Capacity Forecasting

Why this drives the biggest decisions

At a multi-billion-dollar neocloud, capacity decisions are the largest single decisions the company makes:

• Buying 10,000 H200s at $25k each is a $250M decision.
• Building a new datacenter site is a $500M+ decision.
• Locking a power-purchase agreement at a site is a multi-year financial commitment.
• Reserving NVIDIA allocation requires forecasting demand 12-24 months out.

Every one of these decisions rests on a forecast. The DS function that produces credible forecasts is the function that justifies cap-ex spending.

Demand forecasting

Inputs to demand forecasting:

• Existing customer growth. Trajectory of current accounts' usage.
• Sales pipeline. Expected new commitments with probability weights.
• Renewal probabilities. Risk-adjusted continuation of existing contracts.
• Market signals. Industry-level AI demand indicators.
• Seasonality. Conference deadlines, holidays, fiscal cycles.
• SKU mix shifts. Customer migration from H100 to Blackwell, etc.

Output: expected GPU demand by SKU by region by month over a 12-24 month horizon, with confidence intervals.

Supply planning

Supply side considers:

• NVIDIA allocation timing for new generations.
• Datacenter capacity coming online.
• Power and cooling capacity at each site.
• Lead times for new equipment.
• Provider partner capacity (for hybrid models).
• Existing fleet retirement schedule.

The supply-side modeling is more deterministic than demand but with its own uncertainties (delivery delays, datacenter buildout pace, etc.).

Buy-vs-build decisions

Faced with forecast demand exceeding existing supply, the company has options:

• Buy more GPUs into existing capacity. Quickest but constrained by existing power and cooling.
• Build new datacenter capacity. Slowest but most strategic.
• Partner / lease capacity. Faster but less margin.
• Decline incremental customers. Walk away from demand the company can't profitably serve.

The DS-supported decision matrix takes forecasts, cost data, and strategic priorities and produces recommended actions. Leadership decides; DS quantifies the tradeoffs.

Uncertainty quantification

Forecasts that don't quantify uncertainty mislead. Methods:

• Bayesian forecasting with explicit posterior distributions.
• Scenario modeling (bull / base / bear cases).
• Monte Carlo simulation across assumption ranges.
• Confidence intervals on point forecasts.
• Stress testing against adverse scenarios.

The DS communication to leadership emphasizes the uncertainty as much as the central estimate. Decisions are made under uncertainty; pretending otherwise is harmful.

Methods used

The forecasting toolkit:

• Time-series methods (ARIMA, Prophet, neural-forecasting) for established trend extrapolation.
• Hierarchical Bayesian models for borrowing strength across segments.
• Pipeline-stage models for sales-pipeline forecasting.
• Survival analysis for renewal probability.
• Causal models when interventions (pricing changes, marketing pushes) interact with forecasts.
• Ensemble methods to combine the above.

✓ Check yourself

Answer from memory before opening each one.

1. You plan GPU purchases from a demand forecast. If the forecast is 20% too high, and if it is 20% too low, what happens in each case, and what does that suggest about which number to plan on?

   Show answer

   Too high leaves capital idle: on 10,000 H200s at $25,000 that is about $50M of unused hardware. Too low means turning away demand or leasing at lower margin. The costs differ, so plan on a chosen quantile of the distribution, not the mean. See why this drives the biggest decisions and uncertainty quantification.

2. Three pipeline deals ask for 1,000, 500 and 200 GPUs, with win probabilities 0.6, 0.3 and 0.8. What is the expected demand, and why not plan for 1,700?

   Show answer

   Expected demand is 600 + 150 + 160 = 910 GPUs. Planning for 1,700 builds for every deal closing, which is unlikely and leaves idle capacity. Also report the range, because the outcome is one of several discrete combinations. See demand forecasting.

3. Buying more GPUs into existing sites is the quickest option. What limits it?

   Show answer

   Power and cooling at those sites, plus delivery lead times. When they bind, you must build, lease or partner, which are slower or lower margin. See supply planning and buy-vs-build decisions.

4. Why use a hierarchical Bayesian model for demand by SKU and region rather than a separate ARIMA for each cell?

   Show answer

   Many cells have little history, so separate models are noisy. A hierarchical model shrinks sparse cells toward the group and borrows strength across segments. See methods used.

5. Leadership asks for one number for next year's demand. What do you give them?

   Show answer

   A central estimate with an interval and the bull, base and bear cases behind it, plus the stress test that hurts most. A single number hides how wrong it can be, and the decision is made under that uncertainty. See uncertainty quantification.

6. An ARIMA or Prophet model fitted to two years of history says demand will grow steadily. Why might it mislead for a neocloud?

   Show answer

   Trend extrapolation assumes the structure that produced the history continues. Demand shifts when customers move from H100 to Blackwell, when a large deal lands or when pricing changes, and none of that is in the trend. Combine it with pipeline, renewal and causal models. See methods used.

Takeaway

• Capacity decisions are the biggest the company makes, a few hundred million dollars at a time, and each rests on a forecast.
• Demand forecasts combine customer growth, pipeline weighted by probability, renewal risk, seasonality and SKU mix shifts.
• Supply is more deterministic than demand but is gated by allocation timing, power, cooling and delivery delays.
• Faced with a gap, the options are buy into existing capacity, build, partner or decline demand, each with different speed and margin.
• Report forecasts with intervals, scenarios and stress tests, because asymmetric costs make the mean the wrong planning number.
• Use hierarchical Bayesian models for sparse segments, pipeline and survival models for deals and renewals, and trend methods only where history is stable.

Forecasts decide how much supply exists; ranking decides who gets it. → Marketplace Ranking & Matching

Further reading

• Prediction intervals — Hyndman and Athanasopoulos, Forecasting: Principles and Practice, 3rd ed.. How to compute and interpret intervals around a point forecast.
• Time series analysis — statsmodels documentation. ARIMA, exponential smoothing and state-space models in one reference.
• Prophet — Meta, repository and documentation. The additive trend and seasonality forecaster, with its model description and install notes.
• Forecasting at Scale — Taylor and Letham, 2017. The paper behind Prophet and its argument for analyst-tunable forecasts.
• PyMC learn — PyMC documentation. Hierarchical Bayesian models for forecasts across sparse segments.