Why are global superpowers and the world’s most powerful tech giants unanimously backing open-weight AI? Are they suddenly benevolent rather than profit-seeking, or have they recognized a pattern repeated throughout computing history? They back open-weight AI because open-source systems consistently out-adapt gated alternatives once baseline utility is reached. No single R&D team can match the speed of millions of developers optimizing, pruning, and deploying models for specialized hardware and niche workflows. Betting against open weights is betting that a single vendor can out-innovate the rest of the world combined. Open-weight models win on distribution because they turn AI from a rented API into owned, programmable code. When developers can fine-tune weights for specialized tasks and deploy them directly on the compute of their choice, they can eliminate artificial rate limits, data exposure, and vendor lock-in. As millions of builders optimize models for targeted workloads, the collective capability of the open ecosystem outpaces what any centralized provider can maintain. What does this mean for the future of AI infrastructure? As open-weight capabilities approach parity with closed APIs, model weights become commoditized software. Value shifts from model ownership to the execution layer, which is the decentralized compute that powers inference. Global leaders back open weights because they recognize that long-term operational leverage belongs to whoever hosts scalable execution while allowing builders total customization control. The future belongs to open infrastructure that delivers unconstrained, private, and resilient execution at global scale. That is where we align. We are providing the decentralized network necessary to power the world's open AI ecosystem that world leaders are openly promoting.

One of the biggest improvements we've made to subnet consensus is moving away from a pure node 1:1 weighted model toward stake-weighted consensus. Previously, proposals were primarily evaluated based on how many validator nodes participated. This treats every validator equally regardless of the amount of delegated stake they secure. With this update, consensus becomes economically weighted. Each validator's delegated stake is allocated across the subnet nodes they own. When a proposal is created, those stake allocations are snapshotted and remain fixed for the lifetime of that proposal. As attestations come in, the network measures approval by the percentage of delegated stake represented, not just the number of nodes voting. This better aligns governance with economic security. A validator responsible for securing significantly more stake now has proportionally more influence than one with minimal delegated stake, making consensus reflect the actual security backing the subnet rather than simply counting identical votes. Importantly, this doesn't eliminate node participation requirements. We intentionally kept a minimum node-count quorum alongside stake weighting. Even if enough stake is represented, consensus still requires participation from a sufficient number of validator nodes. This prevents consensus from being dominated by only a handful of heavily staked nodes and preserves decentralization and liveness. The result is a hybrid model that combines: • Economic security through stake-weighted attestations. • Decentralization through minimum validator participation. By snapshotting stake at proposal creation, the system also prevents stake movements during voting from changing proposal outcomes, ensuring every proposal is evaluated against a stable and deterministic view of the validator set. This change makes subnet consensus significantly more robust by aligning voting power with the security validators actually provide, while still maintaining the distributed participation guarantees that keep the network healthy.

Another update complete: Frameworkized Subnet Template The subnet template has evolved from a reference implementation into a developer framework for building decentralized AI networks on Hypertensor. This update introduces reusable framework components that handle the common decentralized infrastructure layer, allowing developers to focus on application logic. What's included: • Reusable server framework for P2P networking, peer discovery, consensus startup, telemetry, P2P connection management, and node lifecycle • Merkle DAG framework with signed immutable DAG nodes, multi-head support, state materialization, synchronization, reconciliation, orphan recovery, and pluggable storage backends • DAG + GossipSub base classes that handle publishing, validation, replication, synchronization, parent selection, and message routing • Reusable request/response protocol framework for P2P stream protocols and DAG synchronization • Network API bridge for external services, AI workers, dashboards, and local applications • Consensus, telemetry, scoring, and runtime utilities for production decentralized AI operations • Example implementations for DAG replication, peer state publishing, commit/reveal workflows, monitoring, and server lifecycle management The goal is simple: Give developers the substrate required to build a decentralized network that handles proof-of-useful-work AI workloads, so that builders can focus on application-layer logic. This provides a reusable foundation for decentralized inference networks, agent systems, marketplaces, data networks, and other distributed AI applications built on Hypertensor.

