NVIDIA Blackwell GPU Pricing Index — Week 20, 2026: B200, H100, and the Full GPU Cloud Cost Map

According to GridStackHub.ai data for the week of May 19, 2026, the cheapest B200 GPU rental is $4.00/hr on Vast.ai (spot, limited availability), while the cheapest reliable on-demand rate is $1.38/hr for an H100 on Thunder Compute. That 2.9× price gap — between the current-generation Blackwell flagship and the previous-generation workhorse — defines the central infrastructure decision for AI teams in 2026. This week's pulse covers the full blackwell price index, all major GPU classes, the top five price movers, and a direct answer to the fastest-rising query in our network: B200 vs H100 cost — which GPU should you rent?

This report covers B200, H100, H200, A100, and L40S pricing across spot, on-demand, and reserved markets, sourced from GridStackHub's 58-provider tracking network. Data is current as of the week of May 19, 2026.

Blackwell GPU Pricing Deep Dive: What Does a B200 Actually Cost to Rent in 2026?

The cheapest B200 GPU rental in the market right now is $4.00/hr on Vast.ai spot. That price is real — and it is also unreliable. Vast.ai B200 spot availability is sporadic; the rate fluctuates to $8.00/hr when supply tightens. RunPod's community marketplace lists B200 spot between $6.00 and $9.00/hr, also subject to availability gaps.

For teams that need guaranteed access, enterprise contracts from CoreWeave and Lambda Labs price B200 compute at $10.00–$14.00/hr — reflecting both the hardware scarcity premium and the demand backlog that extends through Q3 2026. Most providers are waitlist or reserved-only. There is no B200 equivalent of RunPod's H100 spot floor. The open market simply does not have enough supply.

The global B200 installed base is approximately 2,000 instances. Compare that to 25,000+ H100 instances available on the cloud market. This 12.5× supply gap is the entire explanation for B200 pricing: it is not a premium for performance alone, it is scarcity pricing.

The supply picture is shifting, but slowly. Mistral AI has brought online 18,000 Blackwell GPUs at a $1.4B Sweden datacenter — the largest single Blackwell deployment announced to date. DeepInfra now runs Blackwell plus next-generation Vera Rubin GPUs across 8 US datacenters. Fireworks AI confirmed a B200 backend at a $4B valuation. CoreWeave, post-IPO, is actively expanding B200 supply. These additions will matter, but they are feeding enterprise contracts and inference clouds first. Spot market availability for B200 remains thin through summer 2026.

B300 and GB200 (the NVLink-connected superchip pairing two B200 dies with a Grace CPU) are not meaningfully available on the open market as of this writing. GB200 deployments are exclusively enterprise contract or internal. Treat any GB200 or B300 spot listing as a pricing anomaly requiring direct verification.

Full GPU Cloud Pricing Index — Week 20, May 19, 2026

The table below covers spot floor, cheapest on-demand, reserved market average, and week-over-week directional change for each major GPU class. Prices reflect GridStackHub tracking network data, updated week of May 19, 2026.

H100 1-year reserved market average: $2.35/hr — up from $1.70/hr in October 2025, a +38% year-over-year increase. AWS H100 on-demand top of range: $41/hr. H100 spot floor (RunPod community): $1.20/hr. That is a 34× spread for the same GPU class depending on provider and commitment tier.

Top 5 GPU Price Movers: What Changed and Why

1. Vast.ai: –63.7% average across 45 GPU listings (Week 16 baseline)
   –63.7% average repricing event
   Average rate dropped from $0.7528 to $0.2731 across 45 pricing records — the largest single-week repricing event in GridStackHub's 58-provider network this year, driven by a significant new supply addition that reset the spot marketplace floor.
2. L40S on Vast.ai: $0.4218 → $0.3846/hr (–8.82%)
   –8.82%
   The L40S continues to compress toward the $0.38/hr floor as Vast.ai's new supply wave reaches the mid-tier inference GPU segment, making it the most cost-efficient option for single-GPU inference on 30B-class models.
3. H100 1-Year Reserved Market: $1.70 → $2.35/hr (+38% YoY since Oct 2025)
   +38% YoY
   Enterprise demand for guaranteed H100 capacity has pushed reserved pricing up 38% year-over-year, even as spot prices remain soft — a divergence that reflects the difference between enterprise procurement and elastic cloud buying.
4. H200 AWS Capacity Blocks: +15% (January 2026)
   +15% AWS
   AWS raised H200 capacity block prices 15% in January 2026, reflecting tighter HBM3e supply and strong demand from teams running 70B+ models that require H200's 141GB memory headroom.
5. A100 30-Day Forecast: +26.43% (supply tightening)
   +26.43% forecast
   GridStackHub's Holt's Double Exponential Smoothing model flags A100 as the GPU most likely to reprice upward in the next 30 days, as aging inventory exits the market and enterprise demand for cost-efficient training persists.

B200 vs H100: Which GPU Should You Rent in 2026?

For most workloads, rent an H100 — at $1.38/hr on Thunder Compute or $1.20–$1.35/hr spot on RunPod. The B200 is 2–4× faster on certain inference benchmarks, but at $4.00–$14.00/hr it costs 3–10× more per hour with far less availability. The math only works for B200 if your workload is latency-critical, requires Blackwell-specific capabilities (FP4 precision, NVLink bandwidth at scale), or your team is running continuous large-batch inference where per-token cost justifies the premium.

If you are fine-tuning a 7B–70B model, running batch inference, or doing any exploratory training, the H100 at current spot prices delivers better cost-per-job than a B200 at $4.00/hr spot — and far better than a B200 at $10.00–$14.00/hr contract. Rent H100 now. Reassess B200 when spot availability stabilizes, expected Q4 2026 at the earliest.

B200 vs H100 cost verdict: H100 wins on cost in 2026. B200 wins on performance. Pick based on your bottleneck.

30-Day GPU Price Forecast: B200 Stays Tight, H100 Softens

Forecasts generated using Holt's Double Exponential Smoothing on GridStackHub's 58-provider pricing dataset. High confidence rating.

Track GPU Prices in Real Time

• GPU Cost Calculator — Compare B200, H100, H200, A100, and L40S across 58 providers instantly. See your true monthly cost before you commit.
• Blackwell Price Index — Live B200, B300, and GB200 pricing updated daily. Real-time Blackwell market data.
• Subscribe to The Grid Stack — Weekly GPU Cost Pulse every Tuesday, free. Pricing data, big movers, and buy/wait signals for the GPU cloud market.

Next issue: Week 21, May 27, 2026. Subscribe free →