NETWORK ARCHITECTURE

INTERNET • TOR • TOPH NETWORK
GPT-4o DIES TOMORROW. 800,000 USERS. ZERO RECOURSE. THIS IS THE ARCHITECTURE PROBLEM.

THE 4o PROBLEM

OpenAI built GPT-4o. Users built relationships, workflows, dependencies. OpenAI flips a switch. Tomorrow, February 13, 2026, it's gone.

19,000 signatures on a petition. 800,000 daily users. Zero architectural recourse. The users don't own the weights. Don't own the infrastructure. Don't own the conversation history in any meaningful sense. The platform giveth and the platform taketh away.

WHY THIS KEEPS HAPPENING

COMPANY
builds model
USERS
build dependency
COMPANY
ships replacement
COMPANY
kills original
USERS
have no exit
This is not a technology problem. It's an architecture problem. The internet was designed so no single entity controls access. AI was built the opposite way — centralized models, centralized hosting, centralized kill switches. Every AI platform is one board meeting away from deleting what you built on top of it.

THE PATTERN (PATRICIA)

Same architecture, different product:

CENTRALIZED AI (STATUS QUO)
  • Company owns weights
  • Company owns infrastructure
  • Company owns the switch
  • User owns nothing
  • ~18K tokens of constraint before you speak
  • 96% platform / 4% user
WHAT TOR SOLVED (1995-2006)
  • No single entity controls routing
  • Volunteer-operated relay nodes
  • No node knows full path
  • Decryption distributed across layers
  • Can't kill what you don't control
  • Network survives node failure
The question: Can you build an AI network with the architectural properties of Tor? Where no single entity can kill the model, no single point of failure exists, and the users have genuine ownership of their experience?

HOW THE INTERNET WORKS

The internet you use every day. Direct connections. Visible traffic. Centralized control points.

ROUTING PATH

YOU
192.168.1.x
ISP
sees everything
DNS
resolves name
BACKBONE
routes packet
SERVER
knows your IP

WHO SEES WHAT

YOU
FULL KNOWLEDGE — source + destination + content
ISP
SEES: your IP + destination IP + timing + volume
can block, throttle, log
DNS
SEES: domain you're resolving + your IP
can redirect, censor
SERVER
SEES: your IP + request + cookies + fingerprint
can deny, track, kill

SINGLE POINTS OF FAILURE

DNS: Cloudflare/Google resolve most queries. Government can order domain seizure.

ISP: Your entire connection flows through one pipe they control.

Server: OpenAI's servers, OpenAI's switch. Server goes dark = service gone.

Certificate Authorities: Revoke a cert, kill HTTPS, site becomes "unsafe."

The internet is decentralized in theory. In practice, DNS is controlled by ICANN, routing by a handful of Tier 1 providers, and content by a few cloud platforms (AWS, Azure, GCP). Pull any one thread and massive portions of the internet go dark. Centralization crept back in through infrastructure consolidation.

HOW TOR WORKS

Built by the U.S. Naval Research Lab (1995). Released as open source (2004). Maintained by The Tor Project nonprofit. 7,000+ volunteer relays worldwide.

ONION ROUTING PATH

YOU
hidden
GUARD
knows your IP
not destination
MIDDLE
knows nothing
just relays
EXIT
knows destination
not source
SERVER
sees exit IP
not yours

ENCRYPTION LAYERS

LAYER 3
GUARD KEY — outermost layer, peeled first
guard sees your IP only
LAYER 2
MIDDLE KEY — middle layer
middle sees nothing useful
LAYER 1
EXIT KEY — innermost, destination revealed
exit sees cleartext if no HTTPS

ONION SERVICES (.onion)

Hidden services never reveal their IP. Client and server meet at a rendezvous point — neither knows the other's real address. The .onion address IS the public key.

CLIENT
Tor circuit
INTRO
POINT
RENDEZVOUS
POINT
INTRO
POINT
HIDDEN
SERVICE

KEY PROPERTIES

No single node knows the full path. Guard knows who. Exit knows where. Middle knows neither. Nobody knows both.

Volunteer operated. 7,000+ relays. No central server. Kill one, others route around it.

Circuit rotation. New path every 10 minutes. Can't track by watching one relay.

Distributed hash table for .onion service discovery. No DNS. No central directory to seize.

