Why Isn't Kimi (or Any AI Company) Building a P2P Distributed AI Network? Here's My Plan — Tear It Apart

Kimi K3
1

Hey everyone,

I’ve been thinking about something that honestly confuses me, and I want to put it out here for discussion.

The Observation

Right now, every AI company — Kimi, OpenAI, Anthropic, Google — follows the same model: massive data centers, billions of dollars in GPU clusters, and users pay subscriptions or API fees to access them. Meanwhile, there are hundreds of millions of powerful PCs sitting idle worldwide — gaming rigs with RTX 4090s doing nothing at 3 AM, developer workstations with 32GB RAM running a text editor.

Why isn’t anyone connecting all of that?

My Plan: TithNet

I’m working on a concept called TithNet — a peer-to-peer AI inference network. The idea is simple:

  1. User installs a desktop app (Windows/Mac/Linux)

  2. The app runs a lightweight node in the background, contributing idle GPU/CPU power to the network

  3. In return, the user gets free AI — chat, code assistance, image generation

  4. The more powerful your hardware, the better your experience (faster responses, priority routing)

The philosophy: VLC for AI. Install it, use it, free forever. No subscription, no API key, no account required.

Technical Approach (High Level)

  • Model sharding across multiple peer nodes — no single PC needs to hold the full model

  • Smart routing — the network finds optimal nodes based on latency, GPU capability, and load

  • Contribution-based tiers:

    • Bronze (4GB RAM, basic CPU) → standard speed

    • Silver (16GB RAM, decent GPU) → fast

    • Gold (RTX 3060+) → very fast

    • Diamond (RTX 4080/4090) → near-instant

  • Encryption in transit — prompts don’t sit on any central server

  • Local fallback — a small on-device model handles basic tasks offline

  • Single app includes AI chat + a built-in code editor (Monaco-based) with autocomplete and AI-assisted debugging

Monetization (Without Charging Users)

  • Sponsorships from the developer/open-source community

  • Grants (AI research, open-source foundations)

  • Users stay free forever — that’s non-negotiable

My Question to Kimi / Moonshot

Kimi already has strong models (K3 is impressive). You already have a developer community. You already have infrastructure. So why not explore a hybrid approach?

Imagine if Kimi offered a desktop app where users could optionally contribute idle compute, and in return get unlimited free access or boosted quotas. You’d solve two problems at once: reduce your inference costs AND grow a loyal, sticky user base that literally powers the network.

Is there a fundamental reason this doesn’t work that I’m missing? Is it a technical limitation, a business decision, or just something nobody has prioritized yet?

Where I Need Honest Criticism

I’m not posting this to promote anything — I genuinely want to know what’s wrong with this plan. Here are the areas I suspect might have problems, but I want real developer feedback:

  • Latency: Can model sharding across consumer internet connections actually deliver acceptable response times? Data center interconnects run at 400Gbps+. Home internet is 50-500 Mbps. Is the gap too large?

  • Reliability: Nodes go offline unpredictably. How do you maintain inference quality when your “servers” can disappear mid-generation?

  • Model quality ceiling: Consumer GPUs can run 7B-13B parameter models well. But the models people actually want (GPT-4 class, K3 class) are much larger. Can sharding across consumer hardware realistically run 70B+ models?

  • Security: Even with encryption, the actual tensor computations happen on someone else’s machine. Is there a real privacy guarantee possible here?

  • Incentive problem: BitTorrent works because seeders already have the file. In AI inference, contributing costs real electricity. Is “free AI” a strong enough incentive for people to leave their PCs running?

  • Coordination overhead: The routing/scheduling layer itself needs reliable infrastructure. Doesn’t that defeat the purpose?

  • Trust: How do you verify that a node returned correct results and not garbage? Byzantine fault tolerance for AI inference seems unsolved.

What I’m Looking For

  • Technical feedback from anyone who has worked on distributed inference, federated learning, or P2P systems

  • Honest opinions from the Kimi team or community on whether this direction has been considered internally

  • Criticism I haven’t thought of yet — the harder, the better

I’d rather have my plan destroyed by honest feedback now than waste years building something fundamentally broken.

Thanks for reading. Looking forward to the discussion.