According to GridStackHub.ai data for the week of June 15, 2026, the cheapest B200 GPU rental drops to $2.58/hr on Vultr — down another 7.5% week-over-week as Blackwell supply crosses the inflection point. Three weeks ago, CoreWeave listed B200 at $4.26/hr. Today that same workload costs 39.4% less. The consumer spot market is still hot: RTX 4090 spot on Vast.ai surged a third consecutive week, and RTX 3090 is up again after last month's 131% spike. We track 32 providers and 396 pricing records every week so you don't have to.
The Number That Matters This Week
B200 at $2.58/hr. Just three weeks ago CoreWeave listed B200 at $4.26/hr — making this a 39.4% price collapse in 21 days. That's not a discount. That's a regime change. Blackwell is no longer premium-only; it's the new default inference tier.
Blackwell Deep Dive: B200 / B300 / GB200
Blackwell pricing is in free-fall relative to its launch trajectory. B200 and B300 are converging fast, with the B300 now representing the cheapest entry into Blackwell architecture.
| GPU | Provider | On-Demand | WoW Change | 1-Yr Commit | Memory |
|---|---|---|---|---|---|
| B200 192GB | Vultr | $2.58/hr | ▼ -7.5% | $1.99/hr | 192GB HBM3e |
| B200 192GB | Lambda Labs | $2.91/hr | ▼ -4.6% | $2.24/hr | 192GB HBM3e |
| B200 192GB | CoreWeave | $3.10/hr | ▼ -3.2% | $2.39/hr | 192GB HBM3e |
| B300 144GB | Lambda Labs | $2.31/hr | ▼ -5.7% | $1.79/hr | 144GB HBM3e |
| GB200 NVL72 | CoreWeave | $7.40/hr | ➡ Flat | — | 1.44TB HBM3e |
| GB200 NVL72 | Crusoe | $6.95/hr | NEW | — | 1.44TB HBM3e |
Week 25 Blackwell headline: CoreWeave's B200 at $4.26/hr (Week 22) was the ceiling. At $3.10/hr today it's already down 27% — and Vultr at $2.58/hr shows where the floor is heading. Crusoe entering GB200 territory with a cheaper $6.95/hr list is the first real competition on the NVL72 tier.
Full GPU Index — Week 25
32 providers · 396 pricing records · Updated weekly
| GPU | Cheapest Provider | On-Demand | WoW Change | 1-Yr Commit | Memory |
|---|---|---|---|---|---|
| GB200 NVL72 | Crusoe | $6.95/hr | NEW | — | 1.44TB HBM3e |
| B200 192GB | Vultr | $2.58/hr | ▼ -7.5% | $1.99/hr | 192GB HBM3e |
| B300 144GB | Lambda Labs | $2.31/hr | ▼ -5.7% | $1.79/hr | 144GB HBM3e |
| H100 SXM 80GB | CoreWeave | $2.14/hr | ▼ -2.3% | $1.74/hr | 80GB HBM3e |
| H200 141GB | Lambda Labs | $1.84/hr | ▼ -2.6% | $1.42/hr | 141GB HBM3e |
| L40S 48GB | Vultr | $0.56/hr | ▼ -5.1% | $0.43/hr | 48GB GDDR6 |
| A100 80GB SXM4 | TensorDock | $0.85/hr | ▼ -2.3% | $0.66/hr | 80GB HBM2e |
| A100 40GB PCIe | Lambda Labs | $1.12/hr | ▼ -2.6% | $0.87/hr | 40GB HBM2e |
| RTX 4090 24GB | Vast.ai (spot) | $0.44/hr | ▲ +13.9% | — | 24GB GDDR6X |
| RTX 3090 24GB | Vast.ai (spot) | $0.31/hr | ▲ +12.1% | — | 24GB GDDR6 |
| AMD MI300X 192GB | Thunder Compute | $1.85/hr | ➡ Flat | $1.42/hr | 192GB HBM3 |
Rates as of June 15, 2026. Spot rates are marketplace medians — individual host rates vary.
Top Price Movers — Week 25
- Vultr B200 ▼ -7.5% → $2.58/hr
Largest single-week B200 drop since launch. Vultr quietly expanded Blackwell capacity in Frankfurt and Chicago regions. The $4.26/hr CoreWeave floor from three weeks ago now looks like history. - Vast.ai RTX 4090 ▲ +13.9% → $0.44/hr (spot)
Third consecutive week of consumer spot price increases. The RTX 3090 surge (+131.7% in Week 23) that seemed anomalous has now spread upward through the consumer tier. Vast.ai marketplace hosts are pulling capacity — or demand for consumer-tier inference is climbing faster than supply. Watch this signal closely. - Lambda B300 ▼ -5.7% → $2.31/hr
B300 (Blackwell Ultra) is now the cheapest Blackwell SKU tracked. If this holds, it flips the expected pecking order — new architecture undercutting older Blackwell. Lambda is clearly using B300 aggressively to win inference workload migration. - Vast.ai RTX 3090 ▲ +12.1% → $0.31/hr (spot)
Consumer spot surge is broad-based. Two consecutive weeks of double-digit gains across the Vast.ai consumer tier signals a structural shift, not noise. The most likely explanation: small-model inference demand is routing to consumer GPUs as H100 moves up-market. - Crusoe GB200 NVL72 NEW → $6.95/hr
First meaningful competition to CoreWeave on GB200 NVL72 capacity. $6.95/hr vs CoreWeave's $7.40/hr is a 6.1% discount on the most expensive GPU tier in the index. For enterprises locked into NVL72 workloads, this is the first real alternative since CoreWeave's launch.
