Chutes — subnet 64

Chutes leads this batch's whole tracked field on one measure and sits mid-pack on another, and those two facts do not point in quite the same direction — which is exactly the kind of gap this dossier exists to pull apart.

What it does

Chutes is a subnet that runs a marketplace for AI inference on rented GPU hardware. A developer packages an application as a "chute" — the project's own comparison is to a single FastAPI web application — built on top of a container image, and defines individual functions inside it called "cords," which work like that application's individual routes. [source] [source] [source]

Anyone can become a user of the platform by registering a Bittensor hotkey and creating API keys, then calling chutes over a simple bearer-token header. [source] A developer controls which hardware a chute is allowed to land on — how many GPUs, how much VRAM, an optional price ceiling per GPU-hour — and is billed for real time spent on that hardware: a one-time deployment fee, then the actual hourly rate while an instance is warm, but nothing for the time it spends downloading a model before it is ready to serve requests. [source] [source] [source] [source] [source]

At this snapshot Chutes sat in this batch's highest earnings band, taking 12.2% @ 8871732 of the day's emission — third highest of the whole batch, 3 @ 8871732 — while also carrying 13589 @ 8871732 holders of its own token, the most of any contest this course tracks, 1 @ 8871732.

How the incentive mechanism works

The repository pinned for this dossier, chutesai/chutes, is the command-line tool and developer kit for the chutes.ai platform — what a developer installs to register, build images, and deploy chutes. [source] By that same repository's own README, the code that runs as a miner and the code that runs as a validator — the two pieces that between them would define what gets scored and how a chute's operator gets paid — are published in two other repositories entirely, and neither is pinned in this course's corpus, so this dossier cannot describe how that scoring works or how it turns into an on-chain result. Stated plainly rather than guessed at: the mechanism could not be read. [source]

What the pinned client does show is one piece of the trust model around it: GraVal, a graphics-card validation library that checks the GPU a chute claims to be running on is the real thing, through VRAM-capacity checks and matrix multiplications seeded by that device's own information. [source] It runs as middleware inside the chute itself, decrypting traffic sent from outside and performing further checks — filesystem checks, device-info challenges, pings — before that traffic is served. [source]

What miners are actually optimising

Because the code that would answer this sits outside what this dossier can read, the exact formula that turns a chute's performance into pay is not something this analysis can state as fact. What the pinned client does establish is the demand side of the market: a developer's node-selector settings — GPU count, minimum VRAM, an optional price ceiling — decide which hardware a chute is even allowed to run on, and a developer only pays for GPU-seconds actually spent warm and serving requests. [source] [source] [source] [source]

Unverified On the visible evidence alone, the more plausible reading is that a chute operator's earnings track being cheap, available, and verifiably running the claimed hardware — the three things GraVal and the billing model actually check — rather than any graded measure of output quality, the way a contest that scores individual answers against a benchmark would. This course's pinned source does not confirm that reading; it is this author's inference from what is visible, not a documented claim, and the two repositories that would settle it sit outside this corpus.

Why it earns

Under this network's own design, a contest's share of the day's pay against every other contest is set by one formula and nothing else: share_i = p_i × (1 − b_i) / Σ_j p_j × (1 − b_j), where p is that contest's own moving-average alpha price and b is a penalty for pay an owner withheld to its own keys. [source] [source] Mechanism quality does not appear anywhere in that formula directly — a better-designed contest can only move this number by first moving price, by convincing enough people to actually buy and hold the token — so mechanism and price are not two things this snapshot, or any snapshot, could ever fully pull apart on their own.

That formula also settles Chutes' own headline gap rather than deepening it. Its two most-watched numbers sit far apart — 3 @ 8871732 by earnings, but 1 @ 8871732 by number of holders, the only member of this batch to lead this course's whole tracked field on that second measure — and how many separate wallets hold a token plays no part in the formula above at all. A contest can reach a given price with many small holders or a few large ones, and either produces exactly the same earnings rank; the two rankings measure distribution and price, not two views of one underlying quality, so they were never going to move together. false @ 8871732 is what this batch's own automated flag reports for that gap, and it does not consider a two-place difference large enough to raise — consistent with a difference the formula itself says is not comparable in the first place.

