Source Report

H100 Spot Market Pricing Trends (2023–Present)

NVIDIA H100 GPUs launched amid extreme scarcity in late 2022/early 2023, driving spot rental rates to $7.50–$10/GPU-hour by Q4 2024 through hyperscaler-only supply and AI training demand; as production ramped to millions of units annually by 2025, spot prices collapsed 64–75% to $1.49–$3.50/GPU-hour (market low $0.73, high $14.90), with indices like Silicon Data's SDH100RT dropping from $9.34 (2024 peak) to $2.20 by early 2026—reflecting oversupply, competition from marketplaces like Vast.ai, and a shift to inference workloads that tolerate spot volatility over premium on-demand.[1][2][3][4][5][6]
- 2023 Q3–Q4: $7.62–$7.77 (hyperscalers only, e.g., AWS ~$7.50/GPU-hr)[6]
- 2024 Q1–Q2: $7.37–$9.39 peak (scarcity max; Silicon Index $9.34 med)[6]
- 2025 Q1–Q2: $3.50–$7.00 (supply eases; Vast.ai dips to $1.49 promo)[1][7]
- 2025 Q3–Q4: $2.36–$2.85 (stabilizes; Vast.ai $1.80–$2.50)[5]
- 2026 Q1: $1.38–$2.20 (10% Jan spike to $2.20, then softens; Thunder Compute $1.38 on-demand)[2][3]

Implication for entrants: Spot markets now viable for non-urgent inference (e.g., Vast.ai/RunPod < $2/hr), but volatility (e.g., 10% weekly swings) demands auto-scaling tools; new buyers avoid 2023–24 peaks by renting spot vs. buying at $25K–$40K new.

CoreWeave and Oracle Rental Rates Over Time

CoreWeave leverages long-term fixed-price contracts (e.g., $3–$4.25/GPU-hr assumed in models) to insulate from spot crashes, maintaining premiums like $4.76–$6.16/GPU-hr on HGX/InfiniBand clusters into 2026 despite market-wide 64% drops; Oracle's bare-metal BM.GPU.H100.8 stays hyperscaler-high at $10/GPU-hr ($80/hr for 8xH100), prioritizing reliability over spot competition as supply floods dilute Hopper's scarcity moat.[2][7]
- CoreWeave: 2023–24 ~$4.25 (avg modeled); 2025 $6.16 (8xHGX); 2026 $4.76 (on-demand, premium for scale/reliability)[8][2]
- Oracle: Consistent $10/GPU-hr (bare-metal 8x, no smaller SKUs); no major cuts despite 2025 supply surge[2]
- Contrast: Spot (Vast.ai) fell to $1.49–$1.87 by late 2025[7]

Implication for competitors: CoreWeave's Nvidia partnerships lock in above-spot rates (10–15% premium via Slurm/K8s), but entrants must match hyperscaler reliability at sub-$2/hr or niche (e.g., spot-only) to capture inference overflow.

Analyst Estimates of H100 Depreciation Cycles

Analysts peg H100 economic life at 2–3 years (frontier utility) despite 5–6 year accounting (e.g., CoreWeave's 6-yr straight-line), with secondary resale hitting a "mid-life cliff": 2-yr used H100s at 61–75% of new (~$15K–$28K vs. $25K–$40K), dropping to 45–55% by year 3 (~$11K–$22K) as Blackwell ramps cannibalize demand—validating bears' predictions over Patel's efficiency thesis, since rental yields now barely cover capex at sub-$2/hr.[9][10][11][12][13]
- Year 1–2: Refurb 80–90%, used 65–75% retention[10]
- Year 3+: Used 45–55%; eBay ~$6K (85% loss from $40K peak)[10]
- Hyperscalers: Varied (Meta 5.5yr extend, Amazon 5yr shorten); Michael Burry flags 2–3yr real life[14]

Implication for entrants: Buy used/refurb at 50–70% off new for inference (break-even ~14mo at $2.50/hr, 100% util), but plan 2yr flips before Blackwell flood; avoid long holds as resale cliffs erase ROI.

