Streamlining Constellation Network for Data and Inference

In his post today, Ben lays out where Constellation is going: the whole network aimed at AI, with operators paid in DAG for compute and inference work. Getting there means building something new and closing some things this community has built and run for years. I know that for many of you these changes land close to home, and you deserve the full picture: where we see the world going, what we are building, every change we are making, when each one takes effect, and what it means for you.

Where the world is going

AI is advancing faster than any technology this industry has dealt with. Models now write and read code at a level that was out of reach two years ago, and demand for the compute to run them keeps climbing.

Keeping funds safe has become more challenging. The tools that read a codebase and find a flaw in hours are available to anyone, and open source code is open to them by definition. DeFi protocols are the most exposed, because they hold other people's funds where a single bug can reach them. Attacks on DeFi are rising across the industry, and the protocols that survive are the ones with large security teams and continuous audits. We believe the result will be a steady consolidation of DeFi onto a few chains with the resources to defend it. Small teams running their own exchanges will not keep up.

Trusted data has become more valuable. As more of what we read and build comes from machines, the record of what a model was asked, what it answered, and who did the work becomes essential. That is what Constellation was built for. Data integrity has been our purpose since day one, and an AI-driven world needs more of it every year.

Both consequences point the same way. Put our effort into data and inference, where demand is growing, and take it out of DeFi, where the risk grows faster than any small team can defend.

The inference network

We are building an inference network on Constellation. It runs as a metagraph, uses DAG as its currency of exchange, and lets people with their own hardware serve AI models and be paid for the work.

Most AI today runs in a handful of data centers, and demand is growing faster than those data centers can grow. At the same time a great deal of capable hardware sits idle in homes and offices: gaming GPUs, Apple Silicon Macs, workstations. Open models now run well on that hardware. We believe the people who own that hardware should be able to share it and be paid for it, the way a household sells solar power back to the grid. Scarce compute gets used instead of sitting idle.

Here is how it works. You install a small application on a machine you already own. It downloads the open models you choose to serve and announces itself to the network. Constellation Gate AI, our AI platform, routes requests from its customers to operators serving the right model. The network records each unit of work and pays each operator in DAG. Gate is the first customer, and every request it sends to an operator is paid for by a Gate user.

This opens the network up. Today you need a server, 250,000 DAG and a place on the validator allowlist to take part. In the inference network you need a computer with a capable GPU or an Apple Silicon chip and a good connection. There is no allowlist and no cap on how many can join. The work is paid for by real customers, and it runs on hardware people already own rather than in rented cloud data centers.

For DAG, this is the job Ben described. Compute flows through the network, the value of that compute flows through DAG, and operators are paid in it for work they delivered. It is our primary focus from here.

We will open the network to a group of beta operators before the public launch. If you have hardware you would like to contribute and want to be among the first, tell us here and we will reach out when we are ready.

Changes to the network

Building that network well means running a smaller, more focused one. Constellation grew to carry an exchange, test networks with over 100 nodes, a rewards platform and more than 200 mainnet operators, all funded from emission and hosted at operators' cost. Most of it was built for an ecosystem of outside teams and DeFi activity, which is not where the network is going. Trading, swapping and the other DeFi primitives belong on purpose-built networks with the security teams to defend them, and as attacks on DeFi grow each year that matters more. The changes below bring the network down to what it needs for data and inference and remove the parts that were costing the most to run and to defend.

• PacaSwap is closed and will not reopen. Liquidity providers are paid out in full and SWAP is reissued on Base.
• Testnet and IntegrationNet are retired on October 15th.
• Mainnet runs on 30 active validators, chosen each quarter by delegation, from October 15th.
• The emission schedule changes once a new one is ratified with the Stardust Collective.
• The Lattice frontend closes on October 15th.
• Inactive metagraphs are shut down.
• DAG balances, the data metagraphs and the core protocol do not change.

PacaSwap is closed for good

PacaSwap has been offline since an exploit on September 9th. We restored the funds taken, and it stays closed. Its positions are settled as follows.

If you provided liquidity, you receive your share of every pool you were in, in DAG, DOR, UP and USDC.dag, as the pools stood at the final snapshot on September 9th. The payout goes to the address you deposited from. A table showing what each address receives will be published on the PacaSwap site by October 1st, and payouts are complete by the same date. You do not need to sign anything.

