KubeTEE is the AI Factory of the Bittensor network. It turns decentralized GPU clusters into a confidential factory for state-of-the-art AI services.
SN 90. Rank 29 of 32 by emission
KubeTEE
Emission1.1%of the block reward · rank 29 of 32
Participants2 miners · 6 validators
Registration0.0005 TAO
Verified2026-09-09links checked and scores merged
The programcompute · github.com@030c992 · read 2026-09-19
The work
Confidential GPU capacity. A miner runs a Kubernetes cluster of 8-GPU worker nodes with Intel TDX and NVIDIA confidential computing, posts a 100 TAO deposit on its own hotkey, and is onboarded by hand. What is paid for today is capacity that passes a readiness check, priced per GPU-hour; job execution and attestation are not yet scored.
own docs · raw.githubusercontent.com
miners earn Bittensor emissions for providing confidential compute capacity and reliably executing Armada-scheduled jobs
How it is scored
A dollar target, not a benchmark. Each cycle the validator reads every miner cluster through Rancher, passes or fails a binary readiness verdict (hotkey binding, high-availability topology, 8-GPU workers, GPU passthrough, runtime handler), and for each ready miner computes GPU count times the card price per GPU-hour times tenure times window hours. Tenure is fixed at 1.0 while the reliability gate is unwired.
code · raw.githubusercontent.com · validator/miner_scoring.py
usd_target_per_hour x tenure x window_hours / usd_per_alpha; tenure: float = 1.0 # 0..1: 1.0 while the reliability gate is not wired
Sources disagree The README's economics section says emissions are weighted by attested TEE health, job-execution quality and uptime and that no attestation means no emissions; its What Ships Today table and the code at the pinned commit put only Rancher readiness in the weight path, with attestation, Armada metrics and health designed and not yet scored.
How the pay splits
Proportional to dollar target, capped by the pool. Each ready miner's earned dollars over the epoch's emission pool in dollars is its weight; the remainder is set on the owner UID, which under recycle returns the alpha to unissued supply. On 2026-09-20 the dashboard showed two clusters: the owner's staging cluster UID 56 at 64 percent of the pool and BTLABS UID 97 at 32 percent, $31.48 recycled.
code · raw.githubusercontent.com · validator/miner_scoring.py
Whatever miners do not earn is recycled to the owner UID (set as weight on the owner's hotkey). Weights are normalized to sum 1.0.
Who judges
Six validators were active on the 2026-09-09 chain snapshot. Commit-reveal is on; weights may be set once per 100 blocks. The validator reads cluster inventory through a GET-only Rancher client, TAO/USD from Taostats, and Targon's payout feed to clamp the price card downward; the subnet-owner validator publishes the price snapshot and dashboard to Hippius. A cycle without a trustworthy price, inventory or metagraph read is skipped.
chain · hyperparams-2026-09-19.json
commit_reveal_weights_enabled True, weights_rate_limit 100, yuma_version 3; chain-2026-09-09.json active_validators 6
How often it turns
Emission settles every 360 blocks, about 72 minutes. The README says the validator sets weights once per epoch, boundary-aligned; the config's window is one hour by default, so a miner earns its hourly target per cycle. The Hippius dashboard on 2026-09-20 read epoch 25,293. A new registration is immune from deregistration for 5,000 blocks, about 16 hours.
chain · hyperparams-2026-09-19.json
tempo 360, immunity_period 5000; config.py window_hours=_get_float("KUBETEE_WINDOW_HOURS", 1.0)
Who buys
Not an enterprise yet. The only recorded income is idle GPU capacity sold as confidential inference of GLM-5.2, GLM-5.3, GLM-5.3-Flash and Ornith-1.5-397B through sayGM, subnet 28, with the alpha earned swapped to SN90 and recycled daily. The validator's card prices GPU-hours from $2.00 for an RTX 6000 to $9.75 for a B300; paid jobs in TAO or alpha are a Phase 2 item.
own docs · kubetee.ai
KubeTEE serves inference on sayGM (Bittensor SN28). The alpha earned is swapped to SN90 and recycled
What has been gamed
No incident is recorded. With one production miner and the owner's staging cluster, the guardrails are structural: the GPU card is clamped downward to what Targon pays per card, floored at 75 percent of the card; a miner absent for sustained cycles is deregistered under a guard; the 100 TAO deposit is measured, not enforced; a cycle without a trustworthy price or inventory read keeps the previous weights.
own docs · raw.githubusercontent.com
A cycle that cannot get a trustworthy price, Rancher inventory, or metagraph read is skipped rather than guessed
4 of 7 fields rest on the chain or the code. A manifest is the site's reading of the subnet's program, not a review; each field links its source. All programs · What a field means · Correct a field
What it says
The subnet in its own words: the line the index took from its front door, and what the team has said on the record.
What it shows
Explained from zero by the index, from the subnet's own materials, for someone who has never heard of Bittensor.
