Blackwell Supply Tightening: B200 Prices Climb +25% as Hyperscaler Demand Absorbs Capacity
B200 average rates hit $6.83/GPU-hour in June 2026 — a 25% climb from early May — as hyperscaler contracts absorb available Blackwell capacity faster than the supply chain can replenish it. The 30-day forecast points to $8.54/GPU-hour. For inference teams, the window to lock in favorable Blackwell rates is closing.
The Numbers: June 15, 2026
Across 21 tracked B200 providers, the distribution looks like this:
| Metric | B200 | B300 | B100 |
|---|---|---|---|
| Avg rate (per GPU/hr) | $6.83 | $5.63 | $4.09 |
| Min rate (per GPU/hr) | $3.87 | $4.48 | $3.99 |
| Max rate (per GPU/hr) | $8.60 | $6.94 | $4.19 |
| 30-day forecast | +25.0% | Stable | Stable |
| Cheapest provider | Vultr | Corvex | Lambda Labs |
| Cheapest rate | $3.87/hr | $4.48/hr | $3.99/hr |
Source: GridStackHub.ai live pricing index, 21 B200 providers tracked, June 15, 2026. Vultr B200 = single-GPU on-demand, US region.
Why Prices Are Rising Now
1. Hyperscaler absorption. Microsoft Azure, Google Cloud, and AWS all expanded their B200/GB200 NVL72 allocations through Q1 2026 earnings cycles. When hyperscalers lock in 8-GPU node contracts, they reduce per-GPU supply available to spot markets. Vultr and Lambda Labs now show multi-day lead times on new B200 provisioning in us-east-1.
2. CoWoS packaging ramp. NVIDIA's CoWoS-L packaging bottleneck has been easing steadily through H1 2026. More B200 silicon is reaching cloud providers — but absorbed by enterprise and government contracts before hitting commodity cloud tiers.
3. Inference demand outpacing training. The transition from frontier model training (which peaked in 2024–2025) to LLM inference at scale creates durable demand for Blackwell-class GPUs. B200's 3.5× inference throughput advantage over H100 means cost-per-token economics favor Blackwell even at a premium hourly rate.
The B300 Wildcard
B300 appears stable, averaging $5.63/GPU-hour across 15 providers. For multi-70B-model inference deployments, B300 at 83% of B200 cost with 90% of throughput is increasingly the better value play. Corvex is offering B300 at $4.48/GPU-hour on-demand as of June 9.
What Teams Should Do Now
- Run B200 cost analysis before July. At $6.83/GPU-hour on-demand, a 1,000-hour commitment on B200 costs ~$6,830/GPU. If the 30-day forecast materializes, that commitment in 30 days costs ~$8,540 — a $1,710 difference per 1,000 hours per GPU.
- Evaluate B300 as a cost hedge. Corvex B300 at $4.48/GPU-hour with 90% of B200 throughput is compelling for batch inference workloads.
- Check reserved instance pricing on CoreWeave. CoreWeave's spot B200 (8-GPU node) runs $4.48/GPU-hour — below the on-demand average. If you have predictable usage windows, preemption risk may be worth the 34% discount.
Forecast Confidence
High. The B200 forecast trajectory (+25% in 30 days) is grounded in persistent hyperscaler demand absorption and a supply chain that hasn't reached equilibrium. Risk: a new TSMC CoWoS allocation round or hyperscaler contract cancellation could reset the trajectory. Neither is our base case.
Want the full 30-day B200 forecast with provider-by-provider projections?
Pro Stack gives you 30-day forecasts, market floor prices, and savings opportunity alerts for all Blackwell models.
Start Pro Stack — $99/moSources & Attribution
All data points traced to named, verifiable sources. Proprietary data from GridStackHub.ai demand intelligence (no PII). Public data from government and industry research organizations.