If you hold or have staked SWAP, the same balance is reissued as an ERC-20 token on Base and sent to you on October 1st. Balances in governance locks and unclaimed rewards count at face value. The private key behind your DAG address controls the same address on Base, and the tokens arrive there with no action from you. Stargazer supports Base, so they appear in the same wallet you use for DAG. Any other Base wallet, such as MetaMask, works too if you import the same key.

The SWAP contract has a fixed supply equal to the clean snapshot total, no owner and no admin key. Governance voting power ends with the metagraph.

Testnet and IntegrationNet are retired

Both networks stop on October 15th and their share of mainnet emission ends. Public test networks were built for a steady stream of outside teams launching metagraphs. From October the protocol team runs smaller networks as it needs them for testing and integration.

If you run a test network node, rewards are paid through October 14th so you have time to cancel your hosting. Shut your node down from October 15th. There is nothing to withdraw.

Mainnet runs on 30 active validators

Mainnet has over 200 operators on its list and does not carry the traffic to justify that size. It runs securely and far more efficiently with 30 active validators, and a set that small can be upgraded in one coordinated release. From October 15th, the DAG community decides which nodes make up that set, through delegation.

Any node can qualify by holding at least 1,000,000 DAG in its wallet, up from 250,000 today. Every qualifying node can receive delegated DAG. Each quarter, qualifying nodes are ranked by the total DAG delegated to them, and the top 30 become the active validators for that quarter and earn validator rewards. The ranking runs again every quarter, so the active set follows where the community places its stake.

The first ranking is announced on October 9th and takes effect on October 15th. If you run a mainnet node and want to be considered, make sure its wallet holds 1,000,000 DAG before October 9th. If your node makes the 30, install the release when it ships and keep running. If it does not, it stops earning validator rewards on October 15th. It stays qualified as long as the wallet holds 1,000,000 DAG, and it can join the active set in a later quarter as its delegation grows.

If you delegate, your delegation is now your vote for which nodes are active, as well as your stake. It keeps earning rewards whether or not the node you chose is in the 30. You can move a delegation to another node at any time without withdrawing it first.

Emission moves to a new schedule

The current emission schedule was designed for a large operator set and does not fit a smaller network. Over the coming weeks we will work with the Stardust Collective to ratify a new schedule, with reduced inflation and fixed emissions as the goals. Validator rewards and delegator yield are set through that process, and we will publish them once the Collective has voted.

The Lattice frontend closes

Lattice has been part of this network for years, first as a launchpad, then as the DeFi hub of the ecosystem, and most recently as the channel that paid Testnet and IntegrationNet operators and as the bribing platform for PacaSwap, bribing in the DeFi sense of the word. Each of those roles ends with the changes above. We are shutting down the Lattice frontend on October 15th and ending our support for it.

LTX and its contracts on Ethereum are untouched, and LTX governance remains active. The LTX community is free to take that organization in a different direction if it chooses. If you staked in one of the EVM staking programs on Lattice, you can withdraw at any time from that program's contract. If you hold veLTX, the locked form of LTX that carries voting rights, you can withdraw your LTX when your lock ends. We keep one Lattice page online for these withdrawals and for LTX governance.

Inactive metagraphs are shut down

Metagraphs that are no longer producing snapshots are removed from the network. Their allowlist entries are cleared and their nodes are taken offline. The data metagraphs in production are not affected.

What does not change

Your DAG balance, the Hypergraph, the data metagraphs, transfers and DAG's listings on existing exchanges all stay as they are. Nothing in this set of changes touches the token or what you hold.

Where this leaves us

This is a lot of change at once. For many of you it touches something you have been part of for years, and I am grateful for everything you have put into this network. We chose to make these changes now, in one step, rather than let them arrive slowly.

What we get for it is a network with far less to defend and all of its effort on the two things it was built for and the world needs more of: trusted data and the compute to produce it. The inference network turns a machine in your home into part of that, paid in DAG for work real customers asked for. That is the most exciting thing I have worked on at Constellation, and these changes are what make it possible. I hope you will build it with us.

---

This article reflects our current vision and plans, which may evolve as we execute. It is intended as general information, not investment advice or a solicitation. Statements about future plans involve risks and uncertainties, and actual outcomes may differ.