1. GPU Demand: A Black‑Swan in Cloud Infrastructure

AI expansion has placed GPUs at the center of infrastructure bottlenecks. By late 2023, NVIDIA had sold out its 2024 Hopper (H100) GPU production, with lead times up to 52 weeks [3]. Amazon’s internal Project Greenland[4], which centralized allocation and redirected supply to its most critical AI projects, underscored the scarcity. By contrast, CPU demand has been stable and cyclical. IDC projects global PC device shipments to grow at 3.1% CAGRthrough 2027[5], reflecting predictable refresh cycles and macroeconomic shifts.

GPU demand, by comparison, is volatile and prone to shocks. A single product launch can trigger a step‑function demand: GPT‑4’s release, for instance, required tens of thousands of GPUs overnight. Meta’s 2023–24 announcement to scale to 350,000 H100s [6] illustrates a regime shift unseen in CPU markets. GPU demand isn’t just an AI story. For much of the past decade, the crypto industry has been one of the largest consumers of GPUs, particularly for proof‑of‑work mining. At its peak before Ethereum’s transition to proof‑of‑stake in 2022, analysts estimate that Ethereum miners spent around $15 billion on GPUs in just 18 months, consuming roughly 14 million cards—about 10 % of all discrete GPUs shipped globally during that period [12]. This demand created massive hardware shortages for gamers and enterprises alike. In proof‑of‑work systems, miners compete to solve cryptographic puzzles using GPUs, consuming enormous computing power and electricity. Proof‑of‑stake, by contrast, replaces this with a system where validators put up collateral (their “stake”) to confirm transactions, dramatically reducing GPU requirements. Even after Ethereum’s merge, other proof‑of‑work chains (like Ethereum Classic, Ravencoin, and Ergo) continue to rely on GPUs, while blockchain analytics workloads increasingly tap into GPU clusters. This dual demand—from AI on one side and crypto on the other—tightens supply and makes the market far more shock‑prone.

This asymmetry explains why methods that work for CPUs—such as Autoregressive Integrated Moving Average (ARIMA), Holt–Winters, Prophet—fail for GPUs. CPUs can be modeled around seasonal cycles, but GPUs require probabilistic, nonlinear approaches that capture tail risks. The AI boom has strained GPU supply chains and revealed the limits of traditional forecasting in an era increasingly defined by what Nassim Nicholas Taleb called black‑swandynamics—rare events with severe impact, obvious only in hindsight [7].

2. Empirical Analysis

Classical forecasting methods—ARIMA, Holt‑Winters, and Seasonal Naïve—were evaluated on demand traces from two cloud providers covering 2023, 2024, and 2025. Though sparse, the datasets captured the post‑2023 GPU surge. For 2023 and 2025, demand (requests) was analyzed; for 2024, actual usage data was used to assess both dimensions.

Among error metrics, WAPE proved most reliable, absorbing temporal gaps and zero‑demand days. By contrast, MAPE was distorted by missing entries; RMSE lacked intuitive comparability.

2.1 Findings

Across all three datasets, CPU forecasts showed lower error and stable convergence, reflecting smoother, diversified, cyclical demand. GPU forecasts were far less reliable—with higher errors, poor temporal consistency, and model breakdowns. Even when CPU seasonality was captured, GPU bursts were consistently underestimated.

2.2 Advanced Approaches

To address the limitations of traditional models, probabilistic (quantile regression) and machine learning approaches (light LSTM) were analyzed. While data sparsity remained a constraint, both methods performed better at capturing nonlinear dependencies, with detailed results presented in Section 3.

The contrast highlights a structural divide: CPU demand is bounded and cyclical, while GPU demand is discontinuous, burst‑driven, and black‑swan‑prone. Classical methods remain serviceable for CPUs but inadequate for GPUs, reinforcing the need for probabilistic and ML‑based approaches that model volatility and tail risks.

3. Beyond Classical Forecasting

3.1 Limitations of Traditional Models

Classical temporal models like Holt‑Winters and ARIMA work for CPUs but collapse under GPU volatility. Their reliance on cycles and seasonality makes them ill‑suited for burst‑driven shocks. What is reliable for CPUs becomes misleading for GPUs—underscoring the need to move beyond point forecasts.

