The Defendable Vast.ai Utilization Signal Rail

Seven signal classes · the workhorse thesis · the yield-analysis standard.

Vast.ai is a core compute utilization intelligence rail. DefendableOS treats it carefully · honestly · with strict claim discipline. Listing rates are NOT resale value. Availability is NOT occupancy. Public visibility is NOT booked revenue. First-party rental receipts ARE operating evidence the platform can issue claims from.

The seven signal classes

• VAST_PUBLIC_LISTING_RATE · observed market hourly asking rate · proves current public observation · does NOT prove actual paid rental.
• VAST_PUBLIC_SUPPLY_OBSERVATION · number/category of visible offerings · proves apparent supply · does NOT prove demand.
• VAST_FOUNDER_MACHINE_LISTING · founder's listed machine + advertised rate · proves asset listed · does NOT prove paid revenue.
• VAST_FOUNDER_RENTAL_RECEIPT · completed rental or payout record · proves actual paid utility evidence · does NOT prove resale market value.
• VAST_FOUNDER_OCCUPANCY_HISTORY · rental duration / booked time · proves actual utilization history · does NOT prove future guaranteed utilization.
• VAST_WORKLOAD_TEST_RECEIPT · self-test, benchmark or validation receipt · proves operational/rental readiness · does NOT prove market demand.
• VAST_DERIVED_YIELD_ANALYSIS · modeled gross/net annualized yield from captured receipts · proves analysis based on disclosed inputs · does NOT prove guaranteed future earnings.

Of these seven, only VAST_FOUNDER_RENTAL_RECEIPT and VAST_FOUNDER_OCCUPANCY_HISTORY count as first-party operating evidence. All others are signal-only.

Operator-captured market snapshot · 2026-05-23 · source: vast.ai market explorer

CAPTURED 2026-05-23 · OBSERVATION ONLY · NOT PAID YIELD. A 21-card snapshot across mid-tier and pro-workstation classes. 472 listed machines, 70 vast.ai-verified hosts. Columns include Listed/Verified counts, Online %, Rented % (occupancy), Peak/Median DLPerf (vast.ai's normalized deep-learning performance score), Asking $/hr, and DLPerf/$ (performance per dollar).

Highest observed occupancy in this snapshot: RTX 5880 Ada 87.7% · RTX 4080 81.8% · RTX 5060 78.6% · RTX PRO 4500 78.2%. Highest observed performance per dollar: RTX PRO 4500 (494.9) · RTX 5060 (353.8) · RTX 4000 Ada (327.9). All values are instantaneous observations · not guarantees of sustained demand, future yield, or paid earnings.

Doctrine: Listed is not sold. Offered is not rented. Rented is not future yield. A snapshot is what survived at one moment · not what will be true tomorrow.

The RTX 3090 workhorse thesis

A prior-generation 24GB card can remain a productive AI compute asset. The thesis does NOT state rental rate, occupancy, or expected earnings as permanent truth · the platform captures Vast.ai public-rate snapshots with timestamp + source URL, collects founder-owned rental receipts when available, compares purchase/resale basis against actual paid yield when both inputs exist, and uses the record to determine whether HOLD/RENT, SELL, REDEPLOY, or REVIEW is the most defensible next action.

The yield analysis standard

Every VAST_DERIVED_YIELD_ANALYSIS carries the same input contract. If a field is missing (purchase cost, listed rate, actual paid rate, utilization period, gross earnings, electricity assumption), the analysis returns EVIDENCE_INCOMPLETE with a list of what to capture next. The platform refuses to model yield from observation data alone.

Closing doctrine

A listing rate is what was asked at a moment. An occupancy percentage is what was observed at a moment. A rental receipt is what was paid for time used. A yield analysis is what receipts and electricity together support. A deed publishes only what survives proof.