restricted, expensive corporate APIs when unrestricted, free alternatives are practically as good? We think the author is dead right that hoarding weights and training ever-larger monoliths is a losing game. Open-source communities iterate faster because their engineers don't wait for product reviews or legal clearance. Unless companies like Google and OpenAI pivot toward lightweight models, aggressive distillation, and open ecosystem collaboration, their multi-million-dollar leads could evaporate before the year is out.