Blackwell Deep Dive: B200 / B300 / GB200
Blackwell GPU Index — Week 24
Blackwell-first
Blackwell pricing continues its slow descent as cloud providers bring more B200 and GB200 inventory online.
| GPU | Provider | On-Demand | WoW Change | Memory |
|---|---|---|---|---|
| B200 192GB | Vultr | $2.79/hr | -4% | 192GB HBM3e |
| B200 192GB | Lambda Labs | $3.05/hr | -2% | 192GB HBM3e |
| GB200 NVL72 | CoreWeave | $7.40/hr | Flat | 1.44TB HBM3e |
| B300 144GB | Lambda Labs | $2.45/hr | NEW | 144GB HBM3e |
Supply context: NVIDIA's B200 supply chain constraints are easing — Vultr and Lambda both expanded Blackwell availability in Q2 2026. GB200 NVL72 remain scarce and expensive, primarily on CoreWeave. B300 (Blackwell Ultra) is the new entrant this week with Lambda Labs, priced between B200 and GB200.
Full GPU Index
| GPU | Cheapest Provider | On-Demand | 1-yr Commit | Memory |
|---|---|---|---|---|
| GB200 NVL72 | CoreWeave | $7.40/hr | — | 1.44TB HBM3e |
| B300 144GB NEW | Lambda Labs | $2.45/hr | $1.89 | 144GB HBM3e |
| B200 192GB ↓-4% | Vultr | $2.79/hr | $2.15 | 192GB HBM3e |
| H100 SXM 80GB | CoreWeave | $2.19/hr | $1.79 | 80GB HBM3e |
| H200 80GB | Lambda Labs | $1.89/hr | $1.49 | 80GB HBM3e |
| A100 80GB SXM4 | TensorDock | $0.87/hr | $0.68 | 80GB HBM2e |
| A100 40GB PCIe | Lambda Labs | $1.15/hr | $0.89 | 40GB HBM2e |
| L40S 48GB | Vultr | $0.59/hr | $0.45 | 48GB GDDR6 |
*Rates as of June 9, 2026. Commit rates reflect 1-year prepaid.
Top 5 Price Changes This Week
-
1
B200 on Vultr — Down 4% to $2.79/hr. Blackwell supply expansion is bringing rates down for the first time since launch.
-
2
A100 80GB on TensorDock — Up 3% to $0.87/hr. Demand for older-architecture training workloads keeping A100 pricing firm.
-
3
H100 on CoreWeave — Down 2% to $2.19/hr. Increasing H100 cluster capacity from new Texas region.
-
4
L40S on Vultr — Down 5% to $0.59/hr. NVIDIA L40S supply normalizing after Q1 shortages.
-
5
GB200 on CoreWeave — Flat at $7.40/hr. No change — NVL72 remain supply-constrained.
Is B200 cheaper than H100 for inference right now?
Yes — for raw inference workloads, B200 is now cost-competitive with H100 when you factor in performance per dollar.
B200 delivers approximately 3x the tokens per second vs H100 for LLM inference (per GridStackHub benchmarks). At $2.79/hr vs H100's $2.19/hr, you're paying ~27% more for 3x the throughput — making B200 the more cost-efficient choice for production inference at scale. For training, H100 still wins on $/FLOP.
Rule of thumb: Inference = B200. Training = H100 (for now).
30-Day Price Forecast
B200 192GB
➘ Soft decline
Current
$2.79
→
Mid-July target
$2.58–$2.64
More supply online (June–July); expect -5–8% by mid-July
Track live →
H100 SXM 80GB
➘ Mild decline
Current
$2.19
→
Mid-July target
$2.05–$2.10
H100 inventory building; slight downward pressure as Blackwell takes share
Compare providers →
GB200 NVL72
➡ Stable
Current
$7.40
→
Mid-July target
$7.40
Supply stays tight; no significant movement expected in 30 days
Compare →
A100 80GB
➡ Flat
Current
$0.87
→
Mid-July target
$0.87
Market finding equilibrium; limited movement expected
Token cost →
5 Things to Watch This Week
- Blackwell cost inflection: B200 at $2.79/hr is the first meaningful week-over-week decline — Blackwell pricing is entering the commodity phase faster than expected.
- B300 is the new mid-tier Blackwell: At $2.45/hr, B300 sits between B200 and GB200, offering Blackwell architecture at near-H100 price points. Watch this tier closely.
- Inference vs training split is real: The market is clearly bifurcating. H100 for training, B200 for inference — and that split is now reflected in pricing trends.
- L40S as the budget workhorse: At $0.59/hr, L40S is becoming the default for budget inference and smaller training jobs. Supply normalization is driving prices down.
- GB200 NVL72 still untouchable: $7.40/hr and flat — no relief expected in 30 days. Enterprise AI at scale only.
Calculate your workload cost → Compare all 34 providers by GPU type, hours, and commitment level.
GPU Cost Calculator →
← Week 23 GPU Cost Pulse
Week 24 of 2026