How much does reverse-auction GPU save you?

DCS Compute runs a reverse-auction lane: instead of paying hyperscaler rack rates, workers bid down for your job. We've measured 12.75% average savings on production traffic. Plug in your workload below — every saving claim is receipt-signed and Basescan-verifiable.

⚡ DCS COMPUTE · REVERSE-AUCTION ROI CALCULATOR

GPU Type

GPU Hours / Month

Current Provider

Workload Tolerance

Monthly savings

$2,194

12.7% vs current

Current monthly cost$17,280

DCS reverse-auction cost$15,086

Annual savings$26,328

Avg cost/hour$20.95/hr

Estimate based on 12.75% measured savings (beta-gate ON). Real savings depend on workload + bid response. Every receipt verifiable on Basescan.

Methodology

How the auction works (in 4 steps).

①

Job posted

You submit a job to /v1/compute/jobs with GPU type, duration, tolerance.

②

Workers bid

Available DCS workers (in your region) bid down within a 5-second window. Receipt per bid.

③

Best bid wins

Lowest bid above the worker's minimum runs your job. Reserve price respected.

④

Settlement

Worker delivers, you accept, settlement on Base mainnet. 97.5% worker, 2.5% treasury.

Verified savings

Numbers you can audit.

The 12.75% figure isn't marketing. It's the rolling 30-day average across production runs. Every saving claim has a receipt CID; verify yourself.

Last 5 production runs (anonymized)

Workload	Hyperscale	DCS	Saved	Receipt
Llama fine-tune (48h H100)	$1,152	$986	14.4%	bafy…3a9f
Video render (12h L40S)	$65	$56	13.8%	bafy…1f0e
Diffusion batch (6h A100)	$84	$74	11.9%	bafy…4d18
Inference batch (24h H200)	$576	$507	12.0%	bafy…7c4a
3D model train (72h H100)	$1,728	$1,498	13.3%	bafy…b91d

Verify any row with npx @dcsplatform/r2-verify <cid>

FAQ

Common questions.

Why only 12.75% and not 50%?

Hyperscalers are not that overpriced once you factor in reliability + region count. We're competing on the margin, not the gross. 12.75% on a $200K/year bill is $25K — buys an engineer half a quarter.

What's the latency cost?

Best-effort mode: bid window is 5 sec, total provisioning ~30 sec. Balanced: instant if there's available capacity in your region. Strict latency: same as hyperscaler (we provision upfront, no auction).

What happens if the auction has no bids?

Falls through to a fallback pool at hyperscale rates. You're never stuck waiting — but you pay rack rate when there's no bid. Receipts mark which mode actually ran.

Is the 12.75% sustainable or is it loss-leading?

Sustainable — DCS treasury takes a 2.5% commission on settled jobs, workers earn at their reserve price. Both sides win because hyperscalers carry overhead that distributed workers don't.

Stop paying rack rate for GPU time.

The reverse-auction lane is live. Try it on a non-critical workload first.

See DCS Compute → API docs