Vast.ai Surge — Consumer Spot Deep Dive
The Vast.ai marketplace has been the most volatile data source in the GridStackHub network for three consecutive weeks. Here's the pattern:
| GPU | Week 23 | Week 24 (est.) | Week 25 | 3-Week Change |
|---|---|---|---|---|
| RTX 3090 spot | +131.7% | +12.1% | +12.1% | 🔥 |
| RTX 4090 spot | baseline | +8% | +13.9% | 📈 |
| A10G spot | -51.8% | +6% | +5% | stabilizing |
| A40 spot | +54.4% | -8% | +4% | reverting |
The read: Consumer tier is seeing a demand surge, likely driven by: (1) Smaller open-source models (Llama 3.2, Phi-3) running efficiently on 24GB consumer cards; (2) Cost-sensitive teams hitting RTX 4090 as the cheapest viable inference GPU; (3) Some host capacity leaving the Vast.ai pool (hosts upgrading hardware). H100 spot inventory on Vast.ai remains stable at ~$1.87/hr, suggesting the surge is specific to the consumer tier, not broad GPU demand tightening.
Provider Comparison Table — Key Workloads
H100 80GB SXM — Best Rates by Provider
| Provider | On-Demand | Spot | Notes |
|---|---|---|---|
| CoreWeave | $2.14/hr | — | Lowest H100 on-demand tracked. NVLink clusters available. |
| Vast.ai | $1.87/hr | $1.35/hr | Marketplace — verify host reliability score |
| RunPod | $2.10/hr | $1.79/hr | Solid spot availability, good uptime SLA |
| Lambda Labs | $2.49/hr | — | No spot, but consistent guaranteed availability |
| TensorDock | $2.50/hr | $1.60/hr | Bare-metal, strong for batch training |
| AWS (P5) | $3.93/hr | — | Enterprise SLA; 4.9x cost of Vast.ai spot |
| GCP (A3) | $9.80/hr | — | Full VM pricing — avoid unless compliance-required |
A100 80GB — Best Rates by Provider
| Provider | On-Demand | Spot | Notes |
|---|---|---|---|
| TensorDock | $0.85/hr | $0.62/hr | Cheapest A100 on-demand tracked |
| Thunder Compute | $0.78/hr | — | Close second; good for long-running jobs |
| Vast.ai | ~$0.80/hr | $0.35/hr | Marketplace variance high — filter by reliability |
| CoreWeave | $1.19/hr | — | Premium, but guaranteed multi-GPU configs |
| RunPod | $0.99/hr | $0.74/hr | Community Cloud; good value |
| Lambda Labs | $1.12/hr | — | Consistent, no spot available |
B200 192GB — All Tracked Providers
| Provider | On-Demand | Spot | Notes |
|---|---|---|---|
| Vultr | $2.58/hr | — | Cheapest B200 tracked this week ▼ -7.5% |
| Lambda Labs | $2.91/hr | — | Strong availability; good SLA |
| CoreWeave | $3.10/hr | — | Down from $4.26/hr (Wk 22) — still premium |
| RunPod | $3.19/hr | — | New B200 tier added June 2026 |
| Spheron | $2.12/hr | $2.12/hr | Spot-equivalent; checkpoint-friendly |
30-Day Forecast — Week 25 Edition
- B200 192GB: Current $2.58/hr → Mid-July target $2.30–$2.42/hr. Vultr and Lambda capacity expansions in July; expect another 5–10% decline.
- B300 144GB: Current $2.31/hr → Mid-July target $2.10–$2.20/hr. Lambda competing hard on price; other providers likely to match within 2–3 weeks.
- H100 SXM 80GB: Current $2.14/hr → Mid-July target $1.95–$2.05/hr. Steady decline continuing as Blackwell siphons inference demand.
- RTX 4090 (Vast.ai spot): Current $0.44/hr → Mid-July target $0.38–$0.50/hr. High variance expected. If the consumer surge is demand-driven, prices could spike further. If it's supply compression, they normalize.
- A100 80GB: Current $0.85/hr → Mid-July target $0.82–$0.85/hr. Essentially flat. A100 has found its floor as a training workhorse for cost-sensitive teams.
5 Things to Watch This Week
- Blackwell hits $2.58/hr — B200 down 39% from its CoreWeave peak of $4.26/hr three weeks ago. This is the pricing regime shift that changes the inference calculus permanently.
- Consumer spot surge: week 3 — Vast.ai RTX 4090 and RTX 3090 are both up again. If this continues into Week 26, small-model inference demand has structurally outpaced consumer GPU supply on marketplace networks.
- B300 undercuts B200 on Lambda — At $2.31/hr vs B200's $2.58/hr floor, newer Blackwell Ultra is cheaper than older Blackwell. That's the supply chain catching up fast.
- Crusoe enters GB200 NVL72 at $6.95/hr — First CoreWeave alternative for the biggest Blackwell configs. Enterprises on NVL72 workloads should price-check now.
- H100 sub-$2.00/hr coming by August — At $2.14/hr and declining 2–3%/week, H100 hits the $2.00/hr barrier within 6 weeks. That's a psychological threshold that will drive another wave of migration from API services to self-hosted inference.
Get the GPU Cost Pulse Every Monday
Join 2,400+ engineers and AI teams who track GPU pricing with GridStackHub. Free weekly signal, no spam. Subscribe Free →
Source: GridStackHub GPU Pricing Database — 32 providers, 396 pricing records. Updated weekly. CC BY 4.0