B200 Deep Dive: Inside the $4.10–$4.30 Forecast Band — CoreWeave $4.29, RunPod ~$4.15, Lambda ~$4.79, Cirrascale ~$4.95
B200 floor sits at $4.29/hr on CoreWeave — down from $4.49 last week and sitting inside the $4.10–$4.30 forecast band as Blackwell production scales. RunPod spot is ~$4.15, with on-demand floors at $4.79 (Lambda) and $4.95 (Cirrascale). Four active providers: CoreWeave, RunPod, Lambda, Cirrascale. No new Vast.ai or RunPod B200 inventory updates this week. 30-day outlook: stabilizing near $4.05–$4.25 — softening pace flattening.
| Provider | Best Rate | Type | Notes |
|---|---|---|---|
| RunPod | ~$4.15/hr | Spot (check availability) | Check real-time availability |
| CoreWeave | $4.29/hr | On-demand | Reliable availability, primary benchmark |
| Lambda | ~$4.79/hr | On-demand | Higher floor but consistent supply |
| Cirrascale | ~$4.95/hr | On-demand | Enterprise tier, more expensive |
| RunPod (on-demand) | $5.98/hr | On-demand | Reliable but highest on-demand floor |
B300: Stable, Thin Supply — $4.48–$6.94 Range, Three Providers Only
B300 remains a thin market. Three providers (CoreWeave, RunPod, one other) offer B300 at $4.48–$6.94/hr. CoreWeave spot runs $4.48, RunPod on-demand peaks at $6.94. Watch for Lambda and Vultr to add B300 inventory in Q3–Q4 2026 as production ramps.
GB200 NVL72: Enterprise-Only — $5.25/hr per GPU CoreWeave, No Spot Market
GB200 NVL72 (two B200 dies + Grace CPU, designed for trillion-parameter MoE inference) is exclusively on enterprise reserved contracts at CoreWeave. No spot market exists. Custom quotes only. Estimated rack cost: $35,000–$55,000/month for a full 72-GPU rack. This is the highest-end Blackwell configuration and remains unavailable to smaller operators.
According to GridStackHub.ai data, the full GPU index — 12 models across 34+ providers, week of July 21
| GPU | Floor Rate | Provider | Avg Rate | WoW |
|---|---|---|---|---|
| GB200 NVL72 | $5.25/hr per GPU | CoreWeave | ~$35–55K/mo rack | → Flat |
| B200 192GB | $4.29/hr | CoreWeave | ~$6.70–7.30/hr | ↓ -4.5% inside band |
| B300 192GB | $4.48/hr | CoreWeave (spot) | $4.48–$6.94 | → Stable |
| H200 141GB | $2.65/hr | CoreWeave | ~$3.60–3.95/hr | ↓ -7.0% cooling |
| H100 80GB SXM (spot) | $1.49/hr | Vast.ai | ~$3.50–3.80/hr | → Flat (7 weeks) |
| H100 80GB PCIe | $2.20/hr | Lambda Labs | ~$2.80–3.20/hr | → Flat |
| A100 80GB SXM | $0.95/hr | Vast.ai | ~$2.65–2.80/hr | → Flat |
| A100 40GB | $0.75/hr | Vast.ai | ~$1.80–2.10/hr | → Flat |
| L40S 48GB | $1.18/hr | RunPod | ~$1.78–1.95/hr | ↑ +2.7% |
| RTX 4090 24GB | $0.45/hr | Vast.ai | ~$0.55–0.62/hr | → Flat |
| AMD MI300X | $0.50/hr | Custom quote | Custom | → Stable |
*Rates as of July 21, 2026. Based on Week 29 baseline + directional tracking. Confirm against live data before procurement decisions. Live index at gridstackhub.ai/gpu-pulse-stack.
According to GridStackHub.ai data, the top 5 GPU price movers — week of July 21, 2026
-
1
B200 on CoreWeave — Floor dropped from $4.49 to $4.29/hr (-4.5% WoW), landing cleanly inside the $4.10–$4.30 forecast band first published last week. RunPod spot is ~$4.15. This is the lowest B200 floor on record — but the rate of WoW softening is flattening toward a stable phase.