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
Sources for this explainer
- https://kubetee.ai
- https://kubetee.ai/dashboard
- https://llm.kubetee.ai
- https://s3.hippius.com/kubetee-validator/index.html
- https://raw.githubusercontent.com/KubeTEE-AI/kubetee-subnet/main/README.md
- https://raw.githubusercontent.com/KubeTEE-AI/kubetee-subnet/main/docs/TOKENOMICS.md
- https://raw.githubusercontent.com/KubeTEE-AI/kubetee-subnet/main/docs/GPU-NODE-REQUIREMENT
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.
How legible it is
Four audiences, four questions each, rated on what a first-time reader can find in five minutes. The scale measures findability, not quality.
On-chain identity 7/7- subnet_name
- KubeTEE
- github_repo
- github.com
- subnet_contact
- pierre@kubetee.ai
- subnet_url
- kubetee.ai
- discord
- discord.gg
- description
- KubeTEE AI Factory: Confidential Computing TEE Multi-Cluster K8s
- additional
- x.com
Stakers and validators●●●●● 4.3
Should I allocate here? · rank 1 of 32 for this audience
Q1What it is●●●●● 4
KubeTEE is the AI Factory of the Bittensor network. It turns decentralized GPU clusters into a confidential factory for state-of-the-art AI services
GPU hour prices sit one click away
Q2Who it is for●●●●● 5
Total recycled $10k since Aug 23
with dated swaps and extrinsic ids, data refreshed September 9, revenue comes from the owner miner serving inference on sayGM
Q3How it resists gaming or fails●●●●● 4
Validator Scoring section with a What Ships Today table separating live from designed, feed failure contracts, attestation not yet in the weight path, an 86 KB README that needs expertise
Q4Identity and documentation●●●●● 4
identity 7 of 7, github, url and discord resolve, Tokenomics document covers recycling, owner conviction and the 41 41 18 split, contact pierre@kubetee.ai not on the site
Miners●●●●● 3.3
Can I compete, and what wins? · rank 18 of 32 for this audience
Q1What it is●●●●● 3
KubeTEE onboards each miner today
miners provide an attested TEE cluster for inference and Armada jobs, onboarding procedures intentionally unpublished, permissionless onboarding is a Phase 1 item
Q2Who it is for●●●●● 4
NVIDIA H100
"H200" "B200" "B300" and RTX PRO 6000 Server on Intel TDX Xeon, 8 nodes per cluster, a 100 TAO deposit, USD price card sets competitiveness, registration cost not stated
Q3How it resists gaming or fails●●●●● 4
readiness gate then usd_target_per_hour times tenure times window hours over usd per alpha, dashboard shows one miner at $454 per hour and 25.4 percent of the pool, guarded deregistration on absence
Q4Identity and documentation●●●●● 2
no runnable miner guide, the README says onboarding procedures are not published, semantic versioning defined for the validator image, repo pushed 2026-09-04, no GitHub releases
Buyers and enterprises●●●●● 2.8
Can I use this today? · rank 16 of 32 for this audience
Q1What it is●●●●● 4
LiteLLM API - Swagger UI
live OpenAI compatible gateway, front page says confidential inference for GLM-5.2, GLM-5.3 and Ornith-1.5-397B is available through sayGM, no sign up path or example
Q2Who it is for●●●●● 3
Enterprise-Grade
and regulated workloads named, prices exist only as a per GPU hour card on the validator dashboard and as sayGM token prices, USDC and TAO billing listed as coming, no pricing page
Q3How it resists gaming or fails●●●●● 2
support is a Discord server with a warning that the team never DMs first, no status page, no SLA, the README calls a consumer aligned validator the protocol native SLA
Q4Identity and documentation●●●●● 2
Swagger UI on the gateway is the only API reference, the chain contact pierre@kubetee.ai appears nowhere on the site, the README gives a first name and an X handle
Newcomers●●●●● 3.5
What is this and why does it matter? · rank 15 of 32 for this audience
Q1What it is●●●●● 4
An AI factory, run where no one can read your data
with a What is KubeTEE section that explains the host cannot read workload memory, consistent with the chain description
Q2Who it is for●●●●● 3
Every service runs inside a hardware-secured Trusted Execution Environment: the host cannot read workload memory
privacy stated as the difference, no comparison with a centralized cloud
Q3How it resists gaming or fails●●●●● 4
validator dashboard timestamped 2026-09-10 with epoch payouts, a revenue dashboard with dated swaps, a live Swagger endpoint, no benchmark of the AI services themselves
Q4Identity and documentation●●●●● 3
subnet_name matches the site, url resolves, the What is KubeTEE section serves as an about page, the description is a tag string of product terms, not a sentence
Provenance
How this rating came to be. Verified means the scorer fetched the page and the words are on it; secondhand means concluded from code, absence, or a third party. The link badge is a separate automated check made before publication.
scored by agent a 2026-09-09, rubric v1.0merged 2026-09-09links verified 2026-09-09: 33 live, 0 unreachable, 2 manual
Something wrong? Corrections of fact are applied as they arrive during the window; rating disputes are batched at its close. How to file one · GitHub issue · email.