With the amount of new subnets being added it can be hard to get up to date information across all subnets, so data may be slightly out of date from time to time
With the amount of new subnets being added it can be hard to get up to date information across all subnets, so data may be slightly out of date from time to time
Nodexo is Bittensor Subnet 106 (SN106) — a permissionless, decentralised GPU compute network built on the principle that a compute market should be based on cryptographic proof, not on vendor promises. Where centralised cloud providers ask renters to trust that the hardware behind a dashboard is real, Nodexo proves it. GPU operators, called miners, connect machines to the network and continuously demonstrate possession of their silicon through a process called ZkGEMM Proof of Hardware. Each machine is bound to a persistent hardware fingerprint derived from the SHA-256 hash of its GPU device identifiers and system identifier, recorded on-chain at registration. If hardware is swapped, the fingerprint changes and the discrepancy is immediately detected.
Validators — which require only commodity CPU hardware, no GPUs — issue continuous timed challenges to every miner. Each challenge is a nonce-seeded GEMM workload sized to the claimed GPU class, with a response deadline that only the stated silicon can meet. Responses commit outputs under a Merkle root, and validators audit randomly selected leaves, making fabrication computationally equivalent to simply doing the work. Challenges are issued with fresh nonces on an ongoing basis, meaning proof of hardware is a live property the fleet maintains rather than a one-time ceremony. Proof history accumulates into a public reliability score visible to any renter.
Once a machine is verified, it enters the network’s resource routing model. Rental demand takes priority: when a renter requests capacity, verified free machines are instantly available for SSH-based rental. When no rental is pending, idle capacity is automatically routed into managed training workloads so GPU operators keep earning rather than waiting. This design eliminates the false idle-market problem common to other compute networks — rented and busy inventory stays visible as active network state, giving renters an accurate picture of true availability.
Subnet 106 is structured as a protocol with many access points. Because validator proof-checking is CPU-only work, any operator can run a validator and open their own rental storefront over the shared, network-owned fleet. Rather than routing all demand through a single platform frontend, Nodexo exposes every verified machine through any door — meaning competition among storefronts improves service without fragmenting supply. This architecture is deliberately designed for a world where renters include not just human developers but autonomous AI agents that can rent compute programmatically without accounts or stored credentials.
Nodexo is Bittensor Subnet 106 (SN106) — a permissionless, decentralised GPU compute network built on the principle that a compute market should be based on cryptographic proof, not on vendor promises. Where centralised cloud providers ask renters to trust that the hardware behind a dashboard is real, Nodexo proves it. GPU operators, called miners, connect machines to the network and continuously demonstrate possession of their silicon through a process called ZkGEMM Proof of Hardware. Each machine is bound to a persistent hardware fingerprint derived from the SHA-256 hash of its GPU device identifiers and system identifier, recorded on-chain at registration. If hardware is swapped, the fingerprint changes and the discrepancy is immediately detected.
Validators — which require only commodity CPU hardware, no GPUs — issue continuous timed challenges to every miner. Each challenge is a nonce-seeded GEMM workload sized to the claimed GPU class, with a response deadline that only the stated silicon can meet. Responses commit outputs under a Merkle root, and validators audit randomly selected leaves, making fabrication computationally equivalent to simply doing the work. Challenges are issued with fresh nonces on an ongoing basis, meaning proof of hardware is a live property the fleet maintains rather than a one-time ceremony. Proof history accumulates into a public reliability score visible to any renter.
Once a machine is verified, it enters the network’s resource routing model. Rental demand takes priority: when a renter requests capacity, verified free machines are instantly available for SSH-based rental. When no rental is pending, idle capacity is automatically routed into managed training workloads so GPU operators keep earning rather than waiting. This design eliminates the false idle-market problem common to other compute networks — rented and busy inventory stays visible as active network state, giving renters an accurate picture of true availability.
Subnet 106 is structured as a protocol with many access points. Because validator proof-checking is CPU-only work, any operator can run a validator and open their own rental storefront over the shared, network-owned fleet. Rather than routing all demand through a single platform frontend, Nodexo exposes every verified machine through any door — meaning competition among storefronts improves service without fragmenting supply. This architecture is deliberately designed for a world where renters include not just human developers but autonomous AI agents that can rent compute programmatically without accounts or stored credentials.
The Nodexo product is a live GPU rental marketplace accessible at nodexo.ai, backed by a continuously verified fleet of NVIDIA GPUs spanning consumer and professional tiers. The live inventory includes configurations from GeForce RTX 4090 (24 GB) and RTX 5090 (32 GB) up through RTX A6000, A100-SXM4-80GB (in 1×, 2×, and 4× multi-GPU configurations), H100 80GB HBM3, and RTX PRO 6000 Blackwell — all listed with real-time reliability scores and trust state. Every machine on the inventory has passed cryptographic hardware challenges; renters can select by GPU class, VRAM, and multi-GPU count and provision a root SSH session in minutes.
