KubeTEE

SN90 KubeTEE rents out GPU factory floors whose workrooms are sealed so that even the landlord cannot look inside, paying the owners of those rooms a dollar-denominated hourly rate in the subnet's token.

The commodity, explained from zero

The commodity is confidential compute: hours on high-end GPUs, the processors that run AI, inside a trusted execution environment, a sealed compartment in the chip that proves what code ran and hides the memory even from the machine's owner. The front page calls it "An AI factory, run where no one can read your data."

It has a price because some buyers, the site says regulated and enterprise workloads, cannot put their data on an ordinary rented server where the operator could read it. The validator dashboard carries a price card in dollars per GPU hour, from $2.00 for an RTX 6000 to $9.75 for a B300.

Who buys today: not an enterprise, yet. The revenue dashboard is explicit: "KubeTEE serves inference on sayGM (Bittensor SN28). The alpha earned is swapped to SN90 and recycled." The factory's spare hours are sold through another subnet's reseller market, and the README says that channel gets only idle capacity. Here the comparison breaks: a landlord collects rent from tenants, but this landlord's only recorded income is from subletting empty rooms to a neighbour.

Why it is on Bittensor at all

The stated reason is paying for the factory by consumption: buyers spend the subnet's token to use it, spent tokens return to the unissued supply and are re-emitted, and emission, the subnet's share of newly minted TAO, is split 41 percent to miners, 41 to validators, 18 to the owner. The tokenomics document names one measure of whether it is working: the subsidy ratio, emission value over miner pay, "must be monotonically down." Cost against a conventional cloud is not compared.

How the work gets done

Miners provide a Kubernetes cluster, software that runs many containers across machines, of eight nodes with data-center GPUs, post a 100 TAO deposit that is held but not slashed, and, today, are onboarded by hand; the README says "KubeTEE onboards each miner today" and that self-service comes in Phase 1. Validators check a readiness gate and convert a dollar target into weight: usd_target_per_hour times tenure times window hours, divided by dollars per alpha, with outside price feeds as a hard dependency; if a feed fails, the cycle skips. The README's own table says TEE attestation, job metrics, and serving probes are "designed / to be built" and "not yet in the weight path."

How you would know it works

Two dated dashboards: revenue at https://kubetee.ai/dashboard ("Total recycled $10k since Aug 23," refreshed September 9, with swap extrinsics) and validator payouts at https://s3.hippius.com/kubetee-validator/index.html, timestamped 2026-09-10, showing a $2,338 epoch pool. The gateway at https://llm.kubetee.ai is a live Swagger page.

What is missing

There is no pricing page; the only prices are the validator's GPU-hour card and sayGM token rates, and USDC and TAO billing are listed as coming. The on-chain contact pierre@kubetee.ai appears nowhere on the site; the README gives a first name and an X handle. Miners cannot join without the operator, and there is no runnable miner guide. Support is Discord, with no status page or SLA.

Go deeper

• Front door with What is KubeTEE: https://kubetee.ai
• README with What Ships Today: https://raw.githubusercontent.com/KubeTEE-AI/kubetee-subnet/main/README.md
• Tokenomics: recycling and the split: https://raw.githubusercontent.com/KubeTEE-AI/kubetee-subnet/main/docs/TOKENOMICS.md

Metaphor: a factory whose workrooms are sealed so even the landlord cannot look in. Every claim is drawn from the evidence set or the subnet's own materials; "(inferred)" marks a conclusion rather than a quote. Corrections.