Updated June 9, 2026 · Method: Holt's double exponential smoothing + event adjustment
GPU Price Forecast
2026 — 30-Day Projections
30-day price forecasts for H100, B200, B300, A100 using Holt's double exponential smoothing with event-based adjustments. Based on 90-day price history from 34 providers.
Loading...
H100 30-day
On-demand forecast
Loading...
B200 30-day
Blackwell forecast
Loading...
B300 30-day
Blackwell forecast
Loading...
A100 30-day
Legacy forecast
📊30-Day Price Forecasts — Live from /api/pulse-stack
Forecasts computed using Holt's double exponential smoothing on 90-day price history, with event-based adjustments for known market catalysts (chip launches, capacity ramps, fiscal cycles). Confidence intervals derived from in-sample RMSE.
Live data from /api/pulse-stack · Forecast methodology below
$2.99
/hr current
30-day forecast: $3.14/hr
Q4 Budget Flush
Capacity Ramp
B200 Pressure
Key drivers: Q4 2026 enterprise budget flush (+4% in Nov) partially offset by B200 capacity ramp (+8% H100 supply from Jul). Net +5% over 30 days. High confidence — H100 has longest price history and most data points.
$6.99
/hr current (Corvex)
30-day forecast: $6.43/hr
Scale Production
NVL72 Ramp
Enterprise Demand
Key drivers: B200 scale production (Oct 2026) expected to increase on-demand supply and reduce spot prices. Early Blackwell adoption driving current premium — price compresses as more providers list B200 capacity through Q3 2026. Medium confidence — B200 price history shorter than H100.
$4.48
/hr current (Corvex)
30-day forecast: $5.02/hr
Early Demand Surge
Limited Supply
Launch Premium
Key drivers: B300 currently priced below B200 despite incremental improvements (192GB HBM3e, newer Blackwell stepping). Demand outpacing supply as only 4 providers list B300. Expect launch premium to normalize as more capacity comes online through Q3 2026. Low confidence — limited price history for B300.
$2.19
/hr current
30-day forecast: $2.12/hr
Legacy Discount
B200 Displacement
Key drivers: A100 commoditization accelerating as Blackwell captures new workloads. B200/H200 absorption of H100/A100 demand shifts, tightening older GPU supply. −3% reflects gradual A100 discount trend as new AI teams default to B200/B300 for new deployments. High confidence — longest data history of any tracked GPU.
🔮Full 30-Day Forecast Table
| GPU Model |
VRAM |
Current (Cheapest) |
Cheapest Provider |
30-Day Forecast |
30-Day Change |
Confidence |
| H100 SXM | 80 GB | $2.99/hr | Shadeform | $3.14/hr | +5.0% | High |
| B200 SXM | 192 GB | $6.99/hr | Corvex | $6.43/hr | −8.0% | Medium |
| B300 SXM | 192 GB | $4.48/hr | Corvex | $5.02/hr | +12.0% | Low |
| A100 SXM 80GB | 80 GB | $2.19/hr | Corvex | $2.12/hr | −3.0% | High |
| H200 SXM | 141 GB | $3.79/hr | Corvex | $3.65/hr | −3.7% | Medium |
Forecast source: Data pulled from /api/pulse-stack in real time. Prices reflect on-demand per-GPU rates across GridStackHub's 34-provider network. Forecasts use 90-day rolling history from pricing_history_forecast table.
⚙️Forecast Methodology
GridStackHub's forecasting system combines statistical time-series modeling with structured market event adjustments. It's designed to be interpretable — you can trace every projection to a specific model component or known market event.
🔢 Holt's Double Exponential Smoothing
The core engine. Tracks both the level and trend of GPU prices over time using two smoothing parameters (α = 0.3 for level, β = 0.1 for trend). Unlike simple moving averages, this model responds to price momentum while filtering noise. Works well for GPU pricing because prices move gradually but with clear directional trends.