Renters have three distinct paths to capacity. The first is x402 accountless rental: using the open x402 internet-native payment standard, a wallet signature alone authorises payment and provisioning — no account creation, no stored API keys, no onboarding steps that a non-human cannot perform. This makes Nodexo fully agent-native by design, positioning it directly for the growing wave of autonomous AI workloads. The second path is a prepaid credit balance funded in TAO or USDC, against which metered rentals run with no fixed end time and can be terminated at any moment. The third is the conviction allowance system: participants who perpetually lock SN106 alpha stake to the subnet owner’s hotkey receive a daily compute credit proportional to their conviction, while the locked stake itself is never spent and continues to earn its native Bittensor yield — effectively a second yield denominated in compute.
For GPU providers (miners), the Nodexo operator dashboard handles fleet registration, hotkey linking, hardware attestation, and endpoint configuration through a guided web interface and a one-command installer. The setup script installs Docker, the NVIDIA container toolkit, Sysbox, PM2, and the prebuilt ZkGEMM proof extension — then verifies the GPU model against the approved hardware list before registering the machine on-chain. Providers earn through two channels simultaneously: subnet emissions for maintaining provable hardware, and direct rental revenue when their machines are occupied.
A second trust tier is available on hardware that supports it: confidential compute via Intel TDX, AMD SEV-SNP, or NVIDIA Confidential Computing extensions. Machines with supported hardware are flagged separately in the inventory, enabling encrypted execution boundaries for workloads that require stronger data isolation. The subnet’s contracts run on the Bittensor EVM and settle in TAO; deposits flow through a gateway contract that routes an owner share (capped at twenty percent on-chain) and stakes the remainder to the subnet’s validator, meaning every rental payment directly increases the stake securing the network that serves it.
The platform is built for horizontal distribution. Validators operate independent storefronts and compete on interface and reach rather than on captive GPU supply, preventing any single frontend from becoming a chokepoint. For developers and enterprises, a public REST API, a CLI, and an x402-compatible SDK surface the full rental, account, and operator workflow for integration into existing pipelines.
The Nodexo product is a live GPU rental marketplace accessible at nodexo.ai, backed by a continuously verified fleet of NVIDIA GPUs spanning consumer and professional tiers. The live inventory includes configurations from GeForce RTX 4090 (24 GB) and RTX 5090 (32 GB) up through RTX A6000, A100-SXM4-80GB (in 1×, 2×, and 4× multi-GPU configurations), H100 80GB HBM3, and RTX PRO 6000 Blackwell — all listed with real-time reliability scores and trust state. Every machine on the inventory has passed cryptographic hardware challenges; renters can select by GPU class, VRAM, and multi-GPU count and provision a root SSH session in minutes.
Renters have three distinct paths to capacity. The first is x402 accountless rental: using the open x402 internet-native payment standard, a wallet signature alone authorises payment and provisioning — no account creation, no stored API keys, no onboarding steps that a non-human cannot perform. This makes Nodexo fully agent-native by design, positioning it directly for the growing wave of autonomous AI workloads. The second path is a prepaid credit balance funded in TAO or USDC, against which metered rentals run with no fixed end time and can be terminated at any moment. The third is the conviction allowance system: participants who perpetually lock SN106 alpha stake to the subnet owner’s hotkey receive a daily compute credit proportional to their conviction, while the locked stake itself is never spent and continues to earn its native Bittensor yield — effectively a second yield denominated in compute.
For GPU providers (miners), the Nodexo operator dashboard handles fleet registration, hotkey linking, hardware attestation, and endpoint configuration through a guided web interface and a one-command installer. The setup script installs Docker, the NVIDIA container toolkit, Sysbox, PM2, and the prebuilt ZkGEMM proof extension — then verifies the GPU model against the approved hardware list before registering the machine on-chain. Providers earn through two channels simultaneously: subnet emissions for maintaining provable hardware, and direct rental revenue when their machines are occupied.
A second trust tier is available on hardware that supports it: confidential compute via Intel TDX, AMD SEV-SNP, or NVIDIA Confidential Computing extensions. Machines with supported hardware are flagged separately in the inventory, enabling encrypted execution boundaries for workloads that require stronger data isolation. The subnet’s contracts run on the Bittensor EVM and settle in TAO; deposits flow through a gateway contract that routes an owner share (capped at twenty percent on-chain) and stakes the remainder to the subnet’s validator, meaning every rental payment directly increases the stake securing the network that serves it.
