LEAP Analytics GPU INTELLIGENCE TERMINAL
NVDA
—
—
AMD
—
—
B300 SPOT
$7.00/hr
rental
B300 SECONDARY
$50K
purchase
H100 FLOOR
$14K
secondary mkt
Country Engine Sector Engine Credit Analysis LEAP Rating KYC Portal GPU Terminal
GPU MARKET VALUE
RENTAL RATES + DEAL FLOW ⬡
METHODOLOGY
What this tab answers: "We are considering financing a purchase of X GPU units at today's price. What is a credible estimate of secondary market resale value in 12, 24, and 36 months — and what is the absolute floor?" The model is calibrated from observed A100 and H100 secondary transactions, then projected forward. B300 and B200 projections are extrapolated from H100 precedent — clearly labelled.
RESALE VALUE SCHEDULE — $/UNIT
SHOW: LEAPEF QUANT MODEL
GPUPurchase Price +6 Months+12 Months+18 Months +24 Months+36 Months+48 Months FloorFloor %
B300 PROJECTED RESALE VALUE — LEAPEF QUANTITATIVE MODEL
LEAPEF model projection from April 2026. White dot = purchase price today. Dotted line = projected resale. Band = P25–P75 range of outcomes.
Purchase price today LEAPEF projected resale (P50) P25–P75 confidence band Floor
Model note: B300 and B200 projections are extrapolated from H100 precedent at equivalent market age — labelled accordingly in the table above. The same methodology applied to A100 and H100 closely tracks actual secondary market prices (see Methodology tab for validation).
What this tab answers: GPU rental rates and where they are heading. The key signal: H100 spot has found its floor and long-term contracts now price above spot — a chip that should be depreciating is being bid up on committed paper. LEAPEF's own deal flow is plotted alongside public market data as confirmation.
CURRENT RENTAL RATES — EX-AWS · 55 PROVIDERS
GETDEPLOYING.COM · MAR 2026
GPUArchVRAMSpot MinSpot AvgSpot Max†LTCProvidersBuild CostAvg/BuildStage
† Excludes outlier hyperscaler pricing. LTC = Long-Term Contract. Build Cost = 5yr TCO/hr.
RENTAL RATE FORWARD CURVE — PROJECTION FROM APRIL 2026
Dot = today. Dotted = LEAPEF model. ◆ = LEAPEF deal flow (if unlocked).
B300 FLOOR DERIVATION — HOW WE ARRIVE AT $21,500
UPDATE RATES QUARTERLY · TSMC / SK HYNIX / MICRON / ASML EARNINGS
The floor is the minimum credible secondary price — anchored to replacement cost of equivalent compute. Supply chain constraints keep replacement cost elevated. The infrastructure compatibility factor below is the key bear case: B300 requires liquid cooling and high-density power that only ~25–35% of data centres have today. Weights reflect cost significance and substitutability — not statistical regression.
ComponentMeasuresCurrent RateSourceStatusWeightEffectRationale
CoWoS PackagingTSMC packaging capacity~85K wpm · 97% utilTSMC earningsCRITICAL35%↑ FloorPrimary bottleneck. Limits production. Fewer units = higher replacement cost.
HBM3e MemorySK Hynix / Micron ASP/GB~$22/GBSK Hynix / Micron earningsTIGHT30%↑ Floor~30% of chip BOM. Higher ASP raises replacement cost directly.
ASML EUV OutputEUV machines/yr~60 machines/yrASML quarterlyCONSTRAINED15%↑ FloorHard ceiling on new fab capacity. Limits medium-term supply growth.
Helium Spot$/mcf — wafer fab~$280/mcfUSGS / GasworldWATCH10%↑ FloorQatar = 30% global supply. Hormuz closure is tail risk. Shock slows fab output.
Neon Gas$/m³ — lithography~$14/m³USGS / trade pressRECOVERING5%→ NeutralPost-Ukraine normalising. Not currently a constraint.
NVIDIA Units ShippedDC GPU units/quarter~650K/Q (est.)NVIDIA earnings (derived)RAMPING5%↓ FloorMore units = more secondary supply = modest downward pressure.
Infra Compatibility % of DCs able to run B300 without major capex ~25–35% of DCs Uptime Institute / DC surveys BEAR FACTOR — ↓↓ Floor B300 needs liquid cooling (1,200W TDP) + high-density power. Reduces addressable buyer pool from ~80% (H100) to ~30%. Smaller pool = structurally lower floor. Key bear case.
Floor Calculation — B300 · Current Conditions
Base depreciation model floor$21,000
CoWoS critical (35%)+ $3,150
HBM3e tight $22/GB (30%)+ $2,700
ASML constrained (15%)+ $1,350
Helium watch (10%)+ $900
Neon neutral (5%)$0
Units ramping (5%)− $600
Supply factors subtotal$28,500
Infrastructure discount (25–35% buyer pool)− $7,000
DERIVED B300 FLOOR $21,500
Range: $19,000–$24,000 depending on liquid-cooled DC capacity expansion rate. Update quarterly.
Why These Weights
CoWoS 35% / HBM 30% — 65% combined. Only components with no near-term substitutes. CoWoS is a TSMC monopoly. HBM is 3 suppliers, SK Hynix ~50% share.