New update complete: Validator-Based Ownership & Staking This update introduces validators as the primary identity layer under the entire Hypertensor umbrella. Previously, subnet nodes were independently registered, and users' stakes were delegated to them on a per-subnet basis. Now, node operators first register a validator identity that can own and operate nodes across multiple subnets. This makes validators easily identifiable participants to whom users can directly delegate stake. Previously, if a validator operated across multiple subnets, delegators had to discover and stake to each node separately. Now, delegators stake directly to validators and receive emissions generated from all subnet activity associated with that validator. The reputation system also benefits from this change. Instead of reputation being isolated to individual subnet nodes, reputation is now accumulated at the validator level. Validators build a track record across all of the networks they participate in, allowing delegation, rewards, penalties, and consensus performance to contribute to a single reputation profile. Why this matters: • Establishes validators as first-class network participants • Creates a foundation for validators to operate across multiple decentralized AI networks • Aggregates validator reputation, rewards, penalties, and consensus participation • Aligns validator incentives across all subnets they participate in • Enables delegation at the validator level and subnet level, making the data structure more coherent and staking decisions easier for users This update represents a significant optimization of the validator and delegator architecture and its scalability, creating a unified reputation, delegation, and rewards system.

We've been building the decentralized AI economy, the coordination layer for the future of AI. The system has been purposefully iterated in stages. Feature by feature. Here's what we've done: Testnet Vitalik→ Focused on the fundamental consensus mechanism where peers formed consensus based on the LLM transformer blocks they hosted, not a state, but actual AI computation. Testnet Gavin→ Introduced economic mechanisms and refined consensus logic, turning early experiments into an economic system that can coordinate real participants. Overwatch Agents→ We introduced Overwatch Agents, an evolving set of decentralized, staked participants who serve as agents that join each AI network at the P2P level, validate consensus, validate on-chain vs. in-subnet state, and score networks based on decentralization, consensus accuracy, and benchmarking. They act as an evaluation mechanism, benchmarking each AI network. Testnet Tensor→ Focused on testing the full-featured blockchain, including the final consensus mechanism, delegate node staking, and many other iterations of features. P2P AI Template→ A decentralized AI framework for deploying deAI networks using DHTs, gossip, noise encryption, PoS, and more features. This is a blockchain tech stack for building AI networks, but it's purposefully built for AI with no computational limitations, unlike blockchains. This ensures the future of AI is decentralized, reproducible, verifiable, and scalable. This is the development backbone that makes it easy for devs to deploy decentralized AI networks. Testnet Hoskinson→ Testnet Hoskinsons' main purpose was to test the P2P AI template in public with multiple teams and partners running nodes to ensure it was ready for production. All of this leads to one point: We are entering the final phases. We're now actively working with external teams and developers to build the first set of AI networks that will be deployed to the ecosystem. People wonder, what makes an AI network decentralized?: - verifiable decentralization - peer-to-peer execution - trustless coordination Built on the P2P AI Template. Secured by the network. Real deAI.

Testnet Hoskinson is complete! 🎇 All core blockchain infrastructure and decentralized AI networking layer testing have been completed! This milestone marks the transition from foundation-building to ecosystem expansion. What we accomplished: • Over 1,000,000 blocks validated • Over 2,000,000 pubsub heartbeats processed • Over 5,000,000 messages successfully gossiped across the network • Fault-tolerant P2P connectivity validated under real conditions • End-to-end decentralized AI network coordination confirmed The future is decentralized, trustless, and P2P AI, and we've unlocked that. ⇾ Production-grade experimentation for agent swarms and model coordination ⇾ Scalable on-chain incentives for AI infrastructure ⇾ Composable decentralized AI services pooled together across a single network We now move from proving the network… to building on it. The future of AI starts now. The future of AI is trustless. The future of AI is peer-to-peer. The future of AI is private.

