Pricing model v0.2

Pricing

A three-part price: token volume, GPU time billed per second, and a demand surge k by GPU class. The token API stays — on the same GPUs, with no double billing.

One price formula

Price = Fee + max( Min; Base × t × k + Σ tokens × k_tok )

t — hours, billed per second · k — GPU class surge, 0.7–2.0 · k_tok = 1 + 0.5 × (k − 1), band 0.85–1.5 · AMD at the CBA rate

  • Fee and minimum. Booking fee $0.10 per session; per-second billing, 60 s minimum; future-window bookings: 10% prepayment, credited.
  • Each GPU-second billed once. A request is dedicated or shared, never both; cold starts and platform retries are not billed.
  • Hybrid: θ = 0.4. Core = 0.4 × Base × t + r_n × N; price = min(Core, Base × t) × k — total ≤ pure rental.

Base rates per GPU-hour

  • H100$2.70
  • H200$3.30
  • A100$1.50
  • L40S$0.90
  • RTX 4090$0.42
  • Mac Studio$0.35

Monthly: base = 0.95–0.98 × index, step ≤ ±10%. H100 ceiling of $5.40 — below AWS Capacity Blocks ($5.97) and CoreWeave ($6.16).

Surge k: three regimes, bounds 0.7–2.0

  • Quiet hour, utilization < 35%: k down to 0.7 (−30%).
  • Neutral, 35–75% utilization: k = 1.0 — no surge.
  • Peak, > 75%: up to 1.75 by utilization, plus a queue term of up to +0.25; clipped to 2.0.
  • Compute every 1 min → publish every 5 min; at most +0.15 / −0.20 per publication.
  • Quote valid for 10 min, together with the AMD rate; a booking keeps its k for the full term.

Tiers — and who benefits

  • Night window 01:00–08:00

    Yerevan time: k ≤ 0.85, tokens k_tok ≤ 0.75

    Buyer: −15…−30% on batch jobs and training

    Seller: never below the 65%-of-base reserve

  • Spot and bids

    max(0.35, 0.5·k) × base; 2-min stop notice

    Buyer: −50% or more, single clearing price

    Seller: idle time into money, own minimum

  • Reserved and commitments

    1–12 months: k = 1.0, discount 5 / 12 / 20 / 30%

    Buyer: a discount and a predictable bill

    Seller: guaranteed utilization

  • On-demand

    k per hourly block, rise ≤ 0.3 per block

    Buyer: one-off jobs with no term, price shown upfront

    Seller: 60% of surge at peak

Seller payout and platform share

Seller payout = max( 65% of base; 80% of base × min(k, 1) + 60% × base × (k − 1) )

The platform keeps the rest: 20% of the base part and 40% of surge; in quiet hours its share drops to 7%, and the seller never gets less than the reserve.

H100 × 3 hours: quiet · normal · peak

kBuyerSellerPlatform
Quiet · 0.70$5.77$5.265$0.505
Normal · 1.00$8.20$6.48$1.72
Peak 18:00 · 1.50$12.25$8.91$3.34

Includes the $0.10 fee. Platform share of the rate: 7% quiet → 20% normal → 27% peak. The platform funds the quiet-hour discount; the seller never gets less than 65% of base.

PowerPoint · 17 · v0.2

Download the pricing presentation

Скачать презентацию · Ներբեռնել ներկայացումը · Download

The T-Exchange three-part pricing model v0.2 in Russian, English and Armenian: 17 slides, no sign-in required.

Pilot billing units today

Until the staged launch reaches live mode, the pilot rental flow and demo bill GPU compute per GPU-hour, allocated disk per decimal GB-hour and ingress / egress per GB, with an illustrative pilot rate bridge. Rates and the surge above are pricing model v0.2 (draft) and are illustrative, not quotes.

Ask