ASML 15% — sets ceiling on how fast the supply constraint can be relieved. Even if TSMC wanted to expand, EUV machines are the gating factor.

Helium 10% / Neon 5% — smaller contribution but geopolitical shock potential. Qatar helium concentration is a tail risk worth monitoring.

Infrastructure discount — applied after supply factors. H100 floor ratio was 47% because ~80% of DCs could run it. B300 floor ratio is lower because only ~30% can. The discount bridges that structural difference.
LEAPEF DEAL FLOW — ACTIVE TRANSACTIONS ⬡
SESSION ONLY · NOT STORED · PIN GATED
Enter LEAPEF deal prices to plot them on the rental chart above.
Anonymised ◆ markers only — no counterparty or volume visible externally.
1. Purpose
LEAPEF's internal platform for GPU-backed structured finance. Used to build conviction on asset pricing, stress-test assumptions before investor calls, and generate the analytical basis for term sheets. The Methodology tab is shared externally when investors ask "how did you arrive at this number?"
2. Secondary Market Price Model
SECONDARY_PRICE(t) = (P₀ − F_eff) × e^(−λ · t) + F_eff

P₀ = Purchase price today · F_eff = Effective floor (supply-chain derived) · λ = Decay rate · t = Months from purchase
Calibration: A100 (4+ years observed, floor confirmed ~$3,800) and H100 (27 months observed, floor forming ~$14,000–18,000). B300 and B200 are extrapolated from H100 precedent at equivalent age — clearly labelled as projections, not observations.
3. Infrastructure Compatibility — The Key Bear Case
B300 requires liquid cooling (mandatory at 1,200W TDP) and high-density power infrastructure. Only ~25–35% of data centres can host B300 without major capital investment (estimated $150–700K per rack retrofit). This structurally reduces the addressable secondary buyer pool from ~80% of data centres (H100) to ~30% (B300). A smaller buyer pool means lower floor relative to a pure supply-chain analysis. The $7,000 infrastructure discount in the floor derivation accounts for this.
4. Generation Transition Analysis
Each generation transition has caused progressively less depreciation pressure on the prior generation. The inference market absorbs supply faster with each cycle — but B300's infrastructure constraint partially offsets this positive trend.
V100 → A100 (2020)
Strong acceleration. V100 dropped ~60% in 18 months post-A100. Inference market too small to absorb displaced supply. Floor: ~10–15% of launch price.
A100 → H100 (2023)
Moderate impact. A100 declined but stabilised at $3,800–4,500 (32% of launch). Growing inference demand absorbed displaced workloads. Floor held — consistent with model.
H100 → B200 (2024–25)
Minimal impact. H100 LTC rate actually rose after B200 launched. Demand now outpaces supply across generations. B300 benefits from this trend — with the infrastructure caveat.
B300 caveat: The positive trend (less acceleration each cycle) is offset by the infrastructure compatibility constraint. B300's smaller buyer pool means any transition effect hits harder in the secondary market. This is why the infrastructure discount is applied to the floor — not despite the positive trend, but in addition to it.
5. Model Validation — A100 and H100 Back-Test
The same model applied to A100 and H100 closely tracks actual secondary market prices — validating the methodology before applying it to B300.
Solid line = observed secondary market prices. Dotted line = model prediction. Close tracking = well-calibrated model.
6. Monte Carlo Bands
P(t+1) = max[(P₀−F_eff)·e^(−λ·t) + F_eff + σ·√(1/12)·Z, F_eff·0.80] · Z~N(0,1) · 1,000 paths
P50 = central estimate. P25 = cautious case for covenant floor analysis. The cone widens over time — genuine uncertainty, not imprecision.
7. LEAPEF Deal Flow Index
DFI = Σᵢ[wᵢ×(contract_priceᵢ/market_priceᵢ)] / Σwᵢ · wᵢ = units × confidence
DFI > 1.0 = LEAPEF deals above public market (bullish) · DFI < 1.0 = caution signal
PIN-gated. Session-only. Never stored or transmitted. The DFI value and anonymised chart markers are the only external outputs.
Disclaimer

LEAPEF AB internal analytical tool. Not investment advice, a credit rating, or a formal valuation. All projections are probabilistic estimates. LEAPEF AB is a registered AIFM (Finansinspektionen). March 2026.