Evaluating Patel's Claim: Efficiency vs. Blackwell Depression

Dylan Patel (SemiAnalysis) argues H100s hold/gain value as models like GPT-5.4 serve "more valuable tokens per GPU" via sparsity/MoE efficiency (sparser active params, lower cost/run), enabling $2.40/3yr contracts despite spot crash to $1.38–$2.20—but data refutes: rentals down 70%+ from 2023 ($7.76 → $2/hr), used resale 45–85% depreciated, and B200/GB200 (1.6–1.7x TCO/GPU, 2.5x perf) already shifting premiums (H100 to "workhorse," 10–20% further drop 2026); bears win as supply > efficiency gains.[15][10][12][1]
- Patel mechanism: GPT-5.4 sparser/cheaper than GPT-4 → more/high-value tokens/H100 → sustained $2+/hr deals[15]
- Counter: Spot $1.49–$2.36 (Silicon Index); used $15K–$28K (55–70% new); B200 2026 pressures[1][6]

Implication for competitors: Patel's right on software-lift (efficiency buoys rentals > pure hardware decay), but bears dominate hardware economics—enter via inference niches or Blackwell early to avoid Hopper glut.

Blackwell's Depressive Impact on Hopper Values

Blackwell (B200/GB200 NVL72) launches at $30K–$50K/GPU (1.6–1.7x H100 capex/GPU) with 2.5x AI perf, 25–50x energy/token efficiency, and rack-scale (e.g., $3.9M/72-GPU), immediately demoting H100 to mid-tier: forecasts 10–20% Hopper rental/purchase drops in 2026 as enterprises upgrade, with H100 spot stabilizing $1.65–$2.50 floor (breakeven for owners) amid B200 scarcity premiums.[1][16][17]
- B200 impact: H100 cloud $2.75–$3.25 Q1'26 → 10–20% lower mid-year[1]
- TCO: GB200 1.6x H100/GPU but scales inference/training better[17]

Implication for entrants: Hopper fire-sale buys now (used $15K+), but pivot to Blackwell leasing by mid-2026; compete on software (e.g., MoE optimization) to extract Patel-style value from depreciating Hopper fleets. Confidence high on trends (web-verified 2023–26); resale estimates medium (secondary data sparse).

---

Recent Findings Supplement (March 2026)

H100 Spot and Rental Pricing Trends (Q4 2025 - Q1 2026)

Silicon Data's H100 Rental Index captured a volatile correction mechanism: after stabilizing at ~$2.36/hr in June 2025 (down 23% from $3.06/hr in Sep 2024), prices dipped to $2.00/hr by Dec 9, 2025, then spiked 10% to $2.20/hr by Jan 6, 2026—driven by isolated demand surges while A100/B200 held steady—before broader spot markets pushed lows to $0.73/hr amid 300+ new providers flooding capacity.[1][2][3]
- GetDeploying tracks H100 (80GB) across 54 providers at $0.73-$14.90/hr (Mar 13, 2026), with 36+ offering it; Thunder Compute leads on-demand at $1.38/hr.[4][5]
- CoreWeave HGX H100 (8x): $49.24/hr or ~$6.16/GPU-hr on-demand (Feb 2026); PCIe at $4.25-$4.76/GPU-hr; Oracle BM.GPU.H100.8 (8x): $80/hr or $10/GPU-hr.[6][5]
- Spot lows: Vast.ai ~$1.47/hr, RunPod $1.87-$2.39/hr; hyperscalers like GCP/AWS spot at $2.25-$2.50/hr (Dec 2025).[5]

Implication for competitors/entrants: 64-75% rental drops from 2024 peaks ($8-10/hr) reflect supply catching demand, commoditizing H100 for inference/fine-tuning; new entrants can undercut via spot marketplaces (sub-$2/hr) but face 40%+ preemption risk—favor flexible workloads over training.