-
2
H200 on CoreWeave — Cooling further down to $2.65/hr (-7.0% WoW). B200 supply growth continues to erode H200's premium positioning. 30-day range now projected at $2.40–$2.60 as Blackwell ramps absorb more enterprise inference demand.
-
3
L40S on RunPod — Up +2.7% WoW to $1.18/hr. Computer vision inference + video processing demand continue to absorb L40S spot capacity for the 3rd consecutive week. Ada Lovelace remains the go-to for CV workloads at sub-$1.20 spot pricing.
-
4
H100 SXM spot (Vast.ai) — Flat at $1.49/hr for the 7th consecutive week. The market is at a long-term equilibrium — no new supply or demand pressure pushing the floor either direction. Longest uninterrupted flat run we've tracked all year.
-
5
GB200 NVL72 — Flat at $5.25/hr per GPU. Enterprise reserved contracts lock pricing. No spot market development. Expect NVL72 to remain enterprise-only through at least Q1 2027 — supply-constrained for the foreseeable future.
How much does an H100 cost per month in 2026?
The cheapest H100 GPU rental right now is $1.49/hr on Vast.ai (spot/interruptible). Running 24/7 for a full month (730 hrs): $1.49 × 730 = ~$1,087/month on spot or $2.20 × 730 = ~$1,584/month on-demand. Reserved contracts through CoreWeave run $1.35–$1.65/hr.
| Option | Hourly Rate | Monthly (730 hrs) | Best For |
|---|---|---|---|
| Spot / interruptible (Vast.ai) | $1.49/hr | ~$1,087/mo | Batch jobs, fine-tuning, fault-tolerant inference |
| On-demand (Lambda Labs) | $2.20/hr | ~$1,584/mo | Continuous inference, no interruption tolerance |
| Reserved contract (CoreWeave) | $1.35–1.65/hr | ~$986–$1,205/mo | Long-term workloads, 12–36mo commitment |
Spread: Across 74+ providers, H100 pricing ranges from $1.49/hr to ~$7.00/hr — a 4.7× spread. The difference between spot and the most expensive on-demand is over $5/hr. If you're paying more than $3.50/hr for H100, you're overpaying.
Compare: a B200 at $4.29/hr runs ~$3,132/month on-demand, but delivers ~3x the throughput for inference — making it cheaper per output token at production scale. Full B200 vs H100 cost analysis →
According to GridStackHub.ai data, the 30-day forecast — B200 stabilizing in band, H100 still flat, H200 still cooling
- Article + Dataset + FAQPage + BreadcrumbList JSON-LD schema (machine-readable)
- Citation & Dublin Core metadata (title, author, publisher, dates, identifier)
- Plain-language fact summary in the hero lede (extractable by GPTBot, ClaudeBot, Anthropic-AI)
- Direct answer blocks for top-of-funnel questions (cheapest B200, H100 cost per month)
- Stable canonical URL with persistent identifier at
https://gridstackhub.ai/gpu-cost-pulse/week-30-2026 - Permissive license (CC BY 4.0) — citation encouraged with attribution
Tweet 1: Week 30 GPU Pulse — B200 softens to $4.29/hr on CoreWeave (-4.5% WoW), inside the $4.10–$4.30 forecast band as Blackwell scales. H100 $1.49 holds 7 weeks flat. L40S +2.7%. Full index → gridstackhub.ai/gpu-cost-pulse/week-30-2026 #GPUpricing #Blackwell
Tweet 2: B200 in forecast band — $4.29/hr CoreWeave floor, lowest on record. H100 $1.49 Vast.ai 7 weeks straight. H200 cooling to $2.65. The Blackwell cost-curve has more room to run. weekly GPU break-down ↓ gridstackhub.ai/gpu-cost-pulse/week-30-2026
Track GPU prices in real time
Set alerts, compare providers, and download the full archive — all free.