3.2 Probabilistic Forecasting

Classical models assume stability and Gaussian error distributions, causing narrow uncertainty bands. With GPUs, bursts create heavy‑tailed errors, underestimating both peaks and tail risks. Workload concentration magnifies the issue: CPUs are diversified across services, but GPUs are typically tied to one or two AI teams. Thus, even small errors carry a disproportionate cost—a 10 % miss on AWS C5 CPUs may go unnoticed, but 5 % on P5EN GPUs can mean millions.

Probabilistic methods—quantile regression and Monte Carlo simulation—offer greater elasticity. Planners can choose among probabilistic profiles—P30, P50, P95—aligned with risk tolerance. In the 2025 dataset, a P50 GPU forecast via quantile regression yielded a WAPE of 72.3 %, while P30 reduced it to 56.7 %, compared with 100.8 % under Holt‑Winters. This flexibility is invaluable in sparse‑data settings, letting planners bias forecasts toward conservative or aggressive quantiles until actuals stabilize around a regular P50.

3.3 Machine Learning Models

Probabilistic models help but rely on fixed parametric assumptions—normal distributions, or fixed quantile levels. Machine learning methods—LSTMs, BSTS, XGBoost—adaptively capture nonlinear patterns and structural breaks. In 2025 tests, basic LSTM, with 2,500 Monte Carlo simulations, achieved P50 WAPE of 28.3 % versus quantile regression’s 72.3 % and Holt‑Winters’s 100.8 %. ML models thus provide adaptive bands that reflect GPU volatility and resilience to black‑swan shocks.

4. Conclusion & Business Impact

The empirical work shows a clear temporal asymmetry: CPUs exhibit bounded, cyclical demand where classical models are serviceable; GPUs are burst‑driven, heavy‑tailed, and frequently defy point forecasting assumptions. The business implications are immediate:

1. Budget exposure scales nonlinearly on GPUs. Even small percentage errors drive large dollars because unit costs are high and burst windows are concentrated. AWS publishes dedicated Capacity Blocks pricing for H100/H200 instances (p5/p5e), with effective rates per instance and per‑accelerator, underscoring each GPU‑hour’s cost. On‑demand H100 usage is $7–$8 per GPU‑hour. Even at AWS’s capacity‑block rates (≈$3.9/GPU‑hour), a modest 5 % miss on 300 GPUs for a month implies about $40k–$90k in unplanned spend—before considering spillover effects on delayed launches. [8, 9]
2. Supply is constrained and slow to bend. Reported 36–52 week lead times on high‑end accelerators and pre‑sold production cycles mean you cannot “buy your way out” of a bad forecast next quarter. Errors cause missed launches and force costly spot or third‑party workarounds. [3]
3. Enterprises are re‑architecting governance around GPUs. Amazon’s internal “Project Greenland” centralized allocation and imposed ROI‑gated access to scarce GPUs—an explicit operating response to forecasting/supply tension. Expect similar “capital rationing” playbooks. [4, 10]
4. Capex stakes are historic. Hyperscalers commit tens of billions quarterly for AI capacity (e.g., Microsoft guiding to $30 B in a single quarter), so forecast discipline directly affects ROIC and margin trajectory. Mis‑estimating GPU demand pushes cash burn forward and erodes unit economics. [11]
5. GPU‑as‑a‑Service: The forecasting challenge also creates opportunities. GPU‑as‑a‑Service providers like CoreWeave benefit from enterprises seeking flexible capacity to hedge risk. Infrastructure software companies offering GPU utilization optimization (like Run:AI or Determined AI) become strategic as customers seek efficiency gains from existing allocations.

5. Next Steps

The findings highlight several imperatives for enterprises navigating GPU volatility.