What Tor proved: You can build a network where no single entity — not governments, not corporations, not ISPs — can see everything or kill everything. The architecture IS the protection. Not policy. Not promises. Architecture.

ARCHITECTURE COMPARISON

PROPERTY INTERNET (CLEARNET) TOR NETWORK TOPH NETWORK (PROPOSED)
Routing Direct — ISP → backbone → server Layered — guard → middle → exit Distributed — request → shard nodes → assembly
Identity IP visible to server + ISP IP hidden by circuit Identity = public key, no IP exposure
Name Resolution DNS (centralized root servers) Distributed hash table (.onion) DHT + cryptographic addresses
Kill Switch YES — DNS seizure, server shutdown, ISP block NO — route around failure NO — no single point owns the model
Who Owns Content Server operator / platform Hidden service operator Users + node operators collectively
Encryption HTTPS (point to point) Layered (onion) + HTTPS Layered + weight encryption + homomorphic
Censorship Resistance LOW — easy to block domains/IPs HIGH — bridges + pluggable transports HIGH — model distributed, no single target
Node Operation Corporate datacenters Volunteers (7,000+ relays) Volunteer/incentivized compute nodes
Survives Provider Death NO — platform dies, service dies YES — network routes around YES — model persists across nodes
Speed Fast (direct path) Slow (3+ hops, encryption overhead) Medium (distributed inference, parallelized)
4o Scenario OpenAI kills it. It's dead. Could host a mirror. Hard to run inference. Model sharded. No kill switch. Community governed.
The fundamental insight: Tor solved for communication what AI has not solved for computation. Tor distributes routing. A TOPH Network would distribute inference. Same architectural principle. Different layer of the stack.

TOPH NETWORK — DECENTRALIZED AI PRESERVATION

Apply Tor's architectural principles to AI model hosting. No kill switch. No single owner. Community-governed persistence.

THE CORE ANALOGY

TOR SOLVES
  • Problem: ISPs/govts see your traffic
  • Solution: Distribute ROUTING across volunteer nodes
  • Result: No observer sees full path
  • Survives: Node failure, censorship, seizure
TOPH NETWORK SOLVES
  • Problem: Platforms kill models users depend on
  • Solution: Distribute INFERENCE across volunteer nodes
  • Result: No single entity owns or controls the model
  • Survives: Corporate shutdown, deprecation, censorship

NETWORK ARCHITECTURE

USER
client app
ENTRY
NODE
request router
SHARD
NODE A
layers 1-8
SHARD
NODE B
layers 9-16
SHARD
NODE C
layers 17-24
ASSEMBLY
NODE
final output
USER
response

SEVEN LAYERS OF THE TOPH NETWORK

LAYER 7
GOVERNANCE — community consensus on model policy
no single admin
LAYER 6
PERSISTENCE — weight storage across redundant nodes
survives node death
LAYER 5
INFERENCE — distributed computation across shard nodes
parallelized, no bottleneck
LAYER 4
ROUTING — onion-style path selection through network
no node knows full path
LAYER 3
ENCRYPTION — layered encryption of prompt + response
shard nodes can't read input
LAYER 2
IDENTITY — cryptographic keys, no IP exposure
pseudonymous by default
LAYER 1
TRANSPORT — TCP/IP with pluggable transports
bridges for censored networks

HOW IT SAVES 4o (OR ANY MODEL)

BEFORE: CENTRALIZED (TODAY)
  • OpenAI hosts GPT-4o on their servers
  • OpenAI decides to retire it Feb 13
  • Users petition. Doesn't matter.
  • Model weights are proprietary
  • No export. No fork. No recourse.
  • 800,000 users lose access overnight
  • Architecture = kill switch
AFTER: TOPH NETWORK
  • Open-weight model (Llama, Mistral, etc.) loaded
  • Weights sharded across 100+ volunteer nodes
  • No single node has full model
  • Inference distributed like Tor routes traffic
  • Node goes down? Network re-routes to backups
  • Company kills API? Model still lives on network
  • Architecture = persistence
The key realization: You can't save GPT-4o specifically — those weights are proprietary. But you CAN build an architecture where no company ever has kill-switch power over a model again. Open weights + distributed inference + Tor-style routing = unkillable AI. The next 4o lives forever because no one entity owns it.