The platform is built for horizontal distribution. Validators operate independent storefronts and compete on interface and reach rather than on captive GPU supply, preventing any single frontend from becoming a chokepoint. For developers and enterprises, a public REST API, a CLI, and an x402-compatible SDK surface the full rental, account, and operator workflow for integration into existing pipelines.
Nodexo is the product of the Neural Internet and VoidAI team, with Hansel Melo — an early Bittensor miner and entrepreneur known in the ecosystem as Zoro — serving as the driving force behind Subnet 106. Melo founded VoidAI as a Bittensor-native infrastructure company and registered SN106 in 2025 under the name Liquidity Provisioning, at that time focused on cross-chain TAO liquidity coordination via Solana DeFi protocols. The team built out bridging, routing, and concentrated liquidity infrastructure for that initial mission before pivoting the subnet’s focus entirely to compute.
The rebrand to Nodexo was announced in June 2026, with Phase 1 — covering Proof of Hardware, flexible payments, and the Conviction Allowance system — going live on June 23, 2026. The rationale for the pivot was a recognition that decentralised compute, with its clearer cryptographic proof model and direct product-market fit for AI workloads, offered a more durable and scalable incentive structure than liquidity provisioning alone. The shift transforms SN106’s emissions from rewarding capital provision to rewarding verifiable hardware execution, aligning the subnet’s economic model directly with real utility: compute delivered and verified.
The nodexo-ai GitHub organisation hosts the primary operator repository, which is written in Python and covers the full miner, validator, CLI, and setup surface. The team operates a public testnet (netuid 468) alongside the mainnet deployment (netuid 106) and maintains active documentation at nodexo.ai/docs. While the team does not publish a full public roster, it has an active presence on X at @nodexo and a community Discord, and has consistently communicated technical milestones and protocol updates directly through those channels.
Nodexo is the product of the Neural Internet and VoidAI team, with Hansel Melo — an early Bittensor miner and entrepreneur known in the ecosystem as Zoro — serving as the driving force behind Subnet 106. Melo founded VoidAI as a Bittensor-native infrastructure company and registered SN106 in 2025 under the name Liquidity Provisioning, at that time focused on cross-chain TAO liquidity coordination via Solana DeFi protocols. The team built out bridging, routing, and concentrated liquidity infrastructure for that initial mission before pivoting the subnet’s focus entirely to compute.
The rebrand to Nodexo was announced in June 2026, with Phase 1 — covering Proof of Hardware, flexible payments, and the Conviction Allowance system — going live on June 23, 2026. The rationale for the pivot was a recognition that decentralised compute, with its clearer cryptographic proof model and direct product-market fit for AI workloads, offered a more durable and scalable incentive structure than liquidity provisioning alone. The shift transforms SN106’s emissions from rewarding capital provision to rewarding verifiable hardware execution, aligning the subnet’s economic model directly with real utility: compute delivered and verified.
The nodexo-ai GitHub organisation hosts the primary operator repository, which is written in Python and covers the full miner, validator, CLI, and setup surface. The team operates a public testnet (netuid 468) alongside the mainnet deployment (netuid 106) and maintains active documentation at nodexo.ai/docs. While the team does not publish a full public roster, it has an active presence on X at @nodexo and a community Discord, and has consistently communicated technical milestones and protocol updates directly through those channels.
Phase 1 of Nodexo — the live compute marketplace with ZkGEMM Proof of Hardware, x402 accountless rentals, credit-based metered billing, and the Conviction Allowance system — launched on Bittensor mainnet on June 23, 2026. Phase 2 introduces tokenized sub-subnets: each specialised compute market within SN106 issues its own token on the Bittensor EVM, is capitalised through a constant-product liquidity pool denominated in SN106 alpha, and receives a share of miner emissions proportional to net capital flowing into its pool — precisely mirroring how dynamic TAO allocates emissions among Bittensor subnets, one level down. At launch, up to four tokenized sub-subnets are active, with the weakest displaced when a stronger new entrant graduates through the registration and auction process.
Each sub-subnet registration follows a dynamic cost schedule modelled on Bittensor’s own subnet registration — cost rises with demand and decays toward a floor — and the fee is burned, returning value to all SN106 alpha holders. New sub-subnets open a continuous clearing auction: bidders contribute alpha (held as liquidity, not spent), and at graduation the protocol mints the opening token supply, distributes it to bidders pro rata, and seeds the constant-product pool. The result is a self-organising layer of compute markets inside Subnet 106, each with its own community and token, each surviving only as long as it attracts capital and maintains verified hardware — an on-chain market of markets for specialised GPU workloads.
Beyond the sub-subnet layer, Nodexo’s roadmap includes broader hardware support as new GPU configurations are validated, expanded operator tooling, and deeper integration with AI agent ecosystems that use x402 for autonomous compute procurement. The confidential compute tier — already supported for TDX, SEV-SNP, and NVIDIA CC-capable hardware — is expected to expand as more operators bring attestable machines online, opening the network to workloads with stricter data isolation requirements.
