RELATIVE DEVIATION
PAPER ONLY
Explore Architect ↗Uwe’s referral link · Architect is a Traders Guild client
Questions? DM @UweJensCerron on Telegram ↗PRIMARY COMPUTE FEED / BITGET
What does a GPU-hour trade for?
Public exchange quotes for GPU rental-price perpetuals. Select H100 or B200 to compare the market price with its index.
Loading Bitget…
Primary source: Bitget public API. Benchmark: Silicon Data GPU rental-price indices. Snapshots refresh with the market table every 60 seconds; this is not a streaming feed. The dated Architect future and synthetic energy-pair replay below use different contracts and are not repriced from these quotes.
Contract and index explanation ↗COMPUTE MARKET WATCH
Every venue. Actual market status.
trade[XYZ] and Paragon are separate market builders on Hyperliquid. Compare their contracts alongside Lighter, Helix, Bitget and Architect.
Checking public exchange APIs…
Refreshes every 60 seconds while this page is visible. Server snapshots may be cached for 60 seconds. This is a current-market monitor, not a stored historical series. Delisted and paused reference prices are suppressed. “—” means unavailable, never zero. Volumes are venue-reported quote notional, not independently verified organic activity. Different indices, collateral and expiry terms mean price gaps are not automatically arbitrage.
Inspect observations and source requests ↗H100 / CROSS-VENUE RESEARCH
Lighter vs Architect
Compare the instrument before comparing the price. A perpetual and a dated future can carry different benchmarks, funding and maturity exposure.
Loading public Lighter H100 snapshot…
| FEATURE | LIGHTER H100 | ARCHITECT H100 |
|---|---|---|
| Contract | H100 perpetual · discovered via API | NVDA-H100-2026-DEC · dated future |
| Size convention | Awaiting public metadata | 730 GPU-hours / contract · saved specification |
| Benchmark | API index price; provider methodology not yet verified | Compute Desk · saved specification |
| Carry | Awaiting funding observation | Expiry / futures basis; not a perpetual funding comparison |
| Bid / ask | Awaiting observation | Use authenticated “Check books” below |
| Data access | Public mainnet API · no key required | Authenticated books; environment labeled below |
Lighter: fetched snapshots, cached up to 30 seconds, not a live stream. Receipt time is not an exchange timestamp. Depth is visible notional within 1% of the mid from up to 100 orders per side, not guaranteed fill capacity. Funding is shown in raw API units until its period is verified. No annualized yield or cross-venue arbitrage signal is inferred. Synthetic replay below remains separate.
Lighter API documentation ↗ · Inspect snapshot JSON ↗ · Architect specification evidence ↗
A NEW MARKET. A FAMILIAR QUESTION.
What is compute
worth against energy?
Follow the gap between GPU rental futures and gas exposure.
Test the relationship before you trade it.
TRADERS GUILD / EXPERIMENT 001
H100 · DEC 2026
versus
UNG · PERPETUAL
Relative-price research.
Not a calibrated electricity hedge.
Not a calibrated electricity hedge.
Synthetic / convergenceConstructed quotes, not historical returns. No exchange orders.
MODEL, NOT A SIGNAL SERVICEPAPER P&L · AFTER MODELED COSTS
MAXIMUM DRAWDOWN
Peak-to-trough, including open P&L
ROUND-TRIP COST ESTIMATE
01 / RELATIVE PRICES
Two markets. One hypothesis.
H100 UNG
Both start at 100 · index of midquotes · UTC
THE RESEARCH TICKET
Watch the relationship
Compute contracts
Energy contracts
Proposed gross exposure
Move to cover costs*
*Approximation versus compute notional. β is an assumption. Contract rounding leaves unmatched exposure. No measured fill capacity.
⊘ Execution disabled
02 / SIGNAL
How unusual is the gap?
Enter at ±2σ, close near zero. No entry beyond ±4σ. The current observation is excluded from its baseline.
03 / ACCOUNTING
Does the idea pay?
Includes modeled spreads, fees and slippage. Excludes funding, financing, market impact and margin liquidation. Dashed line: no trade ($0).
FROM REPLAY TO OBSERVATION
Connect the idea to the order book.
The demo uses the same native symbols as the Hummingbot monitor. Fetch an authenticated snapshot to inspect the two books. Observations never replace the synthetic chart silently.
Not connected. No live liquidity or executable prices have been verified.
Hummingbot → Architect → Lab
Uses architect_perpetual authentication and throttling. The monitor reads the dated GPU contract by native symbol; it does not pretend it is a perpetual.
Import the paired quotes saved by the Hummingbot or Node collector. Files stay in this browser. Imported data remains unverified.
THE THESIS / AND HOW TO BREAK IT
The useful question isn’t
“will AI get bigger?”
It’s whether the price of its output is moving differently from the cost of its inputs, and whether that difference can be traded.
Read the research note ↗The hypothesis
If compute outpaces energy without a lasting change in demand or capacity, the gap may narrow. The research trade is short compute and long energy. Reverse the legs for the opposite divergence.
The competing explanation
GPU scarcity can lift rental prices while gas stays flat. Power contracts and regional grids can disconnect electricity costs from UNG. A wide gap can be rational and persistent.
What would earn conviction
A relationship that survives held-out data, actual funding and costs, executable two-leg depth, and different regimes. A low z-score alone proves none of those things.
AUDIT THE EXPERIMENT
Every paper decision, visible.
| TIME / UTC | DECISION | DEVIATION | H100 CONTRACTS | UNG CONTRACTS | REALIZED P&L |
|---|
24h maximum hold$500 cumulative loss threshold5m cooldownStale quotes rejected
Stops are decision rules, not guaranteed loss caps. A data gap while holding a position halts the simulation without inventing an exit fill.
WHAT IS VERIFIED
Contract definitions, not returns.
Architect instrument snapshot · September 26, 2026
H100 DEC: 730 GPU-hours per contract, Compute Desk benchmark. UNG: one ETF share per contract. ORNN is not this future’s settlement index.
ABOUT THIS RESEARCH
Compute and energy pair trading, explained.
Created by Uwe Cerron at Traders Guild, this lab studies relative prices between H100 GPU rental futures and UNG natural-gas ETF perpetuals listed in Architect’s instrument catalogue.
What does the model measure?
The model compares log compute prices minus an assumed energy exposure ratio times log energy prices with the previous 60 observations. A deviation is a research input, not proof that the relationship will reverse.
Are these live returns?
No. The default experiments use constructed quotes. Paper results include modeled spreads, fees and slippage, but exclude funding, financing, market impact and liquidation. Imported CSV observations are labeled separately. The dashboard places no orders.
Does natural gas hedge a datacenter’s power bill?
Not directly. UNG represents a natural-gas futures ETF, while electricity costs depend on location and power contracts. The energy leg is a proxy whose usefulness must be tested.
Can I reproduce the experiment?
Read the trade thesis and limitations, inspect the instrument snapshot, or review the Python and Node.js source. The Hummingbot monitor is read-only; authenticated runtime validation remains outstanding.
Reference: Architect’s compute-spread framework and Hummingbot’s Architect connector documentation. Machine-readable research guide.