Targon
SN4Run AI models in a secure cloud where nobody, not even the host, can see your data
Targon runs an open auction for confidential GPU compute. Miners bid down the price of hardware-secured H200s; validators verify the encryption is real before routing the job. The team is Manifold Labs, the same crew behind a confidential compute whitepaper co-authored with Intel.
// Confidential AI, priced by auction.
Targon is a decentralized marketplace for renting GPU compute, with a twist: every job runs inside a hardware-protected enclave so the operator of the machine cannot see your data or model. Miners compete on price in a bidding round, validators check that the GPU genuinely supports Confidential Compute, and the cheapest qualified bidder wins the job.
The simple version: Imagine renting a soundproof studio where the landlord physically cannot listen in, only here the studio is a GPU and the lock is enforced by the silicon itself.
Centralized equivalent: Think AWS Nitro Enclaves, Azure Confidential Computing, or Google Confidential VMs, but priced through an open auction rather than a fixed rate card.
How it works:
- Miners provide GPU hardware with NVIDIA Confidential Compute or Protected PCIe (PPCIE) support, then bid in price auctions to win compute jobs.
- Validators orchestrate the bidding rounds, verify hardware attestation, and award jobs to the most cost-effective qualified provider.
- The problem it solves: Regulated industries (healthcare, finance, defense) want to run AI on proprietary data without exposing it to the cloud operator. Confidential computing closes that gap, but the centralized versions are expensive and require trusting one vendor end to end.
- The opportunity: Confidential AI workloads sit at the intersection of two large trends, enterprise AI adoption and zero-trust infrastructure. Pricing them through competition rather than a fixed cloud SKU is a real lever.
- The Bittensor advantage: A permissionless network of verified secure providers, settling prices in an auction, can produce both lower prices and a wider supplier pool than any single hyperscaler offers.
- Traction signals: Manifold Labs published a confidential compute research piece co-authored with Intel, and the team has been talking publicly about the architecture on the Bittensor Guru podcast and via LinkedIn. The targon.com stats page exposes a live miner order book.
Category: Inference and Compute | Centralized Competitor: AWS Nitro Enclaves, Azure Confidential Computing, Google Confidential VMs
Targon sits at the intersection of two trends, the explosion in AI compute demand and the growing need for verifiable security in cloud infrastructure. Most decentralized compute subnets compete on raw price alone. Targon adds a trust layer, using hardware-level encryption to guarantee that data and weights stay private during computation, then drives price discovery through an auction.
Mechanism:
When a job comes in, miners submit price bids. Validators orchestrate the bidding round, verify that the proposed GPU genuinely supports Confidential Compute via attestation, and award the work to the lowest qualified bidder. Per the team's about-page write-up, this competitive design has produced access to H200 instances at as low as $1.07 per hour, though that price is the team's own claim and not independently audited.
The codebase is written in Go and lives at manifold-inc/targon. A direct GitHub API check shows 1,983 commits across 18 contributors, with the most recent commit dated 2026-05-21, a day before this writeup. The repository description is "A library for building subnets with the manifold reward stack," and recent commits show ongoing maintenance: version bumps, Discord alert tweaks, and minor updates by the same handful of core contributors.
Financially, the picture has shifted since the last snapshot. Price sits at 0.05427 TAO and market cap at 285,615 TAO, down from 302,415 TAO in March. The root proportion is 0.154, meaning the vast majority of pool depth comes from organic alpha demand rather than protocol subsidy. Net 7-day TAO inflow is positive at 627 TAO, but small relative to the 129,849 TAO of root depth in the pool.
The notable change is emissions. Under Taoflow, a subnet's emission share is driven by its smoothed net TAO flows. Targon currently shows an emission share of 0%, meaning recent net flows have not been strong enough to capture a slice of the 3,600 daily TAO. This is a market signal, not a development signal, the codebase is shipping commits, the team is active publicly, but alpha buyers have not been pushing in hard enough to register on the Taoflow curve right now.
Active miner count is 3. That is unusually concentrated for a compute subnet, and almost certainly a function of the hardware barrier: NVIDIA Confidential Compute or PPCIE support is not commodity equipment. It is also worth noting that fewer specialized miners can still serve real demand if those miners are well provisioned, but it does mean the supplier side is thin.
The team's pedigree is unusual for Bittensor. Per the TAO.app about page, Rob Myers (CEO) is described as a former Opentensor Foundation core contributor, James Woodman (COO) as ex-COO of OTF, with Joshua Brown as CTO and Ahmed Darwich as lead engineer. The Intel collaboration and the willingness to publish hardware-attestation research align with that background.
- Emission share at zero: Targon is not currently capturing protocol emissions under Taoflow. Net flows would need to strengthen for that to change. Existing alpha supply is not being diluted by protocol issuance, but the subnet is operating without an emission tailwind.
- Thin miner side: With 3 active miners gated by specialized confidential compute hardware, supplier diversity is limited. That is a feature of the security model and a constraint on capacity at the same time.
- Concentrated stake: HHI of 0.097 and a Gini coefficient of 0.716 on the top 100 stakers indicate a meaningfully concentrated stake distribution. Large unstakes could move price.
- Competition from hyperscalers: AWS, Azure, and Google all sell confidential computing today with established enterprise relationships. Targon's pricing edge has to be real and sustained to win serious workloads off them.
- Unverified revenue claim: A recent podcast described Targon as a top-two claimed-revenue subnet on Bittensor at roughly $10.4M ARR. This is the team's number, not an audited figure, and should be treated as such.
