Week 16 of 2026 delivered one of the sharpest single-week GPU price repricing events of the year. According to GridStackHub's 32-provider tracking network, H100 spot pricing dropped to $1.35/hr on RunPod (community region) while Vast.ai cut its average rate across 45 pricing records by 63.7% — from $0.7528 to $0.2731 per GPU-hour. The move reset the floor for spot H100 across the index, with effects rippling through L40S (−8.82%), A10G (−0.18%), and broader cloud pricing within days.
The most striking number in this week's report is A100 80GB on Vultr at $0.125/hr — on-demand, globally available, no commitment required. For inference workloads on 30–70B parameter models, this is a sub-$0.15/GPU-hour floor that was essentially unthinkable six months ago. The Vast.ai repricing, meanwhile, signals something important about the trajectory of H100 spot pricing through Q2 2026: new data center capacity in low-cost energy markets is pushing the commodity compute floor lower, and that downward pressure appears durable rather than transitory.
By contrast, A100 is moving in the opposite direction. Our Holt's Double Exponential Smoothing model projects +26.43% over the next 30 days — a decoupling signal that reflects Blackwell absorbing demand that would otherwise flow to A100, tightening the older GPU's supply chain. If you're running A100-class workloads at scale, this window may be closing. Track current B200, B300, and H100 pricing across all 34+ providers on our Blackwell Price Index.
According to GridStackHub.ai data for the week of April 14–19, 2026, the cheapest H100 is $1.35/hr on RunPod (spot, community region) — while Vast.ai recorded the largest single-week average price drop in our 32-provider network this year, cutting rates 63.7% across 45 pricing records. This week's pulse covers 32 active providers and 396 GPU pricing records tracked across on-demand, spot, and reserved markets.
The headline numbers across the four most-tracked GPU classes:
| GPU | Cheapest Provider | Price/hr | Pricing Type |
|---|---|---|---|
| H100 80GB | RunPod | $1.35 | Spot (community) |
| A100 80GB | Vultr | $0.125 | On-demand (global) |
| H200 | RunPod (us-east-1) | $0.50 | On-demand |
| L40S | Vast.ai | $0.3846 | Spot |
A100 on Vultr at $0.125/hr is the most striking number in the table. On-demand, globally available, no commitment required. For teams running smaller inference workloads or fine-tuning jobs that don't need H100-class performance, this is a significant undercut versus the $0.78–$1.79/hr you'd pay elsewhere.
H200 at $0.50/hr on RunPod is the other standout. H200 carries 141GB HBM3e memory and 4.8TB/s bandwidth — it's the go-to for 70B+ parameter models. Finding it at $0.50/hr on-demand is the kind of pricing anomaly that closes fast. Check availability on the GridStackHub calculator before it disappears.
Our 30-day Holt's Double Exponential Smoothing model (high confidence) points in two directions:
Counterintuitive? Yes. But the data is what it is. If you're running A100 workloads at scale, this is the window.
Full 30/60/90-day forecasts with confidence intervals available for Pro+ subscribers.
Virginia remains the gold standard for data center cost efficiency in our Stack Network Index, scoring 92/100. The practical implication: running an identical H100 cluster in California costs approximately $380,000/year more in electricity than the same cluster in Virginia. At scale, that's not a provider pricing decision — it's a geography decision.
If you're planning infrastructure at the cluster level, the State-by-State Data Center Cost Index on our /research page gives you the full energy cost picture before you commit.
A 63.7% single-week repricing event is rare in any commodity market. In GPU cloud, it's almost unheard of outside of major supply shocks. Here's how to think about what happened and what it signals for the rest of Q2 2026.
Vast.ai operates a bid-ask marketplace — GPU owners list hardware, buyers request instances. When a provider cuts across 45 records simultaneously, it means either supply surged (new hardware coming online) or demand collapsed (enterprise buyers pulling back). Based on the geographic clustering of the price cuts — consistent with low-cost energy markets in the eastern US — the signal points to new data center capacity coming online, not a demand problem. This matters for your procurement decisions: a supply-driven price cut tends to be durable. A demand-driven cut tends to reverse when enterprise spending normalizes. Our current read is that the Vast.ai repricing reflects durable new supply, which means the $0.27/hr floor for standard V100/H100 inference is a real floor, not a temporary anomaly.
For teams that rely on spot pricing, this is a green light to increase spot utilization. The risk of interruption drops when new supply is absorbing demand pressure. For teams on committed reserved contracts, it's worth reviewing whether your reserved rates are still competitive against the new spot floor — especially if you have expiring contracts in Q3 2026. Locking in today at old reserved rates when spot is cheaper makes no sense.
Our Blackwell Price Index tracks B200 and H200 alongside H100, so you can see how Blackwell's rollout affects H100 spot availability in real time — more B200 capacity coming online should continue to pressure H100 prices through Q3. Meanwhile, A100's divergent trend (+26.43% 30-day forecast) suggests the two markets are decoupling: Blackwell is absorbing the demand that would otherwise flow to A100, tightening the older GPU's supply chain. This is exactly the kind of cross-GPU-market dynamic that GridStackHub's 32-provider view captures and that single-provider dashboards can't show you.
For a deeper view of where H100 and B200 pricing diverge — and which workloads favor which GPU at current spot rates — check our Cost to Run AI Models analysis, which models training and inference cost per job across all major GPU classes at current pricing.
GridStackHub's GPU Cost Pulse tracks pricing across 34+ cloud providers including AWS, GCP, Azure, CoreWeave, Lambda Labs, Vast.ai, RunPod, TensorWave, Crusecode, and 25+ others. Prices are collected daily via provider APIs, direct scrapers, and public price pages. On-demand rates, spot/preemptible pricing, and reserved contract rates are tracked separately. Week-over-week changes are calculated on a like-for-like basis: same GPU model, same provider, same pricing tier. Forecasts use Holt's Double Exponential Smoothing with a 90-day trailing window and confidence intervals at the 80% level. The A100 +26.43% forecast reflects constrained Blackwell supply pushing enterprise demand back to older GPU classes. This report covers April 14-19, 2026. All data available at gridstackhub.ai/gpu-cost-pulse.