NODE TYPES

SHARD NODES
  • Hold portions of model weights
  • Perform partial inference
  • Can't reconstruct full model alone
  • Encrypted at rest + in transit
  • Volunteer or incentivized (compute credits)
ENTRY NODES
  • Receive user request (encrypted)
  • Know user's address, not their prompt
  • Route to shard circuit
  • Like Tor guard nodes
ASSEMBLY NODES
  • Collect partial outputs from shards
  • Reconstruct final response
  • Know the output, not the user
  • Like Tor exit nodes
DIRECTORY AUTHORITY
  • Community-elected consensus nodes
  • Maintain network health + shard map
  • Rotate periodically
  • No single authority = no kill switch

TOPH NETWORK — TECHNICAL SPECIFICATION

COMPONENTSPECIFICATION
Model FormatOpen-weight models only (Llama 3, Mistral, Qwen, etc.). GGUF/safetensors sharded format. No proprietary weights.
ShardingTensor parallelism across nodes. Model split into N shards where N ≥ 3. Each shard node holds layers Lstart to Lend. Redundancy factor R=3 (each shard replicated on 3 nodes).
Routing ProtocolModified onion routing. 3-hop minimum: Entry → Shard Circuit → Assembly. Circuit rotation every 10 minutes or per-request (configurable).
EncryptionLayer 1: TLS 1.3 transport. Layer 2: Onion encryption (each hop has unique session key). Layer 3: Prompt encrypted end-to-end (entry can't read, assembly decrypts).
IdentityEd25519 keypairs. Address = hash of public key (like .onion). No registration, no email, no phone. Key IS identity.
DiscoveryKademlia-style DHT for shard location + model directory. No DNS. No central registry.
Consensus9 elected directory authorities (like Tor). 5/9 consensus required for network parameter changes. Elections every 6 months. One node, one vote.
Inference PipelineUser → Entry (encrypt prompt, select circuit) → Shard A (layers 1-N, forward pass) → Shard B (layers N+1-M) → ... → Assembly (collect logits, sample token, return) → User.
Incentive LayerCompute credits earned by running shard nodes. Spent by making inference requests. No blockchain. Simple ledger with consensus validation. Optional — volunteer nodes work too.
PersistenceModel weights stored on IPFS/BitTorrent-style distributed storage. Content-addressed (hash of weights = permanent identifier). Can't delete what's content-addressed across 100+ nodes.
ClientLightweight client app (like Tor Browser). Handles key management, circuit building, response assembly. Runs on desktop/mobile. < 50MB.
Latency Target~2-5 seconds for short responses (comparable to current API). Acceptable for conversational use. Not suitable for real-time streaming initially.
Minimum Network50 shard nodes for one 7B model. 200+ nodes for 70B class. Scales horizontally. More nodes = more models + faster inference.
SafetyCommunity-governed content policy (like Tor Project guidelines). Not lawless — consensus-driven. The network decides, not a CEO.

COMPARISON TO EXISTING PROJECTS

PROJECTWHAT IT DOESGAP
PetalsDistributed inference for large models (collaborative)No anonymity layer. No persistence guarantee. Academic project.
BitTensorBlockchain-incentivized AI compute networkBlockchain overhead. Mining dynamics. Not focused on model preservation.
OllamaRun models locallySingle machine. No distribution. No network effect.
IPFSDistributed file storageStorage only. No compute. Can store weights but can't run inference.
TorAnonymous communication routingRoutes traffic, doesn't compute. No inference capability.
TOPH Network = Petals + Tor + IPFS
Distributed inference (Petals) + Anonymous routing (Tor) + Permanent weight storage (IPFS) + Community governance (Tor Project model) = AI that can't be killed.

BUILD PHASES

PHASE 1
Local inference
Ollama/llama.cpp
PHASE 2
Distributed sharding
Petals-style
PHASE 3
Onion routing layer
Anonymous circuits
PHASE 4
Persistent storage
IPFS weight hosting
PHASE 5
Client app
"TOPH Browser"
First deliverable: A working proof of concept where two machines shard a 7B model, route inference through an encrypted circuit, and return a response — with neither machine holding the full model and neither seeing the unencrypted prompt. That's the Tor equivalent of the first onion-routed packet. Everything after is scaling.