Testnet Hoskinson is approaching its final stages. Our primary focus during this phase has been testing the decentralized AI subnetwork template and framework. This included validating peer coordination, gossip propagation, and fault tolerance under continuous operation. We are now transitioning into the next stage of development: deeper blockchain testing. This phase will focus on validator incentives, smart contract deployment, and precompile functionality. One of the key properties we tested in the subnet framework was fault tolerance. Because each subnet uses a technology stack similar to modern blockchain networks, it must meet enterprise-grade reliability standards. During live testing, the subnet has been running uninterrupted with: - 1,900,800+ pubsub heartbeats - 4,561,920+ messages gossiped between peers With subnet template testing nearing completion, development is expanding into additional areas of the protocol, including: - Blockchain validator rewards - Smart contract deployment - Precompile smart contracts This marks the transition from infrastructure validation toward preparing the ecosystem for the first real subnets to deploy. In parallel, we are preparing for the stage that follows testnet, building the first subnets alongside other great teams in the AI space that will launch on mainnet.

Testnet Hoskinson 2.0 is now live. This upgrade introduces several protocol and node-level improvements: Core updates - Delegate Accounts for nodes - New node reputation events - Fixed reputation update bug affecting removed nodes - Improved peer system alignment with modern P2P standards - Expanded functionality in precompiled smart contracts Delegate Accounts Delegate Accounts allow nodes to specify a separate address to receive a configurable portion of node emissions. This enables: - Reward routing to teams or operators - Contract-based logic for emissions - Business and organizational distribution models - Separation of operational and treasury accounts Delegate Accounts are optional and fully controlled by node operators; the delegate account holder can update the delegate account address themselves. This is an infrastructure-level primitive designed to enable more flexible and scalable node operations. A key step toward production-grade node coordination and scalable network operation has been achieved!

DHTs are great for lookups, but they don't scale when many peers repeatedly request the same data. Each GET triggers multiple queries across the network, creating redundant traffic and bottlenecks. While we still utilize a DHT for more efficient and optimized peer connectivity, we're optimizing subnets to scale to thousands and millions of nodes, enabling us to provide decentralized AI networks to billions of users. GossipSub flips the model: data is published once, propagated through a mesh, and stored locally by peers. No re-fetching, lower latency, reduced bandwidth, and predictable load, which is why blockchains rely on gossip for real-time data sharing.

HUGE

Big upgrades are landing 🚀 We're rolling out a new subnet template built around three pillars: Connectivity. Scalability. Security. This is about making decentralized AI networks actually work at scale. 🔗 Connectivity Random Walk peer discovery Nodes now passively explore the network by querying random keys in the Kademlia DHT. Why this matters: • Continuous peer discovery • Fresh routing tables so nodes always know which peers are reachable and how to find them • No central coordination • Stronger long-term connectivity Connection maintenance Each node actively maintains connections with up to 64 peers, ensuring sufficient overlap and resilience across the subnet. ⚡️ Scalability & Efficiency The old model leaned heavily on DHT records as a global database. That doesn't scale. When every peer queries everything, everyone pays the cost. Enter GossipSub. We've implemented GossipSub, a peer-to-peer pubsub protocol that forms a stable mesh for real-time message propagation. What changes: • Messages are shared once, then gossiped • No constant record fetching • Lower latency • Dramatically reduced load This is how blockchains efficiently broadcast transactions, now applied to subnets. 🔐 Security In addition to proof-of-stake, we've incorporated Noise into the networking stack. Noise establishes encrypted, authenticated peer-to-peer channels with forward secrecy via cryptographic handshakes. Forward secrecy (or Perfect Forward Secrecy) ensures that even if a node's long-term private keys are compromised in the future, past session data remains secure. Secure by default. High performance. No trade-offs. This is the future of AI. P2P. Trustless. Decentralized.

Testnet Hoskinson Update ✅Over 250k blocks processed ✅Over 2,500 epochs We're not stopping! We're updating the subnet template: 🔹 P2P connection maintenance to ensure peers are always connected to at least 1/4 of the maximum number of peers 🔹 Introducing a gossipsub protocol for P2P communication 🔹 Adding random walk functionality to the KAD-DHT 🔹 Adding support for Secp256k1 and ECDSA on top of Ed25519 and RSA 🔹Introducing a noise protocol for secure P2P traffic Adding additional features to the blockchain: 🔹Introducing delegate accounts to nodes for business logic 🔹More events These updates will be pushed and tested in the coming weeks.

👀

🤫

We named one after him too!

It is!

🤗

👏👏

Thanks Scott! 😁

Thanks for helping out!