Loose, Not Flooded
GPU rentals dipped 5% this week. The reason isn't new H100 supply.
GPU rental prices eased week over week on Aug 22, but the tape does not show a new H100 supply flood. The Compute Tightness Index fell to 47.1 Balanced from 54.8 a week earlier, still firmer than a month ago, while GPU rental prices split by generation. H100 and H200 softened, B200 on-demand rose 3.2%. For NVDA, that is bifurcated pricing power. For startups, it is cheaper to rent H100s this week, less reason to rush to buy.
What the tape shows
The blended Compute Tightness Index (CTI, 0 to 100, higher is tighter) printed 47.1 on Aug 22, regime Balanced. That is down 7.7 points week over week from 54.8 on Aug 15, but up 2.4 points from about 44.7 on Jul 9. The 30-day change is essentially flat at +0.4 points, so this week is a give-back inside a July-to-August climb, not a new downtrend.
Per-GPU on Aug 22:
| GPU | On-demand avg | Spot avg | Spot discount vs on-demand | Listings | CTI | Regime |
|---|---|---|---|---|---|---|
| H100 | $3.86 | $1.75 | 55% | 158 | 40.0 | Loose |
| H200 | $4.32 | $2.97 | 31% | 71 | 59.2 | Balanced |
| B200 | $7.05 | $3.82 | 46% | 43 | 53.2 | Balanced |
Week over week (Aug 22 vs Aug 15):
| GPU | On-demand WoW | Spot WoW | Listings change | Unavailable change |
|---|---|---|---|---|
| H100 | -1.0% | -4.4% | +6 (+3.9%) | 36 vs 118 (-69%) |
| H200 | -2.3% | -13.2% | -3 | 19 vs 54 (-65%) |
| B200 | +3.2% | +0.5% | +3 | 42 vs 88 (-52%) |
A simple three-GPU spot average was $2.85 vs $3.02 a week ago (-5.6%). On-demand averaged $5.08 vs $5.05 (+0.5%). Your $2.59 / -9.4% basket is directionally consistent but larger than this three-GPU cut, likely a different weighting.
Source: Bargo Compute Tightness Index daily history, as of Aug 22, 2026.
Supply vs demand: why it loosened
Not a H100 supply flood. H100 listings rose only 6 week over week, H200 listings fell. The 30-day CTI delta is flat for H100 (-0.1) and H200 (+0.2), firmer for B200 (+1.8). No surge in new supply across the board.
Availability normalized. The biggest WoW move was unavailable counts collapsing across all three GPUs, more capacity showing as available while listings barely moved. That reads as loosening, but the Aug 21 to Aug 22 drop was unusually sharp (H100 unavailable 151 to 36 in one day), so watch for a snap-back next week before calling it structural.
Demand softened short term, still strong medium term. Token Demand Index on this slice was 66.3T tokens per week on Aug 21 vs 70.2T on Aug 14 (-5.5% WoW), but vs 42.8T on Jul 13 (+54.8% MoM). Effective price per 1M tokens was $1.13 vs $1.16 WoW (-2.8%) and vs $2.31 a month ago (-51%). Inference pricing over about 30 days fell broadly, OpenAI -25.5%, Claude -40%, Google -12.2%, while open-source rose 4.2%. Falling token prices with rising volume is the classic Jevons pattern, price down, volume up, not demand collapse.
This pattern echoes recent coverage of how shortages are spreading sideways to CPUs and fab tools rather than easing uniformly, see The Shortage Is Spreading Sideways, and how suppliers are funding expansion differently, see The Supply Side Voted With Its Wallet.
What it means for NVDA pricing power
Bifurcated. B200 on-demand +3.2% WoW and spot +0.5% with CTI 53.2 Balanced shows frontier pricing power intact. H100 at CTI 40.0 Loose and 55% spot discount, and H200 spot -13.2%, show prior-generation discounting. That is consistent with a Balanced overall market, not Tight, where buyers can wait for H100 deals but still pay up for B200.
For context on how hyperscaler commitments frame NVDA demand beyond spot rentals, see Microsoft's 10GW Awakening.
What it means for AI startup CapEx
Cheaper H100 at $1.75 spot and H200 at $2.97 spot directly lowers near-term inference and training burn, so renting looks more attractive than buying this week. That defers CapEx urgency for startups that can stay on rented capacity.
It does not change the contracted buildout. The rental move is spot, the buildout is contracted. Startups renting from neoclouds benefit from lower prices now, but neoclouds themselves live on continuous issuance, so a sustained rental decline would pressure their utilization before it helps renters. Near term, rent vs build tilts to rent. Medium term, the decision still hinges on whether token volume re-accelerates after this -5.5% WoW dip.
What to watch
- Does unavailable stay low next week or snap back, confirming real loosening vs a data blip
- B200 listings and price, if B200 rolls over, frontier pricing is cracking, so far it has not
- Token demand WoW, needs to push back above 70T to confirm this was a soft patch, MoM +55% keeps the uptrend intact
More research at bargo.ai/research.
Sources: Bargo GPU rental prices snapshot Aug 22 vs Aug 15, Bargo Compute Tightness Index history, Bargo inference economics pricing.