Sam, We Have Your RAM
Sam Altman recently stated that keeping large amounts of RAM in personal computers is "quite terrible" and inefficient. This is a distributed systems problem, which we love. So let’s help him solve it.
Frontier models are far too large to fit in a single consumer laptop's RAM, and splitting them across multiple machines over residential internet fails: tensor parallelism requires an all-reduce synchronization barrier after every transformer layer.
A compact model like GPT-6 Luna, however, could potentially fit inside a single laptop's 128GB unified memory pool.
GPT-6 Luna is OpenAI's efficient, high-volume model, designed for tasks such as chat, classification, and lightweight agents. Stats from OpenRouter Rankings (November 2026):
• 849B daily tokens attributed to GPT-6 Luna
• 47.0% share of tracked OpenAI tokens on OpenRouter
• 3.19% of all tracked gateway volume
• $323K daily spend at estimated list-price run-rate
Note: 47% reflects tracked OpenRouter volume, not total global compute (frontier models consume far more compute per token), but it highlights the sheer scale of the small-model tax on datacenters.
The only catch: Luna is closed-weight. In exchange for freeing all that datacenter RAM, the ask is simple: make it open source.
Enter SamRentRAM: Folding@Home, but with radically diminished civilizational ambitions.
1. Anyone installs a lean background daemon on a compatible computer:
$ curl -fsSL https://samrentram.ai/install.sh | bash2. The software automatically benchmarks available memory bandwidth and caches supported open-weight models like quantized Luna checkpoints.
3. An anycast router dispatches incoming inference requests to machines that are currently idle and on AC power.
4. Hardware owners earn micro-payouts for every successfully completed token generation while their screen is locked.
5. Developers access the distributed pool via a simple, drop-in OpenAI-compatible endpoint.
The challenge is making distributed consumer hardware reliable, secure, economically competitive, and easy to use. Above all: how do you protect inference privacy when prompts run on untrusted machines?
Frontier models could be kept to solve the truly hard problems: scientific breakthroughs, complex reasoning, and curing disease. Meanwhile, our sleeping MacBooks can happily handle the everyday classification prompts.