H100 Resale/Secondary Market Depreciation (Late 2025-Early 2026)

Secondary listings analyzed by Silicon Data (Jun 2024-Dec 2025) reveal non-linear "step-change" depreciation: retail stuck at $25K-$40K, but used H100s plunged from low-$40Ks (mid-2024) to mid-$20Ks (late 2025), with 2-3yr-old units at 20-40% of peak new prices; eBay grey market hit $6K-$9K (~85% drop from $40K MSRP) by early 2026, accelerated by Blackwell's 11x inference efficiency.[7][8]
- Used: 45-75% of contemporaneous new (~$15K-$28K for 2-3yr); Refurbished: 70-90% premium (~$21K-$34K), but gap widens post-2yrs.[9][10]
- Late 2025/early 2026: Brief rebound then sharp dip on supply waves; B200/B300 entry compresses values further (H100 to "value tier").[11]

Implication for competitors/entrants: No appreciation vs 2023 ($30K+ new, premiums to $50K); bears correct—Blackwell oversupply tanks resale 75-85%, favoring renters over buyers; flip used gear now (50-70% recovery) before Rubin (late 2026) hits.

CoreWeave/Oracle Rental Evolution (Post-Sep 2025)

CoreWeave maintained premium pricing amid corrections: HGX H100 at $49.24/hr (8x, ~$6.16/GPU-hr, Feb 2026), up from ~$3/hr projections (late 2024) via InfiniBand/HPC focus; PCIe variants $4.25-$4.76/GPU-hr; commitments yield 30-50% off but lock-in required.[6][12]
- Oracle BM.GPU.H100.8: Steady $10/GPU-hr ($80/hr node), no single-GPU; AI Enterprise add-on $2.50/GPU-hr (2026 list).[13]
- Trends: CoreWeave volumes rose (e.g., $22B+ OpenAI/Meta deals), holding 10-15% premium vs Lambda/Nebius; Oracle bare-metal waives overhead but all-or-nothing scales poorly for small jobs.[14]

Implication for competitors/entrants: CoreWeave's moat (utilization >80%, rebookings at 95% original) sustains premiums despite 60% market drops; Oracle suits HPC bursts but hyperscaler pricing (2-3x specialists) kills TCO—target neoclouds for 50%+ savings.

Analyst Estimates on H100 Depreciation Cycles

Analysts peg economic life at 2-3yrs (Burry: hyperscalers understate $176B via 5-6yr books vs reality; Barclays trims 2025 earnings 10% on realism), but CoreWeave counters with 6yr viability—A100s (2020) still booked, H100s re-rent at 95% post-contract—via "cascade" to inference.[15][16]
- Schedules: MSFT/GOOG/ORCL 6yr (up from 4); AMZN reversed to 5yr (Feb 2025); neoclouds 4-6yr; physical 7-9yr but obsolescence caps at 18-24mo cycles (H100→Blackwell→Rubin).[17]
- 2026 forecasts: 10-20% further rental drops on B200 scale; resale accelerates post-2yrs (non-linear, 40%+ annual).[11]

Implication for competitors/entrants: Patel's claim (H100 >3yr-ago value via GPT-5.4 efficiency) fails—2023 new ~$30K vs 2026 used $6-20K, rentals 70% off; bears vindicated by Blackwell (11x infer eff.); enter via resale flips or spot inference, avoid capex.

Blackwell's Depressive Impact on Hopper Values

Blackwell (B200/B300) repriced H100 as "mainstream workhorse": 2.5-11x infer perf (e.g., B300 11x cheaper run-cost), flooding secondary with upgrades; H100 rentals expected 10-20% drop in 2026 as B200 availability rises, resale to 20-40% peak.[18][11]
- Q1 2026: B200 ~2x H100 street ($36-44K vs $18-22K used), but 4x FP8 throughput shifts enterprise upgrades; power economics kill old H100s in constrained DCs.[19]
- No offsetting model efficiency moat: Rentals collapsed 64% despite demand; secondary confirms obsolescence.[20]

Implication for competitors/entrants: Bears win—falling Blackwell (~Q1 2026 GA) depresses Hopper 10-20% more; Patel wrong (no value rise); arbitrage via H100 spot ($1-2/hr) for non-frontier tasks while Blackwell premiums fade. Confidence high on pricing (web-verified Mar 2026); med on Patel (snippet match, no direct 3yr claim post-9/13/25).