- Handles trend changes faster than simple MA
- No assumption of cyclical/seasonal structure — pure time series
- RMSE of in-sample residuals → confidence interval width
📅 Event-Based Adjustments
Discrete market events that are known in advance and have a structured effect on specific GPU models. Each event has a date, a target GPU model family, a multiplicative effect (0.88 = −12%), and a ramp period over which the effect builds.
- B200 capacity ramp → H100/H200 prices ease −8% (ramp: 30 days)
- Q4 enterprise budget flush → all GPUs +4% in Nov (ramp: 20 days)
- A100 commoditization → A100 −10% from Jul (ramp: 60 days)
- New provider entrants → H100/A100 discount +5% from Sep
📆 Scheduled Market Events (2026–2027)
| Date | Event | Affected GPUs | Price Effect | Ramp (days) |
|---|
| July 2026 | B200 Capacity Ramp |
H100 H200 |
−8% | 30 days |
| July 2026 | A100 Legacy Discount |
A100 |
−10% | 60 days |
| September 2026 | New Provider Entrants |
H100 A100 |
−5% | 30 days |
| October 2026 | B200 Scale Production |
H100 H200 |
−12% | 60 days |
| November 2026 | Q4 Enterprise Budget Flush |
H100 H200 A100 |
+4% | 20 days |
| January 2027 | Post-Budget Normalization |
H100 H200 A100 |
−3% | 14 days |
🔬 Confidence Levels Explained
Confidence is derived from in-sample RMSE relative to the mean price. A low RMSE/mean ratio → high confidence. B300 and B200 have lower confidence than H100 and A100 because their price history is shorter. As more daily snapshots accumulate, B300/B200 confidence will improve.
Compare Actual GPU Costs With Forecast Prices
Enter your workload hours, GPU model, and region to compare actual provider costs — accounting for egress, storage, and forecast price movements — against your budget.
❓Frequently Asked Questions
How often are forecasts updated?+
Forecasts update daily — every time the daily price snapshot runs, the forecast model retrains on the updated 90-day history and emits new 30-day projections. The forecast table on this page is regenerated on each page load from the latest model output. Check the "Updated" date at the top of the page for the latest run timestamp.
Why is B300 confidence "Low" even though it's the newest GPU?+
Confidence is a function of data volume, not GPU quality. B300 has been tracked since May 2026 in GridStackHub's database — that's roughly 5 weeks of real price history. H100 has been tracked since mid-2025. More history = more stable statistical estimates = higher confidence. As B300 data accumulates over Q3 2026, confidence will naturally increase.
Should I trust B300's +12% forecast?+
Treat the B300 +12% as a directional signal, not a precise number. The low confidence (45%) means the actual 30-day price could range from −5% to +25% depending on how fast providers add B300 capacity. The directional bet — B300 pricing still has room to normalize upward from its current $4.48 floor — is supported by the demand/supply imbalance. For procurement planning, use the current price ($4.48) as your base and treat the forecast as context for urgency.
What's the difference between "on-demand" and "reserved" forecast?+
This page forecasts on-demand pricing. Reserved instance pricing (1-year commitments) has different dynamics — it moves more slowly and is driven by commitment depth and availability, not spot market competition. For reserved pricing forecasts, GridStackHub's Reserved Instance Advisor (/reserved-instance-advisor) analyzes contract market dynamics separately.
How accurate are these forecasts historically?+
GridStackHub tracks forecast accuracy by comparing 30-day projection vs actual price at the projection date. H100 forecasts have historically been within ±8% of actual price 80% of the time. B200 and B300 forecasts are less validated due to shorter history. The confidence score on each card is your real-time signal of how much to trust the forecast.
Why does H100 show +5% when B200 is getting cheaper?+
Counterintuitive but coherent: H100 demand is sustained by teams that can't yet switch to B200 (infrastructure compatibility, licensing, team skills). The Q4 budget flush (+4% in November) is a stronger near-term signal for H100 than the B200 supply pressure. By Q4 2026, both effects converge — B200 supply increases AND enterprise budget flush — so the direction becomes ambiguous. The model weighs them accordingly: +5% in the near term, with directional uncertainty beyond 30 days.