Inside Zipcode and Ditto’s Push to Build Bittensor’s Core Infrastructure

The latest Bigtensor livestream highlighted how Bittensor is maturing into a network of specialized infrastructure.

Zipcode (SN46), Instant, and Ditto (SN118) each pushed a different frontier, highlighting Bittensor’s growing specialization across capital, inference, and intelligence.

The conversation underscored how the network is evolving through deep vertical expertise rather than broad horizontal expansion.

Reading Bigtensor as One Play

The livestream covered credit primitives, hardware trajectories, benchmarking mechanics, and the coordinated buyback flowing through all three subnets tomorrow.

1. Zipcode’s first vault operationalizes the “Bank of Bittensor” concept: Business lines of credit for vetted subnet owners, structured as on-chain credit agreements where Zipcode custodies subnet keys until the loan is paid off.

The subnet-as-house analogy runs the whole design: register a subnet, borrow against emissions, pay off the loan over time.

2. ZipUSD is pivoting to subnet credit as its yield source: The original plan to fund real estate mortgages hit long sales pipelines, so ZipUSD is now backed by lines of credit issued to vetted subnet owners.

Yields land in the 10-25% range, and holders can elect to receive their yield in $SN118 or other subnet tokens rather than dollars.

3. Instant subnet to launch with GPT-OSS inference as the entry point: The first shipment is straightforward inference on the open GPT-OSS 20B and 120B models, with token-per-second benchmarks setting the floor for miner emissions.

4. Instant is designed to walk miners from consumer GPUs down to custom silicon: The benchmark ladder ratchets up from 4090s to 5090s to H100s to H200s to FPGAs and eventually custom-designed chips. Miners get incentivized to keep pushing the hardware boundary.

5. Managed VNet IR (Virtual Network Integration Runtime) removes the TEE requirement for custom miner hardware: The team brings this pattern from his Microsoft Purview background: Instant’s code tunnels into miner infrastructure via a managed virtual network, runs there, and reports token-per-second back.

Miners can bring FPGAs and custom silicon without needing TEE attestation.

6. The endgame is analog computing chips like Talus: Models baked directly into silicon via resistors, running at the speed of physics rather than through binary translation layers.

Talus already runs Llama 3.1 8B on a 6nm chip, and the long-term vision is a data center compressed onto a single chip.

7. Ditto V8 shipping with generative universes: Fake users with fake families, businesses, relationships, and travel histories generated on demand.

Miners have consumed over 4 billion tokens, run 1 million inference requests, and completed 3,000 bench runs so far.

8. Ditto is switching to sequential campaigns instead of parallel competitions: Running three incentive mechanisms at once dilutes miner incentives to one-third each. Sequential campaigns (memory this month, potentially inverse DittoBench next, fine-tuned Ditto LLM after) keep the reward juicy and the miner focus tight. Comparable to World of Warcraft raids.

9. Ditto’s king Kabaw-57 is already beating OpenClaw and Hermes on Ditto Bench: OpenClaw scored 0.226 on memory, Hermes scored 0.136, Kabaw-57 leads Ditto’s own leaderboard.

The LongMemEval comparison test lands soon with all three agents running on the same base model.

10. Ninja, now Conjectures (SN66), is passing its research to Ditto: A key learning: private held-out datasets do not protect against miner probing, because miners send probing submissions to extract structure from private questions. The team is choosing to keep everything open and rely on iteration instead.

Capital, Intelligence, Inference

What tied the entire livestream together wasn’t any single announcement but the shared philosophy behind them.

Zipcode (SN46), Ditto (SN118), and Instant each went deeper into a narrowly defined problem space instead of chasing broader functionality, reinforcing a future where Bittensor grows through specialization.

If that vision holds, capital, intelligence, and inference won’t just coexist on the network; they’ll become foundational infrastructure layers built by dedicated subnets.

NB: Instant is an upcoming subnet with netuid yet to be announced