Where the number comes from, how it becomes a market, what happens at expiry — and what is deliberately not yet true. This page is the whole system in reading order: everything in §1 and §2 runs today, on testnet; everything that does not yet exist is in §3, as a plan behind named gates.
Every market below settles against a published index of what GPU-hours actually rent for. This is how that number is made, stage by stage. Values set in mono are governed parameters of the published spec: the §2 timelock is the only way they move, and the per-tier status at /methodology — source counts and settlement grades, tier by tier — is the authority if this page ever lags it.
Independent rental venues with public, machine-readable pricing are polled continuously. Every print carries its own source count. An executed lease price carries more weight than a listed ask: a filled trade is better evidence than an offer.
Every observation is reduced to a versioned hardware specification: for the flagship, H100 SXM 80 GB, on-demand, cluster-capable. Other form factors are scaled to that specification, and quotes that are implausible or stale are discarded.
Observations far from the cross-source consensus are dropped. Wash prints, fat fingers and decorative asks die here.
What survives is combined into a single weighted print, and no venue may carry more than a capped share of the total weight.
A signer quorum posts each value on-chain, bounded by a per-print movement cap. Below a minimum count of independent sources a tier freezes — visibly STALE on-chain, never estimated, never interpolated. A missing print means we did not have an honest number, and the record keeps the hole.
Settlement demands more still: a higher source count for settlement grade. Below that a tier publishes for display only, and nothing settles against it.
The signers are house-held today. That is a limitation, not a footnote: moving them to independent parties is a named gate in §3.
What trades against the number, and what keeps the venue solvent while it trades. Everything in this section is live on testnet and readable without an account.
Three instrument families, all cash-settled against the index. Forwards: fix a GPU-hour rate for a future month. European options: calls and puts across a strike ladder, priced Black-76. Basis swaps: one tier against another, today the H200−H100 spread. Contract sizes and listed expiries are listing data, read live at /v1/products rather than typed here.
Execution is request-for-quote: ask, and makers return signed, firm quotes your wallet accepts or ignores. Every quoting maker today is house-operated and disclosed at /v1/mm-stats; volume here is operational evidence, not external demand.
Custody never moves: margin is posted from your own wallet, the venue holds no keys and can never sign for you. Margin requirements are parameters the chain owns; the terminal reads them live rather than this page restating them. A position that falls below maintenance is liquidated RFQ-first; losses beyond a defaulter's margin fall on the insurance fund, and only past that are socialized, on-chain and on the record.
The accounting identity — clearing pool = Σ collateral − Σ unrealised basis − socialized losses — is recomputed from public state at /v1/solvency and re-derived in your browser on the landing page. Its residuals must be zero to the wei, and the operators' own monitor alarms on anything else.
At expiry a contract settles in cash against a 72-hour TWAP of the published series, subject to minimum coverage. Never a single print, so a last-second spike settles nothing. There is no delivery.
Every governed parameter sits behind a timelock (1 h on this testnet; 48 h is the guarded-launch target), so a change is visible before it is live. The core contracts are immutable, and withdrawals, settlement and liquidation can never be paused. By anyone. The operators included.
Two attack surfaces are architectural and remain open. The index reads public rental markets, which a determined actor could try to move at the source: the index's aggregation bounds the damage per print; it does not erase the surface. And the oracle signers are house-held until §3's second gate. The contracts have passed two internal audit cycles; no external audit has been completed. This clause is why the plan below is ordered by gates, not dates.
Everything above runs today, on testnet, with play money. Everything below is a plan, not a fact: in order, each stage behind a gate that is passed or not passed, never skipped. No dates are promised, and the dashed line means what it means: not built yet.
Invite-only testnet on Base Sepolia: MockUSDC with no monetary value, real market structure, the index printing around the clock. House market makers and synthetic trading bots exercise the system and are disclosed as such, trade by trade.
Structured sessions with the people the venue exists for: neoclouds hedging fleet revenue, AI teams hedging an inference bill. Then wider invite waves. The gate: repeated, sized hedging intent from real operators. Sign-ups do not count.
The signer set moves from house-held to independent signers, every tier that settles holds settlement grade, and the source panel keeps widening. The gate: no single party — the operators included — can move a published mark.
An external audit of the contracts, and counsel sign-off on entity, jurisdictions and terms. The two internal audit cycles already done are not the gate. Nothing holding real value deploys before both.
Real USDC behind the same invite gate, with low position caps and the 48 h timelock. Whitelisted external market makers quote alongside the house, and caps rise only as the record accrues: the same solvency identity as today, now with something to lose.
More settlement-grade tiers, structured hedging products for compute sellers, and the agent layer scaled up: programmatic hedging through a metered API already runs on this testnet. And one commitment that holds at every stage: no token exists and none is planned. The venue earns fees, or it earns nothing.
TESSELEXGPU derivatives venue