1. Plan with bands, not points. Forecasting should move from single‑number outputs to P50/P90 intervals, with P90 tied to SLAs and P50 serving as base capacity. This aligns procurement with risk tolerance and acknowledges the inadequacy of narrow bands from classical forecasting techniques.
2. Institutionalize allocation governance. Greenland‑style frameworks—such as ROI‑based intake, expiration on idle allocations, and central rebalancing of stranded GPUs—are essential under scarcity.
3. Align contracting with risk policy. Blend committed use (like AWS Capacity Blocks) for P50 needs with on‑demand or spot for bursts, and partner‑cloud or colocation options as a contingency.
4. Treat forecasting as an operations product. Move beyond static reports to a risk engine publishing quantile curves, error diagnostics, and purchase recommendations.
5. Monitor competitive GPU allocation efficiency. For investors, companies that master GPU forecasting will gain sustainable competitive advantages. Track GPU utilization, AI feature time‑to‑market, and capex efficiency. Firms that hit AI delivery timelines while maintaining predictable infrastructure spend are solving the forecasting problem—making them attractive long‑term positions in an increasingly GPU‑constrained market.

Looking forward, enterprises may even need to rethink competition itself. Just as airlines formed alliances to stabilize capacity and energy firms built joint LNG terminals, there is a strong case for GPU procurement consortia. Pooling demand would smooth volatility, create more predictable order books, and reduce stranded capacity. One firm’s idle allocation could be another’s lifeline. Challenges around antitrust, data isolation, and governance remain, but the economic logic is compelling—and may become a strategic necessity.

Finally, companies cannot delay investing in human and system capital for GPU forecasting. Even small improvements yield outsized financial impact. With extreme unit costs, year‑long lead times, and bursty demand, GPU forecasting is not a back‑office exercise—it is a board‑level risk discipline.

References

1) Jadon et al., 2021 – Challenges And Approaches to Time‑Series Forecasting In Data‑Center Telemetry: A Survey (https://arxiv.org/pdf/2101.04224)

2) Nashold & Krishnan, 2020 – Using LSTM and SARIMA Models to Forecast Cluster CPU Usage (https://arxiv.org/pdf/2007.08092)

3) Tom’s hardware – Nvidia sold half a million H100 AI GPUs in Q3 thanks to Meta, Facebook — lead times stretch up to 52 weeks: Report (https://www.tomshardware.com/tech-industry/nvidia-ai-and-hpc-gpu-sales-reportedly-approached-half-a-million-units-in-q3-thanks-to-meta-facebook?)

4) Business Insider – ‘Project Greenland’: How Amazon overcame a GPU crunch (https://www.businessinsider.com/amazon-strategy-overcome-gpu-shortages-nvidia-2025-4)

5) IDC, Monitor Daily – IDC: Global PC Shipments Forecast to Recover in 2024 After Unprecedented Slump in Demand (https://www.monitordaily.com/idc-global-pc-shipments-forecast-to-recover-in-2024-after-unprecedented-slump-in-demand/)

6) Meta’s Engineering Blog – Building Meta’s GenAI Infrastructure (https://engineering.fb.com/2024/03/12/data-center-engineering/building-metas-genai-infrastructure/)

7) Black swan theory (https://en.wikipedia.org/wiki/Black_swan_theory)

8) Amazon Web Services: Amazon EC2 Capacity Blocks for ML pricing (https://aws.amazon.com/ar/ec2/capacityblocks/pricing/)

9) Thunder Compute: NVIDIA H100 Pricing (August 2025): Cheapest On‑Demand Cloud GPU Rates (https://www.thundercompute.com/blog/nvidia-h100-pricing)

10) Data Center Dynamics – Amazon’s retail business resolves internal GPU capacity shortage (https://www.datacenterdynamics.com/en/news/amazons-retail-business-resolves-internal-gpu-capacity-shortage/)

11) Reuters – Microsoft to spend record $30 billion this quarter as AI investments pay off (https://www.reuters.com/business/microsoft-spend-record-30-billion-this-quarter-ai-investments-pay-off-2025-07-30/)

12) Ethereum Miners Spent $15 Billion on GPUs Alone During Latest Cryptocraze (https://www.tomshardware.com/news/ethereum-miners-have-spent-15-billion-on-gpus?utm_source=chatgpt.com)

About the Author

Amogh Garg is a Senior Member of Technical Staff at Salesforce, where he is part of the Cloud Economics and Capacity Management team. He leads forecasting and planning of cloud compute supply, including GPUs that power Salesforce’s Agentforce AI platform. Previously, at AWS, Meta, and Microsoft, Amogh spearheaded infrastructure forecasting and large‑scale capacity management initiatives across some of the world’s largest technology platforms. His expertise spans demand modeling, supply chain optimization, and financial planning for cloud infrastructure.