Phase 1 of Nodexo — the live compute marketplace with ZkGEMM Proof of Hardware, x402 accountless rentals, credit-based metered billing, and the Conviction Allowance system — launched on Bittensor mainnet on June 23, 2026. Phase 2 introduces tokenized sub-subnets: each specialised compute market within SN106 issues its own token on the Bittensor EVM, is capitalised through a constant-product liquidity pool denominated in SN106 alpha, and receives a share of miner emissions proportional to net capital flowing into its pool — precisely mirroring how dynamic TAO allocates emissions among Bittensor subnets, one level down. At launch, up to four tokenized sub-subnets are active, with the weakest displaced when a stronger new entrant graduates through the registration and auction process.
Each sub-subnet registration follows a dynamic cost schedule modelled on Bittensor’s own subnet registration — cost rises with demand and decays toward a floor — and the fee is burned, returning value to all SN106 alpha holders. New sub-subnets open a continuous clearing auction: bidders contribute alpha (held as liquidity, not spent), and at graduation the protocol mints the opening token supply, distributes it to bidders pro rata, and seeds the constant-product pool. The result is a self-organising layer of compute markets inside Subnet 106, each with its own community and token, each surviving only as long as it attracts capital and maintains verified hardware — an on-chain market of markets for specialised GPU workloads.
Beyond the sub-subnet layer, Nodexo’s roadmap includes broader hardware support as new GPU configurations are validated, expanded operator tooling, and deeper integration with AI agent ecosystems that use x402 for autonomous compute procurement. The confidential compute tier — already supported for TDX, SEV-SNP, and NVIDIA CC-capable hardware — is expected to expand as more operators bring attestable machines online, opening the network to workloads with stricter data isolation requirements.
Big shoutout to Gordon Frayne for his incredible work on his TAO Pill podcast! His deep dives into Bittensor Subnets break down the decentralized AI landscape like few others can. Head over to his YouTube channel to catch the full series for top-tier analysis and insights.
Recorded in July 2026. Gordon Frayne speaks with Hansel from NodeXo, Bittensor Subnet 106, about the project’s return to its original focus on decentralized verified compute and the launch of NodeXo 3.0. Hansel explains how NodeXo is designed to aggregate heterogeneous compute from around the world, including Nvidia, AMD, Apple, consumer and enterprise-grade hardware, while using its verification system to prove that miners are genuinely supplying the computational resources they claim. The conversation explores how NodeXo differs from inference-focused subnets by providing access to raw compute that can be used for AI training, inference, mining other Bittensor subnets and other workloads, as well as its ambition to become a broader compute protocol rather than simply another GPU rental platform. A major focus is NodeXo’s tokenized compute model, where holders can lock Subnet 106 alpha through Bittensor’s conviction mechanism and receive daily compute credits proportional to the amount they have locked, creating direct utility for the token while incentivising greater demand for the network’s available hardware. They also discuss miner incentives, plans to reduce miner burn to attract more compute, automated revenue-driven alpha buybacks, and the longer-term vision of allowing external teams to build their own marketplaces and subnets on top of NodeXo’s verification infrastructure. Hansel also outlines an ambitious real-world expansion in Puerto Rico, where the team plans to combine solar panels, battery storage and physical GPU compute nodes in homes and businesses, allowing excess renewable energy to be monetised through Subnet 106 while adding permanent compute capacity to the network, ultimately supporting NodeXo’s goal of becoming a decentralized settlement and infrastructure layer for global AI compute.
@0xSero available on @nodexo for $12/h
it's genuinely insane.
i lock $7k into 106 once, and i keep it. it's staked, not spent.
that alone drips ~$200 of compute credits every 24 hours, forever. enough to run 4x h100s nonstop.
so a recoverable $7k buys me ~$73k a year of compute, and the principal never leaves my
Welcome back to our weekly Deep Dive!
14 H100s are sitting free right now. No waitlist, no contract.
We just added 22 H100s to the network, and 14 of them are available as you read
this.
The rest of the board is the same story: 34 A100s and 30 RTX PRO 6000s, so 86 of
our 101
Bittensor | $TAO
Like I mentioned before, locking @nodexo alpha gives users access to compute through daily GPU credits.
And this is not just any compute, it is verified GPU compute.
Right now, that means access to 100+ verified GPUs across 16 countries, with approximately
I’ve had many students over the years, both in my real estate business and in medicine, and one thing I’ve learned is that the most valuable systems are rarely one dimensional. There’s sometimes a deeper structure if you know where to look.
A building can house a business,
just got access to this
maybe I’ll leverage some compute on @nodexo to fine tune an open VLA/world model
$TAO stay tuned!