Mechanism evidence

Because mechanism only ever reaches pay by first moving price, Chutes' earnings can never confirm, on their own, that whatever scores a chute's work is any good. This dossier also cannot read that code for Chutes specifically: the pieces that would define what gets scored and how a chute's operator gets paid are published outside what was pinned here. [source] What can be said is only the visible participant count, which is small for the size of the reward: 14 @ 8871732 active producers are dividing a 12.2% @ 8871732 slice of the network's daily payout, against a maximum of 256 @ 8871732 registration slots on the contest. None of that confirms or denies whether the underlying work is good; it only describes how few people are doing it — the exact judge count sits in the technical note at the end of this dossier.

Concentration evidence

13589 @ 8871732 holders is a count of wallets holding Chutes' own token on the open market — it is not a count of anyone with a say in setting a weight on this contest. That power sits with a much smaller permitted set of judges, and one judge's influence over that vote scales with money committed behind it, not with how many separate wallets hold the token. [source] Root proportion — how much of that judging power traces back to money staked into the root network generally, rather than into Chutes' own token specifically — sits at 14.0% @ 8871732 here. [source] A double-digit reading on top of a permitted set this small means a handful of generalist backers, not the broad holder base above, decides how 12.2% @ 8871732 gets split among Chutes' 14 @ 8871732 active producers.

Price evidence

Chutes carries real, sustained market activity: a price of 0.07907444 @ 8871732 TAO per token, 5284835776985 @ 8871732 rao traded over the preceding day, and a market capitalisation of 451469883728254 @ 8871732 rao — down -8.54136796619152 @ 8871732 percent over the trailing week even as it rose 0.3650992761603817 @ 8871732 percent over the trailing month. [source] Because the cross-contest formula above runs on price alone, this is the one leg of the three that mechanically explains Chutes' earnings rank: whatever mix of real demand for its product, its judges' own confidence, or plain speculation is moving that price is, by construction, the actual cause of the rank, and nothing pinned here lets this dossier separate those causes from one another.

What the holder count separately confirms is only that Chutes' own token is widely distributed — the most broadly held of any contest this course tracks — which is evidence about how many people are interested enough to hold some, not evidence about the price they are holding it at or the earnings that price produces. 13589 @ 8871732

Failure modes and attack surface

The clearest structural risk here is arithmetic: a small handful of judges and 14 @ 8871732 active producers are, between them, entirely responsible for a contest earning 12.2% @ 8871732 of the network's daily payout — a small enough group that losing or coercing a handful of keys changes outcomes materially. That same concentration is what makes coordinated agreement among judges hard for the chain to ever catch on its own, since the ranking it produces only ever sees a pile of money-weighted numbers, never who is actually behind them or whether they are acting together. [source]

GraVal's hardware check is the one anti-cheating measure this dossier can point to, and it only covers one kind of dishonesty — claiming GPU hardware that isn't actually there. [source] Whether the code receiving that traffic actually enforces the check correctly, and what happens to a chute that fails it, is governed by code in the separate, unpinned repository this dossier cannot read — so this course can confirm the check exists and describe what it tests, but not confirm it is applied as designed in production. [source]

What to copy, what to avoid

Worth copying: verifying a claimed physical resource — here, a specific GPU — with a challenge the claimant cannot fake without actually having the hardware, run as middleware close to where the work happens rather than trusted on the honour system. [source] [source]

Worth avoiding: publishing only the developer-facing client under the licensed, publication-eligible repository while keeping the code that actually decides who gets paid, and how much, in separate repositories outside it. Whatever the reason, the practical effect is that nobody reading only the pinned, licensed source — this dossier included — can check what the contest's own incentive design actually rewards. [source]